the 23rd european modeling simulation symposium

20 downloads 2777 Views 30MB Size Report
Feb 21, 2011 - Online Collaborative Simulation Conceptual Model Development. Bhakti S. S. Onggo ..... Applications”, Custom Integrated Circuits Conference,. 2005. ...... Amazon's Elastic Computing, or Salesforce (Pedrinaci. 2010).
THE 23RD EUROPEAN MODELING & SIMULATION SYMPOSIUM SEPTEMBER 12-14 2011 ROME, ITALY

EDITED BY AGOSTINO G. BRUZZONE MIQUEL A. PIERA FRANCESCO LONGO PRISCILLA ELFREY MICHAEL AFFENZELLER OSMAN BALCI

PRINTED IN RENDE (CS), ITALY, SEPTEMBER 2011 ISBN 978-88-903724-4-5 I

¤ 2011 DIPTEM UNIVERSITÀ DI GENOVA RESPONSIBILITY FOR THE ACCURACY OF ALL STATEMENTS IN EACH PAPER RESTS SOLELY WITH THE AUTHOR(S). STATEMENTS ARE NOT NECESSARILY REPRESENTATIVE OF NOR ENDORSED BY THE DIPTEM, UNIVERSITY OF GENOA. PERMISSION IS GRANTED TO PHOTOCOPY PORTIONS OF THE PUBLICATION FOR PERSONAL USE AND FOR THE USE OF STUDENTS PROVIDING CREDIT IS GIVEN TO THE CONFERENCES AND PUBLICATION. PERMISSION DOES NOT EXTEND TO OTHER TYPES OF REPRODUCTION NOR TO COPYING FOR INCORPORATION INTO COMMERCIAL ADVERTISING NOR FOR ANY OTHER PROFIT – MAKING PURPOSE. OTHER PUBLICATIONS ARE ENCOURAGED TO INCLUDE 300 TO 500 WORD ABSTRACTS OR EXCERPTS FROM ANY PAPER CONTAINED IN THIS BOOK, PROVIDED CREDITS ARE GIVEN TO THE AUTHOR(S) AND THE WORKSHOP. FOR PERMISSION TO PUBLISH A COMPLETE PAPER WRITE TO: DIPTEM UNIVERSITY OF GENOA, DIRECTOR, VIA OPERA PIA 15, 16145 GENOVA, ITALY. ADDITIONAL COPIES OF THE PROCEEDINGS OF THE EMSS ARE AVAILABLE FROM DIPTEM UNIVERSITY OF GENOA, DIRECTOR, VIA OPERA PIA 15, 16145 GENOVA, ITALY.

II

THE 23RD EUROPEAN MODELING & SIMULATION SYMPOSIUM SEPTEMBER 12-14 2011 ROME, ITALY

ORGANIZED BY

DIPTEM – UNIVERSITY OF GENOA LIOPHANT SIMULATION SIMULATION TEAM IMCS – INTERNATIONAL MEDITERRANEAN & LATIN AMERICAN COUNCIL OF SIMULATION MECHANICAL DEPARTMENT, UNIVERSITY OF CALABRIA MSC-LES, MODELING & SIMULATION CENTER, LABORATORY OF ENTERPRISE SOLUTIONS MODELING AND SIMULATION CENTER OF EXCELLENCE (MSCOE) MISS LATVIAN CENTER – RIGA TECHNICAL UNIVERSITY LOGISIM LSIS – LABORATOIRE DES SCIENCES DE L’INFORMATION ET DES SYSTEMES MISS – UNIVERSITY OF PERUGIA MISS – BRASILIAN CENTER, LAMCE-COPPE-UFRJ MISS - MCLEOD INSTITUTE OF SIMULATION SCIENCES M&SNET - MCLEOD MODELING AND SIMULATION NETWORK LATVIAN SIMULATION SOCIETY

III

ECOLE SUPERIEURE D'INGENIERIE EN SCIENCES APPLIQUEES FACULTAD DE CIENCIAS EXACTAS. INGEGNERIA Y AGRIMENSURA UNIVERSITY OF LA LAGUNA CIFASIS: CONICET-UNR-UPCAM INSTICC - INSTITUTE FOR SYSTEMS AND TECHNOLOGIES OF INFORMATION, CONTROL AND COMMUNICATION

I3M 2011 INDUSTRIAL SPONSORS PRESAGIS CAE CAL-TEK MAST AEGIS TECHNOLOGIES I3M 2011 MEDIA PARTNERS MILITARY SIMULATION & TRAINING MAGAZINE EURO MERCI

IV

EDITORS AGOSTINO BRUZZONE MISS-DIPTEM, UNIVERSITY OF GENOA, ITALY [email protected]

MIQUEL A. PIERA

AUTONOMOUS UNIVERSITY OF BARCELONA, SPAIN [email protected]

FRANCESCO LONGO

MSC-LES, UNIVERSITY OF CALABRIA, ITALY [email protected]

PRISCILLA ELFREY

NASA-KSC, FL, USA [email protected]

MICHAEL AFFENZELLER

UPPER AUSTRIAN UNIVERSITY OF APPLIED SCIENCES, AUSTRIA [email protected]

OSMAN BALCI

VIRGINIA TECH, USA [email protected]

V

THE INTERNATIONAL MEDITERRANEAN AND LATIN AMERICAN MODELING MULTICONFERENCE, I3M 2011 GENERAL CO-CHAIRS AGOSTINO BRUZZONE, MISS DIPTEM, UNIVERSITY OF GENOA, ITALY MIQUEL ANGEL PIERA, AUTONOMOUS UNIVERSITY OF BARCELONA, SPAIN

PROGRAM CHAIR FRANCESCO LONGO, MSC-LES, UNIVERSITY OF CALABRIA, ITALY

THE 23RD EUROPEAN MODELING & SIMULATION SYMPOSIUM, EMSS 2011 GENERAL CO-CHAIRS FRANCESCO LONGO, MSC-LES, UNIVERSITY OF CALABRIA, ITALY PRISCILLA ELFREY, NASA-KSC, USA

PROGRAM CO-CHAIRS OSMAN BALCI, VIRGINIA TECH, USA MICHEAL AFFENZELLER, UPPER AUSTRIAN UNIVERSITY OF APPLIED SCIENCES, AUSTRIA

VI

EMSS 2011 INTERNATIONAL PROGRAM COMMITTEE

TRACKS AND WORKSHOP CHAIRS

MICHAEL AFFENZELLER, UPPER AUSTRIAN UNIV. OF AS, AUSTRIA WERNER BACKFRIEDER, UPPER AUSTRIAN UNIV. OF AS, AUSTRIA OSMAN BALCI, VIRGINIA TECH, USA STJEPAN BOGDAN, UNIVERSITY OF ZAGREB, CROATIA ENRICO BOCCA, SIMULATION TEAM, ITALY FELIX BREITENECKER, TECHNICAL UNIVERSITY OF WIEN, AUSTRIA AGOSTINO BRUZZONE, UNIVERSITY OF GENOA, ITALY SORIN DAN COTOFANA, DELFT UNIV. OF TECHNOLOGY, GERMANY GIANLUCA DE LEO, VMASC-ODU, USA STEPHAN DREISEITL, UPPER AUSTRIAN UNIV. OF AS, AUSTRIA PRISCILLA ELFREY, NASA-KSC, USA WENHUI FAN, TSINGHUA UNIVERSITY, CHINA MARIA PIA FANTI, POLYTECHNIC UNIVERSITY OF BARI, ITALY IDALIA FLORES, UNIVERSITY OF MEXICO, MEXICO CLAUDIA FRYDMAN, LSIS, FRANCE MURAT M.GÜNAL, TURKISH NAVAL ACADEMY, TURKE GRAHAM HORTON, MAGDEBURG UNIVERSITY, GERMANY AMIR HUSSAIN, UNIVERSITY OF STIRLING, SCOTLAND, UK WITOLD JACAK, UPPER AUSTRIAN UNIV. OF AS, AUSTRIA EMILIO JIMÉNEZ, UNIVERSITY OF LA RIOJA, SPAIN BERTHOLD KERSCHBAUMER, UPPER AUSTRIAN UNIV. OF AS, AUSTRIA DANKO KEZIC, UNIVERSITY OF SPLIT, CROATIA PETER KULCZYCKI, UPPER AUSTRIAN UNIV. OF AS, AUSTRIA JUAN IGNACIO LATORRE BIEL, UNIV. PÚBLICA DE NAVARRA, SPAIN FRANCESCO LONGO, MSC-LES, UNIVERSITY OF CALABRIA, ITALY MARINA MASSEI, LIOPHANT SIMULATION, ITALY YURI MERKURYEV, RIGA TECHNICAL UNIVERSITY, LATVIA TANYA MORENO CORONADO, UNIV. NAC. AUT. DE MÉXICO, MEXICO MIGUEL MÚJICA MOTA, UAB, SPAIN GASPER MUSIC, UNIVERSITY OF LJUBLJANA, SLOVENIA GABY NEUMANN, TECH. UNIV. APPL. SCIENCES WILDAU, GERMANY MUAZ NIAZI, COMSATS INSTITUTE OF IT, PAKISTAN TUDOR NICULIU, UNIVERSITY OF BUCHAREST, ROMANIA VERA NOVAK, HARVARD MEDICAL SCHOOL, USA TUNCER ÖREN, M&SNET, UNIVERSITY OF OTTAWA, CANADA FEDERICA PASCUCCI, UNIVERSITY OF ROMA 3, ITALY SEGURA PEREZ, UNIV. NAC. AUT. DE MÉXICO, MEXICO MERCEDES PEREZ DE LA PARTE, UNIVERSIDAD DE LA RIOJA, SPAIN MIQUEL ANGEL PIERA, UAB, SPAIN CESAR DE PRADA, UNIVERSIDAD DE VALLADOLID, SPAIN ROCCO RONGO, UNIVERSITY OF CALABRIA, ITALY STEFANO SAETTA, UNIVERSITY OF PERUGIA, ITALY ROBERTO SETOLA, UNIVERSITY OF ROMA 3, ITALY ROGER SMITH, WILLIAM SPATARO, UNIVERSITY OF CALABRIA, ITALY JERZY W. ROZENBLIT, UNIVERSITY OF ARIZONA, USA ALBERTO TREMORI, SIMULATION TEAM, ITALY LEVENT YILMAZ, AUBURN UNIVERSITY, USA STEPHAN WINKLER, UPPER AUSTRIAN UNIV. OF AS, AUSTRIA LIN ZHANG, BEIHANG UNIVERSITY, CHINA XUESONG ZHANG, JILIN UNIVERSITY, CHINA GERALD ZWETTLER, UPPER AUSTRIAN UNIV. OF AS, AUSTRIA

WORKSHOP ON MODELING & SIMULATION IN HEALTHCARE CHAIRS: WITOLD JACAK, UNIVERSITY OF APPLIED SCIENCES UPPER AUSTRIA (AUSTRIA); WERNER BACKFRIEDER, UNIVERSITY OF APPLIED SCIENCES UPPER AUSTRIA (AUSTRIA); MURAT M.GÜNAL, TURKISH NAVAL ACADEMY (TURKEY); JERZY W. ROZENBLIT, UNIVERSITY OF ARIZONA AGENT DIRECTED SIMULATION CHAIRS: TUNCER ÖREN, UNIVERSITY OF OTTAWA (CANADA); LEVENT YILMAZ, AUBURN UNIVERSITY (USA) CLOUD SIMULATION AND HIGH PERFORMANCE COMPUTING CHAIRS: PROF. LIN ZHANG, BEIHANG UNIVERSITY ,(BEIJING, CHINA) PROF. WENHUI FAN, TSINGHUA UNIVERSITY (CHINA) DISCRETE AND COMBINED SIMULATION CHAIR: GASPER MUSIC, UNIVERSITY OF LJUBLJANA, (SLOVENIA) HUMAN-CENTRED AND HUMAN-FOCUSED MODELLING AND SIMULATION CHAIRS: GABY NEUMANN, TECHNICAL UNIVERSITY OF APPLIED SCIENCES WILDAU (GERMANY); AGOSTINO BRUZZONE, MISSDIPTEM, UNIVERSITY OF GENOA, (ITALY) INDUSTRIAL PROCESSES MODELING & SIMULATION CHAIR: CESAR DE PRADA, UNIVERSIDAD DE VALLADOLID, (SPAIN) INDUSTRIAL ENGINEERING CHAIR: ENRICO BOCCA, SIMULATION TEAM, (ITALY) PETRI NETS BASED MODELLING & SIMULATION CHAIRS: EMILIO JIMÉNEZ, UNIVERSITY OF LA RIOJA (SPAIN); JUAN IGNACIO LATORRE, PUBLIC UNIVERSITY OF NAVARRE (SPAIN) SIMULATION AND ARTIFICIAL INTELLIGENCE CHAIR: TUDOR NICULIU, UNIVERSITY "POLITEHNICA" OF BUCHAREST (ROMANIA) SIMULATION OPTIMIZATION APPROACHES IN INDUSTRY, SERVICES AND LOGISTICS PROCESSES

CHAIRS: IDALIA FLORES, UNIVERSITY OF MEXICO; MIGUEL MÚJICA MOTA, UNIVERSITAT AUTONOMA DE BARCELONA (SPAIN).

VII

GENERAL CO-CHAIRS’ MESSAGE WELCOME TO EMSS 2011 Building on the long success of 22 editions, the 23th European Modeling & Simulation Symposium (also known since 1996 as "Simulation in Industry") is an important forum to discuss theories, practices and experiences on M&S (Modeling & Simulation). EMSS 2011 brings together people from Academia, Agencies and Industries from all over the world, despite the Symposium name referring just to Europe; in fact EMSS 2011 represents a unique opportunity within I3M2011 framework to share experiences and ideas and to generate a new archival source for innovative papers on M&Srelated topics. The Symposium is also meant to provide information, identify directions for further research and to be an ongoing framework for knowledge sharing; the structure of the conferences, strongly based on Tracks, allows to create synergies among different groups keeping pretty sharp each framework; the quality of the papers is the stronghold of this event and even this year it was possible to apply severe selection on the submissions and to guarantee top level papers, therefore the conference in Rome is one of the largest EMSS organized during last years. In fact, another strong feature for the 2011 edition of EMSS is the site: the event is located in Rome, a city that for his history and cultural background was called Caput Mundi (Head of the world). The EMSS 2011 Program, Presentations, People and Place make it a professionally worthwhile and a personally enjoyable experience: so welcome to EMSS 2011.

Francesco Longo MSC-LES University of Calabria

Priscilla Elfrey Kennedy Space Center NASA

VIII

ACKNOWLEDGEMENTS The EMSS 2011 International Program Committee (IPC) has selected the papers for the Conference among many submissions; therefore, based on this effort, a very successful event is expected. The EMSS 2011 IPC would like to thank all the authors as well as the reviewers for their invaluable work. A special thank goes to all the organizations, institutions and societies that have supported and technically sponsored the event.

LOCAL ORGANIZATION COMMITTEE AGOSTINO G. BRUZZONE, MISS-DIPTEM, UNIVERSITY OF GENOA, ITALY ENRICO BOCCA, SIMULATION TEAM, ITALY FRANCESCO LONGO, MSC-LES, UNIVERSITY OF CALABRIA, ITALY FRANCESCA MADEO, UNIVERSITY OF GENOA, ITALY MARINA MASSEI, LIOPHANT SIMULATION, ITALY LETIZIA NICOLETTI, CAL-TEK SRL FEDERICO TARONE, SIMULATION TEAM, ITALY ALBERTO TREMORI, SIMULATION TEAM, ITALY

IX

This International Conference is part of the I3M Multiconference: the Congress leading Simulation around the World and Along the Years I3M

Simulation around the World & along the Years I3M2010, Fes, Fes, Morocco

I3M

www.liophant.org/i3m

I3M

Simulation around the World & along the Years I3M2007, Bergeggi, Bergeggi, Italy

Simulation around the World & along the Years I3M2004, Liguria, Liguria, Italy

I3M

Simulation around the World & along the Years HMS2002, Marseille, Marseille, France

www.liophant.org/i3m

EMSS

Simulation around the World & along the Years ESS1993, Delft, The Netherlands

www.liophant.org/i3m

Simulation around the World & along the Years I3M2008, Calabria, Calabria, Italy

I3M

Simulation around the World & along the Years I3M2005, Marseille, France

I3M

Simulation around the World & along the Years HMS2003, Riga, Latvia

www.liophant.org/i3m

EMSS

Simulation around the World & along the Years ESS1994, Istanbul, Turkey

www.liophant.org/i3m

X

Simulation around the World & along the Years I3M2009, Tenerife, Spain

www.liophant.org/i3m

I3M

www.liophant.org/i3m

I3M

Simulation around the World & along the Years I3M2012, Wien, Wien, Austria

www.liophant.org/i3m

www.liophant.org/i3m

www.liophant.org/i3m

I3M

I3M

www.liophant.org/i3m

www.liophant.org/i3m

I3M

Simulation around the World & along the Years I3M2011, Rome, Italy

Simulation around the World & along the Years I3M2006, Barcelona, Spain

www.liophant.org/i3m

I3M

Simulation around the World & along the Years HMS2004, Rio de Janeiro, Brazil

www.liophant.org/i3m

EMSS

Simulation around the World & along the Years ESS1996, Genoa, Italy

www.liophant.org/i3m

Index



Controllable Equivalent Resistance CMOS Active Resistor with Improved Accuracy and Increased Frequency Response Cosmin Popa

1

A Local Search Genetic Algorithm For The Job Shop Scheduling Problem Mebarek Kebabla, Hayet Mouss, Nadia Mouss

5

Temporal Neuro-Fuzzy Systems in Fault Diagnosis and Prognosis Rafik Mahdaoui, Hayet Mouss, Djamel Mouss, Ouahiba Chouhal

11

Optimization of automated guided vehicle rules for a multiple-load AGV system using simulation and SAW, VICOR and TOPSIS methods in a FMS environment Parham Azimi, Masuomeh Gardeshi

17

A New Method for the Validation and Optimisation of Unstable Discrete Event Models Hans-Peter Barbey

25

Simulation Helps Assess and Increase Airplane Manufacturing Capacity Marcelo Zottolo, Edward Williams, Onur Ulgen

30

Calibration of process algebra models of discretely observed stochastic biochemical system Paola Lecca

36

Denervated Muscle Undergoing Electrical Stimulation: Development Of Monitoring Techniques Based On Medical Image Modelling Paolo Gargiulo, Thomas Mandl, Egill A. Friðgeirsson, Ilaria Bagnaro, Thordur Helgason, Páll Ingvarsson, Marcello Bracale, Winfried Mayr, Ugo Carraro, Helmut Kern

44

3d Segmented Model Of Head For Modelling Electrical Activity Of Brain Egill A. Friðgeirsson, Paolo Gargiulo, Ceon Ramon, Jens Haueisen

50

An Agent-Based Information Foraging Model of Scientific Knowledge Creation and Spillover in Open Science Communities Ozgur Ozmen, Levent Yilmaz

55

Simulation Highway – Direct Access Intelligent Cloud Simulator Egils Ginters, Inita Sakne, Ieva Lauberte, Artis Aizstrauts, Girts Dreija, Rosa Maria Aquilar Chinea, Yuri Merkuryev, Leonid Novitsky, Janis Grundspenkis

62

Simulation of Gesmey Generator Manoeuvers Amable López, José A. Somolinos, Luis R. Núñez, Alfonso M. Carneros

72

Using Semantic Web Technologies to Compose Live Virtual Constructive (LVC) Systems Warren Bizub, Julia Brandt, Meggan Schoenberg

78

Side Differences in Mri-Scans In Facial Palsy: 3-D Modelling, Segmentation And Voxel Gradient Changes Paolo Gargiulo, Carsten Michael Klingner, Egill A. Friðgeirsson, Hartmut Peter Burmeister, Gerd Fabian Volk, Orlando Guntinas-Lichius

87

Exploiting Variance Behavior in Simulation-based Optimization Pasquale Legato, Rina Mary Mazza

93

XI

Accelerated fully 3D iterative reconstruction in SPECT Werner Backfrieder, Gerald Zwettler

100

Study on the servilization of simulation capability Y L Luo, L Zhang, F Tao, Y Bao, L Ren

105

Gyrus And Sulcus Modelling Utilizing a Generic Topography Analysis Strategy for Processing Arbitrarily Oriented 3d Surfaces Gerald Zwettler, Werner Backfrieder

111

Fast Marching Method Based Path Planning for Wheeled Mobile Robots Gregor Klancar, Gasper Music

118

Modeling and Simulation Architecture For Cloud Computing and Internet of Things (IoT) Based Distributed Cyber-Physical Systems (DCPS) Xie Lulu, Wang Zhongjie

127

Transport Network Optimiziation: Self-Organization by Genetic Programming Johannes Göbel, Anthony E. Krzesinski, Bernd Page

137

3D Physics Based Modeling and Simulation of Intrinsic Stress in SiGe for Nano PMOSFETs Abderrazzak EL boukili

144

Simulation model for the calcination process of cement Idalia Flores, Guillermo Perea

150

Job Satisfaction Modelling in Agent-Based Simulations Alexander Tarvid

158

Simultaneous Scheduling o Machines ad Operators in a Multi-Resource Coinstrained Job-Shop Scenario Lorenzo Tiacci, Stefano Saetta

166

Effect of reject option on classifier performance Stephan Dreiseitl, Melanie Osl

176

New discrete Topology Optimization method for industrial tasks Sierk Fiebig

181

3G Mobile Network Planning Based On A Traffic Simulation Model And A CostBenefit Model To Service Los Cabos International Airport Aida Huerta Barrientos, Mayra Elizondo Cortes

187

Providing Semantic Interoperability for Integrated Healthcare Using a Model Transformation Approach Barbara Franz, Herwig Mayr

195

Management of supply networks using PDES Carmine De Nicola, Rosanna Manzo, Luigi Rarità

201

Sugar Factory Benchmark

211

Rogelio Mazaeda, Alexander Rodriguez, Alejandro Merino, Cesar De Prada, Luis Felipe Acebes, Designing and Implementing a Model to Examine R&D section´s capabilities with emphasis on reversed engineering in chemical Factory Neda Khadem Geraili, Mona Benhari

220

On the Use of Minimum-Bias Computer Experimental Designs Husam Hamad

229

A Practical Guide for the Initialisation of Multi-Agent Systems with Random Number Sequences from Aggregated Correlation Data Volker Nissen, Danilo Saft

235

XII

Multi-agent multi-level Modeling – A methodology to simulate complex systems Jean-Baptiste Soyez, Gildas Morvan, Rochdi Merzouki, Daniel Dupont

241

Indoor Pedestrian Navigation Simulation Via A Network Framework John Usher, Eric Kolstad

247

Reconfigurable Human-System Cosimulation Tudor-Razvan Niculiu, Maria Niculiu

254

An Asynchronous Parallel Hybrid Optimization Approach To Simulation-Based Mixed-Integer Nonlinear Problems Kathleen Fowler, Timothy Kopp, Jacob Orsini, Josh Griffin, Genetha Gray

264

Reconstruction of clinical workflows based on the IHE integration profile "Cross-Enterprise Document Workflow" Melanie Strasser, Franz Pfeifer, Emmanuel Helm, Andreas Schuler, Josef Altmann

272

A Methodology For Developing Des Models: Event Graphs And Sharpsim Arda Ceylan, Murat Gunal

278

Revisitation o The Simulation Methodologies And Applications In Manufacturing T. Radha Ramanan, Ihsan Sabuncuoglu

283

Insights into the Practice of Expert Simulation Modellers Rizwan Ahmed, Mahmood Shah

289

Modelling Resilience in Cloud-Scale Data Centres John Cartlidge, Ilango Sriram

299

Advanced Container Transportation Equipment using Transfer Robot and Alignment System Young Jin Lee, Dong Seop Han, Dae Woo Kang, Duk Kyun Lee, Geun Choi, Kwon Soon Lee,

308

Global Context Influences Local Decision

314

Terry Bossomaier, Michael Harrè Modeling And Simulation Of Petri Nets For Complex Scheduling Rules Of Automated Manufacturing Systems Chulhan Kim, Tae-Eog Lee

319

A Business Process Modeling Approach To Support Production Systems Analysis And Simulation Claudia Battista, Giulia Dello Stritto, Francesco Giordano, Raffaele Iannone, Massimiliano M. Schiraldi

325

Online Collaborative Simulation Conceptual Model Development Bhakti S. S. Onggo, Suchismita Hoare

333

dSPACE Based direct-driven permanent magnet synchronous wind power system modeling and simulation Yan-xia Shen, Xiang-xia Liu, Zhi-cheng Ji, Ting-long Pan

340

Missing Data Estimation for Cancer Diagnosis Support Witold Jacak, Karin Proell

345

A New Devs-Based Generic Artficial Neural Network Modeling Approach Samuel Toma, Laurent Capocchi, Dominique Federici

351

An Improved Time-Line Search Algorithm To Optimize Industrial Systems Miguel Mujica, Miquel Angel Piera

357

XIII

Simulation and Model Calibration With Sensitivity Analysis For Threat Detection in Brain Keegan Lowenstein, Brian Leventhal, Kylie Drouin, Robert Dowman, Katie Fowler, Sumona Mondal

363

Simulation And Modelling Of The Flat-Band Voltage For Below 200nm SOI Devices Cristian Ravariu, Florin Babarada

371

Simulation of Human Behavior in Situation of Emergency

375

Samira Benkhedda, Fatima Bendella, Karima Belmabrouk Simulation, Optimisation and Design a Platform for in-vivo Electrophysiological Signals Processing Florin Babarada, Cristian Ravariu, Janel Arhip

380

A Simulation-Based Framework for Industrial Automated Wet-Etch Station Scheduling Problem In The Semiconductor Industry

384

Adrian Aguirre, Vanina Cafaro, Carlos Mendez, Pedro Castro Data Stream Management in Income Tax Microsimulation Models Istvan Molnar, Gyorgy Lipovszki

394

Experimental manufacturing system for research and training on humancentred simulation Diego Crespo Pereira, David del Rio Vilas, Rosa Rios Prado, Nadia Rego Monteil, Adolfo Lamas Rodriguez

400

Modelling and Simulation of a Wireless Body Area Network Prototype for Health Monitoring Yeray Callero, Rosa María Aguilar

410

An Extreme Learning Machine Algorithm for Higher Order Neural Network Models Shuxiang Xu

418

Increasing Availability of Production Flow Lines Through Optimal Buffer Sizing: a Simulative Study Vittorio Cesarotti, Alessio Giuiusa, Vito Introna

423

Using Query Extension And User Feedback To Improve Pubmed Search Viktoria Dorfer, Sophie A. Blank, Stephan M. Winkler, Thomas Kern, Gerald Petz, Patrizia Faschang

433

Simulation of The Vessel Traffic Schedule In The Strait Of Istanbul ùirin Özlem, lhan Or, Birnur

439

New Genetic Programming Hypothesis Search Strategies for Improving the Interpretability in Medical Data Mining Applications Michael Affenzeller, Christian Fischer, Gabriel Kronberger, Stephan M. Winkler, Stefan Wagner

448

On the Use of Estimated Tumor Marker Classifications in Tumor Diagnosis Prediction - A Case Study for Breast Cancer Stephan Winkler, Michael Affenzeller, Gabriel Kronberger, Michael Kommenda, Stefan Wagner, Witold Jacak, Herbert Stekel

454

Automatic Selection of Relevant Data During Ultrasonic Inspection Thouraya Merazi Meksen, Malika Boudraa, Bachir Boudraa

460

CT Based Models for Monitoring Bone Changes in Paraplegic Patients Undergoing Functional Electrical Stimulation Páll Jens Reynisson, Benedikt Helgason, Stephen J. Ferguson, Thordur

465

XIV

Helgason, Rúnar Unnþórsson, Páll Ingvarsson, Helmut Kern, Winfried Mayr, Ugo Carraro, Paolo Gargiulo A High Resolution Distributed Agent-Based Simulation Environment for Large-Scale Emergency Response Glenn Hawe, Graham Coates, Duncan Wilson, Roger Crouch

470

Eigenfrequency Based Sensitivity Analysis Of Vehicle Drivetrain Oscillations Oliver Manuel Krieg, Jens Neumann, Bernhard Hoess, Heinz Ulbrich

478

Improving Job Scheduling on a Heterogeneous Cluster by Predicting Job Execution Times Using Heuristics Hannes Brandstätter-Müller, Bahram Parsapour, Andreas Hölzlwimmer, Gerald Lirk, Peter Kulczycki

488

Agent-Based Simulation Of Electronic Marketplaces With Ontology-Services Maria João Viamonte, Nuno Silva, Paulo Maio

496

Key Issues In Cloud Simulation Platform Based On Cloud Computing Lei Ren, Lin Zhang, Yabin Zhang, Yongliang Luo, Qian Li

502

Research On Key Technologies Of Resource Management In Cloud Simulation Platform Ting Yu Lin, Xu Dong Chai, Bo Hu Li

508

A Framework For Enhanced Project Schedule Design To Aid Project Manager’s Decision Making Processes Sanja Lazarova-Molnar, Rabeb Mizouni, Nader Kesserwan

516

A Novel Approach To Realistic Modeling And Simulation Of State-Varying Failure Rates Sanja Lazarova-Molnar

526

Modeling and Simulation of Order Sortation Systems Fahrettin Eldemir, Elif Karakaya

535

Development Of The Surface To Air Missile Simulator Through The Process Of Component Composition And Dynamic Reconfiguration Of Weapon System Jeebeom Suk, Jaeoh Lee, Yoonho Seo

541

Modeling and simulating a benchmark on dynamic reliability as a Stochastic Activity Network Daniele Codetta-Raiteri

545

A Stochastic Approach To Risk Modeling For Solvency Ii Vojo Bubevski

555

Design and Implementation of a Fuzzy Cognitive Maps Expert System for Oil Price Estimation Mohamad Ali Azadeh, Zeinab sadat Ghaemmohamadi

562

How to benefit more from intuitive power and experience of the human simulation knowledge stakeholder Gaby Neumann

568

Object-oriented modelling and verification aided by model simplification techniques Anton Sodja, Borut Zupancic

574

The exclusive entities in the formalization of a decision problem based on a discrete event system by means of Petri nets Juan Ignacio Latorre-Biel, Emilio Jimenez-Macias

580

Matrix-based operations and equivalente classes in alternative Petri nets Juan Ignacio Latorre, Emilio Jimenez

587

XV

Synthesis of Feedback Controller for Stabilization of Chaotic Henon Map Oscillations by Means of Analytic Programming Roman Senkerik, Zuzana Oplatkova, Ivan Zelinka, Donald Davendra, Roman Jasek

593

A Simulation Study of The Interdependence Of Scalability And Cannibalization In The Software Industry Francesco Novelli

599

Security in sending and storage of Petri nets by signing and encription Iñigo León Samaniego, Mercedes Pérez de la Parte, Eduardo Martínez Camara, Juan Carlos Sáenz-Díez Muro

605

Petri net transformation for decision making: compound Petri nets to alternatives aggregation Petri nets Juan Ignacio Latorre, Emilio Jimenez

613

Improvements in the optimization of Flexible Manufacturing Cells modelled with Discrete Event Dynamics Systems: Application to a real factory problem Diego Rodriguez, Mercedes Perez, Juan Manuel Blanco

619

A Simulation-Based Capacity Planning Model: a Case Study In a Contract Furnishing Sme Nadia Rego Monteil, David del Rio Vilas, Diego Crespo Pereira, Rosa Rios Prado, Arturo Nieto de Almeida

626

PN as a tool for innovation in industry: a review Jesús Fernandez de Miguel, Julio Blanco Fernandez, Mercedes Perez

635

Stochastic Optimization of Industrial Maintenance Strategies Francisco Castellanos, Ann Wellens

642

Design And Development of Data Analysis Modules For The Aermod And Calpuff Simulation Models Ann Wellens, Gamar García

648

Development Of A Simulation Tool For Consecuence Analysis In Industrial Instalations Victor Pérez, Gamar García, Maria Guadalupe Ávila, Francisco Castellanos, Ann Wellens

654

Using Markov Chain and Graph Theory Concepts to Analyze Behavior in Complex Distributed Systems Christopher Dabrowski, Fern Hunt

659

Formal Framework For The Devs-Driven Modeling Language Ufuoma Ighoroje, Oumar Maïga, Mamadou Traoré

669

A Methodology for the DEVS Simulation Graph Construction Adedoyin Adegoke, Ibrahima Hamadou, Hamidu Togo, Mamadou Traoré,

675

Efficient exploration of Coloured Petri net based scheduling problem solutions Gasper Music

681

Plant Capacity Analysis in a Dairy Company, Applying Montecarlo Simulation Joselito Medina-Marin, Gilberto Perez-Lechuga, Juan Carlos Seck-Tuoh-Mora, Norberto Hernandez-Romero, Isaias Simon-Marmolejo

690

GPGPU Programming and Cellular Automata: Implementation of The Sciara Lava Flow Simulation Code Giuseppe Filippone, William Spataro, Giuseppe Spingola, Donato D'Ambrosio, Rocco Rongo, Giovanni Perna, Salvatore Di Gregorio

696

XVI

Neighborhood Concept for Modeling an Adaptive Routing in Wireless Sensors Network Jan Nikodem, Maciej Nikodem, Ryszard Klempous, Marek Woda, Zenon Chaczko

703

A DEVS-based Simplified Business Process Modelling Library Igor Rust, Deniz Cetinkaya, Mamadou Seck, Ivo Wenzler

709

Research on co-simulation task scheduling in Cloud Simulation Platform Chen Yang, Bo Hu Li, Xudong Chai

715

An Optimal Non-Blocking Dispatching in Free-Choice Manufacturing Flowlines by Using Machine-Job Incidence Matrix Ivica Sindicic, Stjepan Bogdan, Tamara Petrovic

722

Simulation of Vascular Volume Pulsation of Radial Index Artery Pichitra Uangpairoj, Masahiro Shibata

728

Coding Tcpn Models Into The Simio Simulation Environment Miguel Mujica, Miquel Angel Piera

734

Developing a Simulation Training Tool from a Medical Protocol Catherine M. Banks, John A. Sokolowski

740

Model Synthesis Using a Multi-Agent Learning Strategy Sebastian Bohlmann, Arne Klauke, Volkhard Klinger, Helena Szczerbicka

747

Service Optimization For System-Of-Systems Based On Pool Scheduling And Inventory Management Driven By Smart Simulation Solutions Agostino Bruzzone, Marina Massei, Enrico Bocca

755

Modeling Of Obesity Epidemics By Intelligent Agents Agostino Bruzzone, Vera Novak, Francesca Madeo, Cecilia Cereda

768

Maritime Security: Emerging Technologies for Asymmetric Threats

775

Agostino Bruzzone, Marina Massei, Alberto Tremori, Francesco Madeo, Federico Tarone, Francesco Longo On The Short Period Production Planning in Industrial Plants: A Real Case Study

782

Agostino Bruzzone, Fracesco Longo, Letizia Nicoletti, Rafael Diaz 792

Authors’ Index



XVII

XVIII

CONTROLLABLE EQUIVALENT RESISTANCE CMOS ACTIVE RESISTOR WITH IMPROVED ACCURACY AND INCREASED FREQUENCY RESPONSE Cosmin Popa University Politehnica of Bucharest, Faculty of Electronics, Telecommunications and Information Technology, Romania [email protected]

(mobility degradation, bulk effect and short-channel effect) limits the circuit linearity introducing odd and even-order distortions, as shown in [4]. For this reason, an improved linearisation technique has to be design to compensate the nonlinearities introduced by the secondorder effects.

ABSTRACT

A new active resistor circuit will be further presented. The main advantages of the original proposed implementation are the improved linearity, the small area consumption and the improved frequency response. An original technique for linearizing the I (V ) characteristic of the active resistor will be proposed, based on the simulation of the Ohm law using two linearized differential amplifiers, a multiplier and a current-pass circuit. The controllability of the active resistor circuit is excellent, existing the possibility of modifying the value of the equivalent resistance by changing the ratio between a control voltage and a control current. Additionally, the value of the simulated resistance is not function on technological parameters, with the result of improved circuit accuracy. The errors introduced by the secondorder effects will be also strongly reduced, while the area consumption of the active resistor will be minimized by replacing the classical MOS transistor with FGMOS (Floating Gate MOS) devices.

II. THEORETICAL ANALYSIS The original idea for implementing a linear currentvoltage characteristic of the active resistor, similar to the characteristic of a classical passive resistor is to simulate the Ohm law using two linearized differential amplifiers and a multiplier circuit. Because of the requirements for a good frequency response, only MOS transistors working in saturation could be used. 2.1. The block diagram of the active resistor The structure of the proposed active resistor is based on four important blocks: two differential amplifiers AD1 and AD 2 with linear transfer function, a multiplier circuit MULT and a current-pass circuit I , the block diagram being presented in Figure 1. The I XY current, which is passing through the I block, is generated by the multiplier circuit, I XY I O I 2 / I 1 , while I 1 and I 2 currents are obtained from the differential amplifiers AD1 and AD 2 , I 1 g m1VO and I 2 g m 2 (V X  VY ) . It results:

Keywords: active resistor, differential amplifiers, linearity error, second-order effects

I. INTRODUCTION CMOS active resistors are very important blocks in VLSI analog designs, mainly used for replacing the large value passive resistors, with the great advantage of a much smaller area occupied on silicon. Their utilization domains includes amplitude control in low distortion oscillators, voltage controlled amplifiers and active RC filters. These important applications for programmable floating resistors have motivated a significant research effort for linearising their current-voltage characteristic. The first generation of MOS active resistors [1], [2] used MOS transistors working in the linear region. The main disadvantage is that the realised active resistor is inherently nonlinear and the distortion components were complex functions on MOS technological parameters. A better design of CMOS active resistors is based on MOS transistors working in saturation [3], [4], [5]. Because of the quadratic characteristic of the MOS transistor, some linearisation techniques were developed in order to minimize the nonlinear terms from the currentvoltage characteristic of the active resistor. Usually, the resulting linearisation of the I  V characteristic is obtained by a first-order analysis. However, the secondorder effects which affect the MOS transistor operation

I XY

IO

g m 2 V X  VY . g m1 VO

(1)

The equivalent resistance of the circuit having the block diagram presented in Figure 1 will be: Rech.

V X  VY I XY

VO g m1 I O g m2

VO IO

(W / L) 1 I O1 . (W / L) 2 I O 2

(2)

The great advantage of the proposed implementation of the active resistor is the very good controllability of the equivalent resistance by the ratio of a control voltage VO and a control current I O . As it is shown in (2), the value of the resistance does not depend on technological parameters, (W / L) 1 and (W / L) 2 representing aspect ratios of the transistors composing the differential 1

amplifiers AD1 ad AD 2 , respectively, while I O1 and I O 2 being biasing currents of these amplifiers.

I d1, 2 (vid ) #

VO

I1 IXY

IO

MULT

2.3. The original linearized differential amplifier The original proposed differential structure from Figure2 is based on a symmetrical structure that assures, in a firstorder analysis, the linearization of the transfer characteristic, equivalent to a constant circuit transconductance. Supposing a saturation operation of all MOS active devices from the previous circuit, it is possible to write that the output current expression is I 2 I X  I Y :

I2

AD2 IXY VX

IXY

IXY

I

(4)

where I O is the biasing current for the differential amplifier. In order to improve the circuit linearity, especially for large values of the differential input voltage ( THD has relatively large values for v id of about hundreds of mV ), a linearization technique has to be implemented.

AD1

IO

1/ 2 IO K 1 / 2 IO K3/2 3 r vid # v # .. , 1 / 2 id 2 2 16 I O

VY

Figure 1: The block diagram of the active resistor



2.2. Classical CMOS differential amplifier The most common approach of a differential amplifier in CMOS technology is based on strong-inverted MOS transistors (usually working in the saturation region), having the most important advantage of a much better frequency response with respect to the weak-inverted MOS differential amplifiers. As a result of the quadratic characteristic of a MOS transistor operating in saturation, the transfer characteristic of the classical CMOS differential amplifier will be strongly nonlinear, its linearity being in reasonable limits only for a very limited range of the differential input voltage. The drain currents of the classical CMOS differential amplifier will have the following nonlinear dependence on the differential input voltage, v id :





K VGS X  VGSY VGS X  VGSY  2VT . 2

I2

(5)

Because: VGS X  VGSY

2 V X  VY

(6)

2VGSO ,

(7)

and: VGS X  VGSY

it results a linear dependence of the output current on the differential input voltage: I2

8 KI O V X  VY ,

(8)

equivalent to a constant transconductance of the circuit: I d1, 2

IO IO r 2 2

4 § Kv 2 K 2vid ¨ id  ¨ IO 4 IO2 ©

1/ 2

· ¸ ¸ ¹

,

having a Taylor expansion around vid limited expressed by:

(3)

gm

0 , fifth-order

2

8 KI O .

(9)

VDD

I2 IX VX

T1

IY

T2

TX

TY

IO

VY

Figure 2: The linearized differential structure The important advantages of the previous circuit is the improved linearity that could be achieved in a first-order analysis and the possibility of controlling the value of the transconductance by modifying a continuous current ( I O ) .

2.5. The current-pass circuit The necessity of designing this circuit is derived from the requirement that the same current to pass between the two output pins, X and Y . The implementation in CMOS technology of this circuit is very simple, consisting in a simple and a multiple current mirrors (Figure 3).

2.4. The second-order effects The linearity (8) of the transfer characteristic of the differential amplifier from Figure 2 is slightly affected by the second-order effects that affect the MOS transistor operation, modeled by the following relations: channellength modulation (10) and mobility degradation (11)). ID

K

K VGS  VT 2 1  OV DS 2 K0 . >1  T G (VGS  VT )@(1  T DV DS )

VDD

IXY

(10)

VX

(11)

VT 

2I D I TG D . K K

2.6. The multiplier circuit The original idea for obtaining the multiplying function is to use two identical square-root circuits, implementing the following functions:

(12)

I OUT 1

2 IO I2

(14)

2 I XY I 1 .

(15)

and:

I OUT 2

I OUT 1 and I OUT 2 being the output currents of these square-root circuits. Using a classical current mirror, it is possible to impose that I OUT 1 I OUT 2 , resulting the necessary multiplying function:

f

¦ ak VX  VY k .

IXY

Figure 3: The current-pass circuit

The last term represents the error which affects the quadratic characteristic of the MOS transistor biased in saturation, caused by the previous presented second-order effects. The result will be a small accuracy degradation of the entire circuit linearity, quantitative evaluated by the superior-order terms in the transfer characteristic of the differential amplifier:

I XY

VY

-VDD

Considering that the design condition O T D is fulfilled, the gate-source voltage of a MOS transistor working in saturation at a drain current I D will be:

VGS

IXY

(13)

k 1

I XY

Because of the circuit symmetry, the odd-order terms from the previous relation are usually cancel out, so the main circuit nonlinearity caused by the second-order effects will be represented by the third-order error term from the previous relation, having much smaller value than the linear term.

I IO 2 . I1

(16)

The important advantages of this implementation are represented by the increased frequency response that could be obtained as a result of the current-mode operation of the multiplier circuit and of the biasing in saturation of all the MOS active devices and, additionally, by the reduced 3

silicon occupied area achieved by using exclusively MOS transistors.

ACKNOWLEDGMENTS The work has been co-funded by the Sectoral Operational Programme Human Resources Development 2007-2013 of the Romanian Ministry of Labour, Family and Social Protection through the Financial Agreement, POSDRU/89/1.5/S/62557.

2.7. Active resistor with negative equivalent resistance Active resistors with controllable negative equivalent resistance cover a specific area of VLSI designs, finding very large domains of applications such as the canceling of an operational amplifier load or the design of integrators with improved performances. In order to obtain a negative equivalent active resistance circuit, the block diagram from Figure 1 must be modified by inversing the sense of the I XY current passing through the I block, resulting an equivalent resistance expressed by:

Rech. ' 

V0 I0

(W / L) 1 I 01 . (W / L) 2 I 02

REFERENCES 1. Z. Wang, “Current-controlled Linear MOS Earthed and Floating Resistors and Application”, IEEE Proceedings on Circuits, Devices and Systems, 1990, pp. 479-481. 2. L. Sellami, “Linear Bilateral CMOS Resistor for Neuraltype Circuits”, Proceedings of the 40th Midwest Symposium on Circuits and Systems, 1997, pp. 1330-1333. 3. E. Ozalevli, P. Hasler, „Design of a CMOS Floating-Gate Resistor for Highly Linear Amplifier and Multiplier Applications”, Custom Integrated Circuits Conference, 2005. Proceedings of the IEEE 2005, 18-21 Sept. 2005, pp. 735-738. 4. K. Kaewdang, W. Surakampontom, N. Fujii, „A Design of CMOS Tunable Current Amplifiers”, Communications and Information Technology, 2004. ISCIT 2004. IEEE International Symposium on, 26-29 Oct. 2004, pp. 519-522. 5. F. Bahmani, E. Sanchez-Sinencio, „A Highly Linear Pseudo-Differential Transconductance, Solid-State Circuits Conference, 2004. ESSCIRC 2004. Proceeding of the 30th European, 21-23 Sept. 2004, pp. 111-114.

(17)

III. CONCLUSIONS A new active resistor circuit has been presented. The main advantages of the original proposed implementation are the improved linearity, the small area consumption and the improved frequency response. An original technique for linearizing the I (V ) characteristic of the active resistor has been proposed, based on the simulation of the Ohm law using two linearized differential amplifiers, a multiplier and a current-pass circuit. The controllability of the active resistor circuit is excellent, existing the possibility of modifying the value of the equivalent resistance by changing the ratio between a control voltage and a control current. Additionally, the value of the simulated resistance is not function on technological parameters, with the result of an improved circuit accuracy. The errors introduced by the second-order effects have been also strongly reduced, while the area consumption of the active resistor has been minimized by replacing the classical MOS transistor with FGMOS devices. As a result of the proposed linearization technique designed for the differential amplifier from Figure 2, the linearity (9) of its transfer characteristic is referring both to small and large signal operation, being limited only by the second-order effects that affect the MOS transistors’ operation. The consequence will be a relatively large range of the input voltage that could be applied across the input pins ( V X and VY from Figure 1), respecting the important restriction of maintaining the estimated circuit linearity.

AUTHORS BIOGRAPHY Cosmin Popa is with University Politehnica of Bucharest, Faculty of Electronics, Telecommunications and Information Technology, Romania. His area of interest includes analog integrated and mixed-signal VLSI designs. He is author of more than 140 scientific papers and of 5 research books.

4

              !"#$ $                  !"  #   !" $  !"

#*#"!,*'""#" #'#$3"$"/ +#' " *? ,"$## #+-."*' - .+#" *+ "',- .#$##"0- . 48=6& ',"'$*)/)3 $ #+' +'%1 "! " #0 "'#*,)"" #//"& 03 "'), '$"'"*##"3$// * #"!! ))! #+#! #+"*'# 1#$+"$/*' '3/*$#'! $,"'$48@6&3"' ) #"$""'""')/,#*/+#" *+ "',# ,*, #" )" , 2" #),0 " #/ ""'"/# +''$*'+ "',-&+& **' ,"$## #+#$"*'.& ! $ ,"'$/+#" *+ "',#$"'# +''$ *'+ "',0))$#$"' ' +' )/,#*0)"$4(  !!!  (#(( - #*A .&*' )" # *'*" 2$!)* /! #+"'"'1 $ ,*' #' ∈ )'8!!! ' *#$"'/ :$)* #+" ,  + 94@6&

0', "'*,)" #" ,/(% & 8)#"#:,)/(%')), /,$/=(-%8%>%=.0' *')*$#=,*' # -'8'>'=.&"' )," #*'$  )#"$! #""*'" # +&8&

"',3*#" #"*# $$#(#$ ,*' # "$/0? - .*'(,")"'+'*',*' ##*#$ #!#*& - .*'('$)*$"'+'"',*' # #)" *$& - .*')" #:*"$,"# #")"$# + 3#,*' #-#)%,)" # 0$.& - 3.*',*' #*##!)*#)" #" " ,&



#:,)/(%')*'$ #+), &? '8?@'>?='=?=  '? '8?8'=?7'>?= (? '>?>'8?@'=?8

'(*" 3 "$", #/ *'$0 "' , # , - &&, # , 2 #+"'*,)" #" ,/ "'"(.0' *' "',"*,,#+/"' ),? 8>= 8 > =

8>=@78=8@ ,-# ".

 +&8& #""%'")#"" #/" #/"' #"#* #  &

&

+)*$/"' + "', $* $ /! /0?

      

3 #+"' ),0 #"$*$/"#$ ! ,),#"$'! $+ "',& '))$+ "', *, #+#" *+ "',0 "'**'#& ' /,'"/ #$"/"" ##$"'* *')*$ )) $"*'" #+#"$! +#" *)" #"B$ +B#$/ ,)3 #+ "& '

/! 0 

#"'/ ")'0))!+#" *+ "',0' *' + #0 "'# # " ))" ##$"",)"" ,)3 " "'+'** 3+#" #& ')*/"'  + "', )#"$ #+ "',8&

$  &) '+#" *+ "',   *'$ #+), #"#*1C * "/"*'$/ #"#*1C 8& #""' # " ))" #C >&3""'))" #C  ", #" #* " # " / $ =&*"*',,/,"'*#"))" #C @&))!"'*, #" #)"""'*',,*"$"")="+#"#0#C 7&3""'*',,+#"$"")@C =8>8=8>=60' *''$ #")"$"*'$'0# # +&>&*'( #, )"$"'" ,**'('"' )" #& '/ "(#,)#""'/ " )" #/"'(#$"'*#$)#""'*#$ )" #& '$/+# #"'*',,)#" "'$ #0' *'"')" #/(*'$$&

& +    ,  ) )#""' *',,0##( )#"" #"'"##)" " #$),"" #0 "' ,%)" " #/(#,4=6&#%( %,*' # ),*',,*#" #F +#&*'( )) #"'*',,:*"! " ,#$*' )" #++#$#" #$ *"*#*")" #/ ("/"#)" #0' *' *#":"%$)#$#"&

 +&>&*'$ $ #+/,*',,/"'),'0$ #  & ! 4 

'   ) #"' + "',"' # " ))" # *# "/#$,!+#"$*',,&

#"' )'**')*$ )) $""' "" #,#+/ #//) #+" #+#"$! +#" *)" #&#" !**'*# " # ,3 #+/," #"#"'# # "# +''$& 0 ,)#" ,)"*',"'$ #$" 3 $*!* #+&"' "0,#"#*!" $/ #?"'# +''$"*"#$"',,!-"  ".& ' */"'" " ")3#""'*' )*/,"# #+*"" #3  "$ #)3  ")& '" """'*"'"'3*#"!# 3$"'"'#"'0'" #& '#+"'/"' " " !/* " * ,)"#*& '0'3 $$!#, *" ""'"3! #+"0#"03 4, #,:6 " ))$ #4886& "#+"' $*$ !## "0'#"'*#"" # """'#"' )3 #C"'0  "  #*$ #,,#"&

( #   ) '+#" *3" # $# #+ "', #+#" *)"?*3,"" ##$ *" #&   #"' "$!0$)""'3$ ( ,- G.*',"'"+#"#! / " #4=6& # G#*',,-$#. *#" "" #+/#+"' #"'#+/#"' $" '//' #+"'& ' " #+  #"$ #"'* 3 *',, #"',) " #/"'" #+/ "+# /"$" #++#/,"'* 30 "'"', ) " #"'+# #"'" #+**$ #+""'  $ #"'(&   ',"" #)"'" #+*'#+ ""'*',,& #+ "',0)*  ,"" #)""'"0*$%  '  -.& #"' ,"" #"0(#$,!*'#& /""'"+#/"'*# $$*',, *)#$ #+"#(*'#+$""'"'& :,)*',,4'=&'&=&'=6*,4&=' &'='&=6 /*# $ #+(8#$>"0))$&

"'# +''$"*" #"*'0 $)""'"*'# 1$! 0* #$,"# * 4886

#"' # +''$* " *)"'*,)$/* +#"$&* " **/)" # $/ #$ "0 "'"',: ,,** 3)" #"'"#+" "'* " *)"'#$"'")*$#,,*' #&

/8HH-?*$."'#0)"'/ ""0 )" ##"'"*#$"'""0)" ## "'/ "*& 03 /A80)#!"'""0 )" # #"' *#"'"''#$ /A0)"' / ""0)" #4886&

    '*" #,*'# ,/)$*" # # "' )) $#"'/ "## #+/"' *',,& 0*',,*'#0 "' )  "!))" #""' / "#/*3,#+ " #$ 3 $-,"/0"" # :*$$ /, #+)$*$.& '#$" #+"'0",,/ "'))" #& - .  .  )" # #"')'3"$ **$ #+""' / "#& #"' "$!0"' ,)#3/*'$"'/ "#/#*" #&

+ "',>'0"'"*'+ "',0'3 *# $$'& ' + "',  ,)# #"'/ " ")"' # " " # "#/,"'"+ 3#!"' / ")'& '# " ""3#,/")& #*' "" #"'# +''$/"'*#"" #  " #$#/"'# +' *"$/"'#:" "" #&

7

'"*'/ # '/"/ :$#,/ "" #

"'" #,*'' +'& '# ")""'/0 #+ " #/"' """+ #0 "' "+ #&

$  ') '**'+ "',   "/"*'$,#$), #"#*1C * *'$/ #"#*1C

"/"" # 8&3"*'$,67C >& #""'84# +''$/"'*#"" #C  "' """ #  =&)*"'*#"" #0 "'#$"#"8C @&)$""'" "C  7&)*"'*#"" #0 "'"" #,#+"':, #$C 9.))$! '#$ ',)#4896C

8

/)" #."'"%#0#" #"'"" # " #$! ,)+#" *+ "', #/ 3"",)"- & /"/ ")')* #+#!.#$"'"" # " #$! + "', #/ 3"",)"&



'0"',)#)/,#*"" " */ "'))$ + "',/"'*"$#*', ),*,)" 3!0 "' ,)+#" *+ "', "'"*#" """'/ ")'/"'0' + "',& "  "),#,), 2-#,/(#,



,)"" #"" #$!"'))$+ "',##*',),

#"#*

 2- .

"#0# " #

 89  >9 J7 J< JD 98 9> 9= 9@ 97 9< 9D 9E 95 89 98 9> 9= 9@ 97 9< 9D 9E 95 89 88 8> 8= 8@ 87

-8989. ->97. -8989. -8989. -87>9. -8989. -8989. -8989. -8989. -8989. -8989. -8989. -8989. -8989. -8989. -897. -897. -897. -897. -897. -877. -877. -877. -877. -877. ->97. ->97. ->97. ->97. ->97.

5=9 88=@ 5@= 5> 8>9D

"+#"$ " #0 "'  #7"",)" 5=E 88D= 8>=E 5@@ 8>9D

"+#"$ " #0 "'  # 7"",)" 5=9 88=@ 5@= 8>9D   

,"'*,)"" #"/"'"8 "*$ *#*$$"'""'))$+ "',)$*$+$ " ## #"#*""$& #,"/* ""# )" ," #& #"'/0"'*"'" # 3!#/"')" ,& "*$*#*$$"'""'

+ "',"#"" #""1 " #""'"#$!"' ,)+#" *+ "', "'"*#" """'/ ")'/"'))$+ "',& ' ,#"'""'*#$)' ,)3!"' *')*&

""'"##"/(%')*'$ #+,*' //"'0# #"' ""'/*$#'! $ ,"'$ #*," #+"*'# 1*##"3"'  ),// * #"!4D6& #"' ))0'3*# $$ +#(%')),#$0'3))$+#" * **'+ "',& '))$+ "',*" #"0 ")& "!+#" *+ "',0 "')" #%$ #*$ #+ G*3,"" ##$ " " ""+!*" # )) $"+""/"" #& #

9

4D6  #&&;#&B", # " *(') *'$ #+?")#""B @  %  " ( 99@.))&875L8&

48=6 ##,,& &3 #$#&#$('&N&B *'+ "',/')*'$ #+B?! %! / ! ' "! A&3&8k @ t ... t P>k @ values of the performance measure of interest, the sample means of the X k ,...,X1 performance measure of interest for each system solution, the probability of correct P^CS` selection, the indifference zone chosen by G the experimenter. An IZ procedure is statistically indifferent to which system solution is chosen among the k competing alternatives when all these alternatives fall within a fixed distance G from the best solution. In a maximization problem the probability of performing a correct selection with at least level of confidence P* is

! Pi i z k | Pk  Pi t G ` t P* .

simulation run are independent and normally distributed, one may pursue variance estimation by simply using classical statistics and computing the sample mean

Xi

VAR>X i @ Si2

(2)

k 1

k

(5)





1 n 2 ¦ X ij  X i . n 1 j 1

(6)

Should this not be the case - as customary in a simulation-based study of practically any real-life system - then one must start from the output stochastic process, organize its data and compute the process variance. For example, for system i let X1,...,X j , ,...,X n

f

³ FT t  h fT t dt

1 n ¦ X ij , nj 1

followed by the unbiased sample variance which is used as variance estimator

Under the hypothesis of normality of the statistics involved, this probability was first computed by Rinott in (Rinott 1978) starting from the following inequality

P^CS` t

(4)

are distributed according to Student’s law. The above integral is set equal to P* and solved numerically for h , for different values of n . Numerical values for h , which is also known as Rinott’s constant, are tabled in (Wilcox 1984). In conclusion, when simulating k alternative system solutions, IZ procedures guarantee the selection of the “best” solution or a “near best” according to a pre-specified probability. From a practical point of view, considering a large number of simulation replications for each solution reduces sampling errors; on the other hand, the computational expense of even one single replication of any simulation model is likely to be cumbersome. Bearing in mind these conflictual objectives, pioneering two-stage indifference-zone ranking and selection (R&S) procedures (Rinott 1978, Dudewicz and Dalal 1975) have been followed by more recent and advanced procedures based on an n -stage logic, with n ! 2 (Kim and Nelson 2001, Chen and Kelton 2005). In our SO approach we also exploit an n -stage IZ R&S procedure where the idea of “efficient” sampling is pursued by basing the number of output observations to be taken from each system on the corresponding variance behavior (i.e. how variance changes as the sample from simulation output grows), given a fixed computing budget. Thus, for our enhancement, it is necessary to establish how such variance should be estimated. If for system i ( i i..k ) the n elementary output X i ˆ ^X ij , j 1..n` returned from a observations

P1 , P2 ,...,Pk

P^CS` ˆ P^Pk

X k 1  P>k 1@ X k  P>k @ and Tk 1 ˆ G h G h

(3)

^

t 0

`

be a weekly dependent stationary output process with

where

94

This stated, our procedure uses a variance-weighted decisional mechanism based on the variance estimator described above to guide the sampling activity on the number of additional simulation output observations to be taken from each system. Practically speaking, when process variance decreases this multi-stage procedure is expected to terminate faster than classical two-stage R&S algorithms because of its auto-adaptive control. In every other case, the number of iterations during which the sample variance either remains constant (during the last x runs) or increases is controlled by an upper bound ( UB ) on the number of additional simulation runs to be carried-out which is given by the well-known formula based on Rinott’s constant

mean P X and variance V X2 . This process is said to be weakly dependent if the lag-j covariance

>

@

J j ˆ Cov X i , X i j ,

j

0, r 1, r 2, 

(7)

satisfies J j o 0 as j o f (Billingsley 1995). If one chooses to organize this data in batches of size k , the sample mean for batch i is given by:

1 k

X i (k ) ˆ

ik

¦X j

(8)

j i 1

and according to the Central Limit Theorem

additional runs

D

X i (k ) o Z ( P X , V 2 (k ) / k ), k o f i

(9)

k 1

§ ©

j· k¹

V X2  2 ¦ ¨1  ¸ J j . j 1

Furthermore, the variables in the following set (11)

lim V 2 (k )

>

lim k Var X i (k )

k of

@

V X2

/k

|

F n21 n 1

(13)

ªF2 º E « n 1 » ¬« n  1 ¼»

1

12 13 14 15

(14)

and thus

>

settings for i = 1 to k do for j = 1 to n0 do X ij m take a random sample of n0 from

10 11

Applying the mathematical expectation to the above formula

ª S 2 (n, k ) º E « X2 » ¬« V X / k ¼»

2 3 4

7 8 9

.

16

E k ˜ S X2 (n, k )

@

V X2

(15) 17

where

S X2 (n, k ) ˆ

k n 1

¦ X i (k )  X n, k n

2

(17)

P* , G , n0 , h , x , UB m select procedure

V X2 , i . (12)

By (Hogg and Craig 1978)

S X2 (n, k )



1

5 6

become independent as k o f and k of

G2

Table 1: Our IZ R&S Procedure

(10)

^ X1(k ),...,X i (k ),...,X n (k ) `

2 2 i

The following pseudo-code provides a high-level description of our approach when considering a maximization problem:

where

V 2 (k )

h S

(16)

18

i 1

19

is the estimator of the (output) process variance.

95

each of the k systems end for X i m compute an estimate of the sample mean of the performance index of interest for system i update stopping condition[ n ] end do Ni max n0 , h2 Si2 G 2 m determine the sample size to take from each system if n0 t max i Ni then max i X i m select system with greatest sample mean as best and stop Else For i = 1 to k do while Ni d UB do X ij m take one additional random

sample for system i X i m compute an estimate of the sample mean of the performance index of interest for system i S i2 m compute a run-weighted estimate of the sample variance of the performance index of interest for system i Ni max n0 , h2 Si2 G 2 m determine the new sample size for system i if N i d n0 or Si2 constant in the last x runs then

20 21 22 23

algorithm based on choosing the candidate solution j among m neighboring solutions j1 , j2 ,..., jm of the current solution i . As for solution comparison and selection, the procedure reported in Table 1 is inserted on line 6 of the above schema.

stop sampling for system i end while End for max i X i m select system with greatest sample mean as best

So doing, our approach avoids relying on too much information obtained in just one stage and, at the same time, allows to save on computing budget.

3. APPLICATIONS IN PORT LOGISTICS Container terminal logistics have received great interest in the scientific literature from both the theoretical and practical standpoint (Stahlbock and Voß 2008). The reason for such concern is straightforward if one considers the number and random nature of operational activities carried-out in these facilities: vessel arrival and berthing, resource assignment and scheduling, container transfer and handling, emergency management (e.g. equipment failure, congestion phenomena, weather conditions) and so on. In a maritime container terminal many different company-based rules, regulations and practices can be the grounds of application for the simulation-based optimization framework previously described. Real case studies are given in companion papers (Legato, Mazza and Trunfio 2008; Legato, Mazza and Trunfio 2010). Here we consider the yard and some organizational and operational issues pertaining to its role within the terminal. We then propose to manage the yard activity with respect to policies and equipment employed for container stacking/retrieval by applying the SO approach.

2.2. The Framework The simulation-based optimization framework now proposed in Table 2 serves a double purpose. On one hand, it offers a common ground where to define and compare the different IZ R&S techniques that, in turn, are recalled throughout this work or in companion papers (Legato, Canonaco and Mazza 2009). On the other, it shows how a simulation engine inserted in an optimization algorithm is often the only practical solution method available when dealing with difficultto-solve combinatorial problems, embedded in realistic dynamic logistic processes characterized by several elements of randomness. Table 2: SO Framework for Solution Generation and Evaluation 1 k , n 0 , stopping condition[0] m select procedure settings 2 i * i m set best solution = initial solution 3 while stopping condition[ n ]= false do 4

n = n +1

5

i1 n , i2 n ,..., ik n m at iteration n

6

7

3.1. Problem Description The purpose of a stacking yard in a terminal is to provide storage space for containerized cargo during import, export or transshipment operations. Whether dedicated or shared among different shipping companies, suitably-sized lots of the yard are generally assigned to each company and equipped with technological means in order to enable the stacking/retrieval of container batches (i.e. a set of containers sharing some common properties).

take/generate k alternative solutions i* best^ i* , >i1 n , i2 n ,..., ik n @` m compare the k alternative solutions at iteration n with current best and, eventually, update the best update stopping condition[ n ]

8

end do

9

i * m return best solution

len gth

tier

As one may observe, on line 5 solutions are either taken or generated. In the latter case, a metaheuristic approach based on a variant of the well-known Simulated Annealing (SA) algorithm (Alrefaei and Andradóttir 1999) has been adopted. Besides discarding the basic assumption according to which the temperature Tempk o 0 as k o f by assuming Tempk Temp k , this approach bears two possible ways of estimating the optimum solution. It either uses the most visited solution or selects the solution with the best average estimated value of the objective function. The effectiveness of this constant temperature approach is not yet consolidated for complex and large practical applications. (Mazza 2008) discusses this issue and introduces a guided-search refinement in the SA

bay

lane

Figure 1: Definition of a Yard Block. A yard is typically organized in zones that, in turn, are divided into blocks. As shown in Figure 1, the size of a block is defined by three dimensions: i) number of lanes or rows (e.g. 6 or 13, along with an extra lane if internal trucks are used to perform container transfer);

96

turns (see Figure 3). These movements are exclusively referred to inter-block (and not inter-zone) crane transfer. In our study, we focus on the new operational scenarios generated by five alternative management policies - all known a priori - for assigning yard cranes to yard blocks and accounting for order, times and routes of the crane transfer. The objective is to select, by way of the SO framework, the policy which allows us to minimize the maximum average time to complete stacking/retrieval operations of suitable batches of containers in the yard.

ii) number of container tiers or stack height for each lane (e.g. 5); iii) number of containers in length (e.g. 20). A vertical section of a block (e.g. 5 tiers * 6 lanes) is normally referred to as bay.

Berth Area

Berth Area

Quay Area

Quay Area

container discharge/loading

container discharge/loading

Yard Area container stacking/ retrieval

Yard Area container stacking/ retrieval

(a) (b) Figure 2: Two Alternative Yard Organizations.

3.2. Numerical Experiments To perform the comparison of five alternative system solutions we consider the corresponding variance patterns with respect to a hypothetical operational scenario in which average container traffic in yard blocks is at a medium level (e.g. not many shipping lines stack/retrieve containers in that area) and average crane transfer times between blocks are high (e.g. in an extensive yard area). Figure 4 illustrates an example of how variance changes as the samples taken from system simulation under different policies grows. Observe that for the first three policies variance behavior is stable, meaning that are no significant changes in variance estimation as the sampling procedure progresses. Thus the algorithm continues adding single observations (or batches or simulation replications) as required by the “stable” variance estimate until the upper bound provided by Rinott’s two-stage procedure is reached (Legato and Mazza 2008). When the variance pattern increases, as for policy n°4, the upper bound is still provided by Rinott’s procedure.

It is worth observing that both the number and size of blocks in a yard affect the average travel time of shuttle vehicles cycling between the quay and the yard areas, as well as the container handling time on the yard. For instance, in the yard organization depicted by Figure 2.(a), the average distance to be covered in order to reach a container is greater than the average distance deriving from the solution portrayed in Figure 2.(b). On the other hand, more container handling equipment can be concentrated in a specific area in the former case, thus returning a smaller service time, whereas this possibility is prevented in the latter case due to potential interference between container movers meant to operate on adjacent yard bays If container stacking/retrieval on the yard is performed by transfer cranes, such as rail-mounted gantry cranes (RMGCs) or rubber-tired gantry cranes (RTGCs), then a common operational issue actually consists in periodically deciding how many and which cranes are to be assigned to a block. This decision usually depends on the expected daily workload in each block and, therefore, on the total crane capacity (measured in time units) required to complete container stacking/retrieval operations.

Variance Behavior

Variance 60 50 40 30

yard block

20 10

yard crane

0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39

90° turn

Policy 1

Policy 2

Policy 3

Policy 4

Observations

Policy 5

Figure 4. Sample Paths of Variance Behavior Instead, in policy n°5 the variance estimate has a decreasing trend and, thus, the algorithm is expected to terminate faster. This expectation is justified by the auto-adaptive control of the procedure which can be monitored according to a step-by-step logic. In this sense, Table 3 provides a trace of the variance behavior for policy n°5. As one may observe, after setting n0 10 , P* 0.90 , G 5 and, thus, h 3.137 , according to Rinott’s procedure the number of runs to consider for system i are

Figure 3: Possible Intra-block Crane Transfer. To do so, cranes must be transferred from one block to another. If we consider RTGCs, these cranes can travel between adjacent yard blocks without any turning motion or by changing lanes. In the former case, crane transfer can take about 10 minutes; in the latter, about 5 additional minutes are required to perform 90 degree

97

Ni



max n0 , h 2 Si2 G 2





max 10, 3.137 2 *120.30 52



48

estimate of the skewness of the sample mean distribution, given that the normality assumption is approximately verified only after a large number of simulation runs - a condition one should avoid, due to the computational burden it is bound to bear.

(18)

So, Ni  n0 38 additional runs must be added to guarantee the predefined probability of correct selection P* 0.90 . Alternatively, as shown in Table 3, our procedure after only one supplementary run at step 11, returns

Ni



max 11, 3.1372 *108.30 52



43

REFERENCES Alrefaei, M.H. and Andradóttir, S., 1999. A simulated annealing algorithm with constant temperature for discrete stochastic optimization. Management Science 45 (5), 748-764. Billingsley, P., 1995. Probability and Measure. Third Edition. John Wiley & Sons, Inc. Chen, E.J. and Kelton, W.D., 2005. Sequential selection procedures: using sample means to improve efficiency. European Journal of Operational Research 166, 133-153. Dudewicz, E.J. and Dalal, S.R., 1975. Allocation of observations in ranking and selection with unequal variances. Sankhya B7, 28-78. Fu, M.C., 2001. Simulation optimization. In: Proceedings of the 2001 Winter Simulation Conference, Peters, B.A., Smith, J.S., Medeiros, D.J., and Rohrer, M.W., Eds, pp. 53-61. December 9-12, Arlington (Virginia, USA). Fu, M. and Nelson, B., 2003. Guest Editorial. ACM Transactions on Modeling and Computer Simulation 13(2), 105–107. Hogg, R.V. and Craig, A.T., 1978. Introduction to mathematical statistics. Fourth Edition. Macmillan Publishing Co., Inc., New York. Kim, S.-H. and Nelson, B.L., 2001. A fully sequential procedure for indifference-zone selection in simulation. ACM TOMACS 11, 251-273. Legato, P., Canonaco, P. and Mazza, R.M., 2009. Yard crane management by simulation and optimization. Maritime Economics and Logistics 11(1), 36-57. Legato, P. and Mazza, R.M., 2008. Selecting the Optimum by Searching and Ranking Procedures in Simulation-based Optimization. In: Proceedings of the 20th European Modeling and Simulation Symposium (Simulation in Industry), pp. 561-568. September 17-19, Campora S. Giovanni, (CS) Italy. Legato, P., Mazza, R.M. and Trunfio, R., 2008. Simulation-based optimization for the quay crane scheduling problem. In: Proceedings of the 2008 Winter Simulation Conference, Mason, S.J., Hill, R., Moench, L., and Rose, O., Eds, 2717-2725. December 7-10, 2008. Miami (Florida, USA). Legato, P., Mazza, R.M. and Trunfio, R., 2010. Simulation-based Optimization for discharge/loading operations at a maritime container terminal. OR Spectrum 32 (3), 543-567. Mazza, R.M., 2008. Simulation-based optimization in port logistics. Thesis (Ph.D). Università della Calabria.

(19)

meaning 32 additional runs (i.e. 43 – 11 previous runs). It, thus, realizes a gain of 6 runs after one single run. Table 3: Step-by-step Trace of Variance Behavior for Policy n°5 N° of observations for policy i=5 Step Ni Sample mean Sample variance 92.34 120.30 10 48 92.39 108.30 11 43 92.96 102.34 12 41 92.48 96.85 13 39 92.20 90.49 14 36 …







In numerical terms, given that both procedures choose policy n°3 as best, in the worst case our procedure returns the same results as Rinott’s two-stage procedure ( ' =0%), while for decreasing variance behavior our procedure is more efficient by 31,25%, as illustrated in Table 4. Table 4: Comparison of Observations Required by Rinott’s Procedure (RP) and Our R&S procedure N° of Our observations Performance Alternatives RP Ours '%) (' 31 31 0% policy 1 27 27 0% policy 2 9 9 0% policy 3 32 32 0% policy 4 48 33 +31.25% Policy 5

4. CONCLUSIONS An n -stage indifference-zone based ranking and selection procedure has been proposed to “hopefully” deliver more efficient sampling than classical two-stage algorithms. Its performance has been tested by some numerical experiments. Rather than just using a classical sample mean, it appears that tracking the variance behavior reveals improvement margins when the variance pattern is decreasing. In the future, a further possibility may lie in investigating how to use an

98

Rinott, Y., 1978. On two-stage selection procedures and related probability-inequalities. Communications in Statistics - Theory and Methods A7(8), 799-811. Stahlbock, R. and Voß, S., 2008. Operations research at container terminals: a literature update. OR Spectrum 30(1), 1-52. Wilcox, R.R., 1984. A table for Rinott’s selection procedure. Journal of Quality Technology 16(2), 97-100. AUTHORS BIOGRAPHY Pasquale LEGATO is an Associate Professor of Operations Research at the Faculty of Engineering (Università della Calabria – Rende, Italia), where he teaches courses on simulation for system performance evaluation. He has published on queuing network models for job shop and logistic systems, as well as on integer programming models. He has been involved in several national and international applied research projects and is serving as reviewer for some international journals. His current research activities focus on the development and analysis of queuing network models for logistic systems, discrete-event simulation and the integration of simulation output analysis techniques with combinatorial optimization algorithms for real life applications in Transportation and Logistics. His home-page is . Rina Mary MAZZA went to the Università della Calabria, Rende (Italia), where she received her Laurea degree in Management Engineering and a Ph.D. degree in Operations Research. She is currently Head of the Research Project Office at the Dipartimento di Elettronica, Informatica e Sistemistica (DEIS, Università della Calabria). She is also a consultant for operations modeling and simulation in terminal containers. Her current research interests include discrete-event simulation and optimum-seeking by simulation in complex logistic systems. Her e-mail address is: .

99

ACCELERATED FULLY 3D ITERATIVE RECONSTRUCTION IN SPECT Werner Backfrieder(a), Gerald A. Zwettler(b) (a), (b)

University of Applied Sciences Upper Austria, Institute of Software Engineering, Campus Hagenberg, Austria (a)

[email protected], (b)[email protected]

regions. In tumor diagnosis tissue pathologies are imaged as hot spots. The amount of activity uptake and the size of the lesion are an important measure for the progress of the tumor disease. Both, distribution and kinetics of the radiotracer are subject to functional imaging, e.g. perfusion images of the human brain after stroke or assessment of the clearance rate in kidneys. Single photon emission computed tomography (SPECT) is a volume imaging technique, visualizing the human body as a series of transversal slices. The photons generated during disintegration of a short lived radionuclide, e.g. Tc-99m, are registered by a gammacamera as projection images. There is a great variety of algorithms for the reconstruction transversal slices from projection data. Filtered back-projection (FBP) in combination with specific filter windows is, due to its high performance, the main method in clinical practice (Herman 2009). With increasing computational power iterative methods, allowing a more accurate modeling of geometrical and physical properties of the imaging process, were introduced into clinical environments. The maximum likelihood expectation maximization (ML-EM) algorithm (Shepp and Vardi 1982) is the foundation of a series of optimized algorithms in emission tomography (Hudson and Larkin 1994). The slice topology of the reconstruction algorithms is a major limitation in image quality of emission tomography. In contrast to x-ray computed tomography the scanner hardware allows major interferences from adjacent slices. Collimator geometry defines a conic volume of response to a single detector position; low count rate and scatter are major deteriorating effects in SPECT imaging. Fully 3D image reconstruction accomplishes the simultaneous reconstruction of the whole image volume, but at the cost of high computational burden (Backfrieder et al. 2002, Backfrieder et al. 2003a, Backfrieder et al. 2003b, Benkner et al. 2004). In the following a fully 3D iterative reconstruction algorithm is described implementing a dual projector back-projector pair for accelerated reconstruction. The newly developed algorithm is based on the OS-EM family providing accelerated convergence (Hudson and Larkin 1994).

ABSTRACT Image quality in single photon emission computed tomography (SPECT) is substantially influenced by scatter and a finite volume of response associated with single detector elements. These effects are not restricted to the image plane, implying a shift in the tomographic imaging paradigm from 2D to 3D. The application of a 3D reconstruction model suffers from huge numerical efforts, affording for high performance computing hardware. A novel accelerated 3D ML-EM type reconstruction algorithm is developed by the implementation of a dual projector back-projector pair. An accurate 3D model of data acquisition is developed considering scatter and exact scanner geometry in opposite to a simple pencil-beam back-projection operator. This dual concept of projection and backprojection substantially accelerates the reconstruction process. Speed-up factors achieved by the novel algorithm are measured for several matrix sizes and collimator types. Accuracy of the accelerated reconstruction algorithm is shown by reconstruction of data from a physical Jaszczak phantom and a clinical endocrine study. In both cases the accelerated 3D reconstruction method achieves better results. The novel algorithm has a great potential to scale fully 3D reconstruction down to desktop applications, especially with the new possibilities employing massive parallel graphics hardware. The presented work is a step towards establishing sophisticated 3D reconstruction in a clinical workflow. Keywords: emission tomography, fully 3D reconstruction, nuclear medicine, high-performance computing 1. INTRODUCTION Nuclear medicine imaging modalities show the distribution of radioactive tracer providing diagnostic information. Main fields of application are tumor diagnostics and in vivo assessment of metabolism. Therapeutic applications are limited to therapy with beta-emitters, e.g. radioiodine therapy of the thyroid. In nuclear medicine imaging the kinetics of radioactive tracer particles within the human body is the basis of diagnostic information. After intravenous application specialized radiopharmaceuticals distribute within the body and finally accumulate in targeted morphological

2. MATERIAL AND METHODS The imaging equation in tomography reads

100

¦a

yi

ij

˜ x j  ei ,

A sequence of intermediate images x(n) is calculated until a stopping criterion is satisfied. During the nth iteration each pixel xj is updated by a multiplicative factor. This factor is the weighted sum of all projection values yi affected by the pixel xi. The correction term depends on the quotient of the measured projection value and the calculated pseudo-projection

(1)

j

it describes the relation between the pixels of the source distribution (=image) x and a single projection values y. Both, the image and projection array are twodimensional, i.e. x-and y-direction in the image, angle and lateral distance in projections, but are represented by a single linear index. A value of the system matrix aij describes the contribution of pixel xj to the projection value yi. This allows the accurate modeling of x x x x

yip

Each measured value contains an error term ei. In the case of radioactive decay and detection of photons this error term is Poisson-distributed. Under the constraint of a Poisson-distributed random process image reconstruction is formulated as a maximization problem of the likelihood L of measured data y

L( y | x)

–

˜ xj

e yi

 ¦ j aij ˜ x j

yi !

i

,

(2)

where the sum represents the expectation value of the respective measured projection value yi- The algorithm aims in maximizing the term in Eqn. 2 by choosing a proper image-vector x. The solution is the iterative MLEM algorithm for tomography by Shepp and Vardi, 1982

x (jn 1)

x (jn ) ¦ aij i

yi ¦ aij ' x (jn' ) j'

.

(4)

2.1. Modeling of the system matrix Each line of the system matrix defines the weights of all voxels to a specific projection value. In a conventional SPECT study the image volume consists of 128 slices, with a matrix size of 128x128 pixels, each. A row consists of 1283=2.097.152 elements. The number of projection values, i.e. the number of lines of the matrix, is calculated from the size of the projection matrix and the angular increment of the detector head, i.e. 128x128x120 for a 3 degrees increment on a circular orbit. In total the system matrix contains 4.12x1012 elements. Even dedicated high-performance-computer (HPC) systems cannot hold this huge amount of data in memory. Since a line of the system matrix considers all elements of the image volume, most of the entries are zero. With careful modeling of the geometrical and physical properties of data acquisition, this leads to a significant reduction of data. Each projection value is related to a flat rectangular region of the detector surface, i.e. the field of view (FOV) divided by the number of elements of the projection matrix. For assessment of the contribution of each voxel to a specific projection value, a point source is positioned at the center of a voxel and the fraction of radiation reaching the detector element is calculated. This corresponds to the ratio of the surface of a sphere, with origin in the voxel and the radius is the distance to the detector element, and the projection of this detector element onto this sphere. This simple geometrical

Figure 1: Sketch of SPECT scanner geometry.

ij

x (jn' ) .

The iteration steps in Eqn. 3 converge to a feasible solution, representing a maximum entropy solution to the imaging equation. To further accelerate the convergence of the algorithm ordered subsets are implemented. The discussed reconstruction model describes the reconstruction of a single image slice. Spatial activity distribution out of the slice is not considered by this model. Finite collimator aperture and scatter have significant contribution from pixels out of the considered slice on the projection values, necessitating a three dimensional (3D) approach to the reconstruction problem for further improvement of image quality. In contrast to FBP the iterative approach allows a simple extension to 3D by covering the whole image volume and projection values of all slices by respective vectors. As a consequence the system matrix A grows o(N6) with the lateral length (N pixels) of the image cube.

Figure 1 shows a sketch of the image plane and the pixels summing up a single projection value.

j

ij '

j'

scanner geometry photon attenuation detector response scatter

¦ a

¦a

(3)

101

consideration leads to a model of the volume of response as a cone targeting to the detector surface. With increasing distance to the detector the cone-width increases and the weight of voxels decreases. The voxel weights at the level of the central slice of the volume of response (VOR), i.e. at the level of the projection value, are shown in Fig. 2.a. The VOR has circular symmetry.

numerical operations is proportional to the non-zero elements of the system-matrix A. To achieve most accurate physical and geometrical modeling the forward projection is implemented by the modeled weights according to Eqn. 4. The accurate assessment of pseudo-projections is crucial, since its ratio to measured projection values yi defines the amount of the correction term. The back-projection operator comprises the projection values considered for the update of a specific pixel. In this novel approach not all elements, as defined by the above model of the system matrix, are considered, but only a subset defined by orthogonal projection onto the detector surface. The lateral distance from the center of the profile is

l

x (jn1)

The camera head is equipped with a lead collimator, to limit the viewing direction approximately to bars normal to the detector surface. The design consists of a regular pattern of thin lead septa, arranged as long thin bore holes or as a honeycomb grid. The modulation factor is the cast shadow of the collimator septa depending on the detector thickness and the ratio of wall thickness of septa and their aperture. Scatter is a further amplification of the voxel weights; usually it is modeled by a zero centered Gaussian distribution. The total pixel weight reads

4 scatterD coll ) geom

,

(6)

where x, y are the coordinates of the updated voxel and - is the rotation angle of the detector head. Only projection values within the slice are considered. The ML-EM algorithm with dual projector and backprojector pair reads

Figure 2: Volume of response of a 128x128 projection matrix (a) and its amplification by a LEGP collimator (b).

ai j

x ˜ cos -  y ˜ sin - ,

x (jn ) ¦ lij i

yi ¦ aij ' x (jn' ) j'

.

(7)

The coefficients lij denote the reduced set of backprojection values. The speed up factor is linear to the reduction of the lij coefficients in relation to the total number of entries in the system matrix entries aij. For a standard 128x128 matrix and a LEGP parallel collimator the speed-up factor is 218.53. This speed up of fully 3D reconstruction implemented together with the ordered subsets concept, the newly developed algorithm is called 3D accelerated ordered subsets expectation maximization (3D-AOS-EM).

(5)

2.3. Physical phantom and patient data Data are collected from a circular clinical standard Jaszczak SPECT phantom on a three headed Philips IRIX camera. The phantom was filled with 600 MBq Tc-99m. Acquisition parameters were: 128 by 128 projection matrix, pixel size 4.4mm, 120 projections on a full circular orbit of 360 degrees and 20s acquisition times in stop and go mode. On the same camera data from a clinical endocrine study, 55MBq I-131 applied activity, were acquired on a 64 by 64 projection matrix over a 565mm FOV, with 60 projections on a circular orbit, and 30s acquisition time per projection.

where the geometrical form factor is ), the attenuation factor of the collimator is D and the contribution of photon scatter by human tissue is 4. Figure 2.b shows the application of the collimation factor to the VOR. 2.2. Dual projector back-projector pair In the previous section the OS-EM algorithm and the modeling of the system matrix is discussed in detail. With the generalization of the reconstruction problem to 3D the computational effort increases substantially, affording for HPC hardware to achieve suitable performance for image reconstruction, to establish it in a clinical environment. The high cost of the ML-EM algorithm is caused by a series of projections and back-projections during each iteration step, cf. Eqn. 3. The sum over all projection values containing the actual pixel can be considered as back-projection. From each intermediate image x(n) pseudo-projections are calculated. The number of

3. RESULTS Results are shown for acceleration of the algorithm in contrast to 3D ML-EM, a comparison of reconstruction methods applied to physical phantom data and a clinical study.

102

signal intensity, manifested by substantial star-artifacts centered at the lesion. The hot lesion is a connected oval region with a small tail at the lower left. This image data cannot clearly support the decision, if this tiny image structure is a real pathology or an artifact. 2DML-EM reconstruction shows a clearly manifested hot lesion in this part of the image. Reconstruction of the image using the newly developed 3D-AOS-EM yields two clearly distinguishable hot lesions.

3.1. Speed up factors The speed-up factors - as a consequence of the implementation of the accelerated back-projection operator - are shown in Figure 3. Results are shown for two collimator types, a low energy general purpose (LEGP) and a high energy high resolution (HEHR) collimator. The speed up factor directly relates to the reduction of entries in the back-projection matrix compared to those in the respective projection matrix, as shown by different factors for the collimators used during the studies. The HEHR collimator has a significantly smaller VOR thus the speed-ups are smaller than those of the LEGP collimator, cf. Fig. 3.

4. DISCUSSION Fully 3D image reconstruction is the most accurate reconstruction model for nuclear medicine emission tomography. The direct implementation of the 3D data model suffers from high computational complexity resulting in long reconstruction cycles, hardly to establish in a clinical workflow. The substantial acceleration of the algorithm by introduction of a dual projector back-projector pair has high potential to scale down the problem from HPC platforms, as already implemented on PC-clusters (Backfrieder et al. 2003b), to desktop hardware. The actual algorithm is implemented as a MATLAB prototype, thus the evaluation of the performance is done on basis of speedup factors. The newly introduced programming interface CUDA to the highly parallel architecture of the graphics-subsystem offers new perspectives to solve computationally intensive numerical problems. In ongoing work the 3D-AOS-EM algorithm will be implemented in the C-CUDA framework.

Figure 3: Speed-up factors 3.2. Phantom data A slice of the Jaszczak phantom comprising 6 sectors of cold rods with increasing diameter is shown in Fig. 4. Slices were reconstructed using FBP, the clinical standard, and the accelerated fully 3D reconstruction with dual projection back-projection (3D-AOS-EM). During iterative reconstruction 15 iterations with 4 subsets were performed. Compared to FBP the contrast of cold spots is significantly increased with 3D-AOSEM. In sector 4 (numbered in order of decreasing diameter) rods are still distinguishable, especially in the distal part of the phantom, since with FBP the whole sector is blurred out.

Figure 5: Endocrine study reconstructed FBP (a), 2DML-EM (b) and 3D-AOS-EM. Figure 4: Reconstruction of a standard Jaszczak phantom using (a) FBP and the novel 3D-AOS-EM algorithm.

The acceleration of the fully 3D reconstruction together with its implementation on desktop systems is a further step towards sophisticated image processing supporting clinical diagnostics.

3.3. Clinical data Data from the clinical study show a transversal slice through the thyroid, cf. Fig. 5. FBP suffers from low

103

reality techniques in the context of surgical planning and navigation. Gerald A. Zwettler was born in Wels, Austria and attended the Upper Austrian University of Applied Sciences, Campus Hagenberg where he studied software engineering for medicine and graduated Dipl.Ing.(FH) in 2005 and the follow up master studies in software engineering in 2009. In 2010 he started his PhD studies at the University of Vienna at the Faculty of Computer Sciences. Since 2005 he is working as research and teaching assistant at the Upper Austrian University of Applied Sciences at the school of informatics, communications and media at the Campus Hagenberg in the field of medical image analysis and software engineering with focus on computer-based diagnostics support and medical applications.

ACKNOWLEDGMENTS Authors want to thank Univ.-Prof. Dr. Michael Gabriel and the radiological technologist from the Institute of Endocrinology and Nuclear Medicine of the General Hospital Linz, Austria, for providing SPECT data.

REFERENCES Backfrieder, W., Forster, M., Benkner, S., Engelbrecht, G., Terziev, N., Dimitrov, A., 2002. Accurate attenuation correction for a fully 3D reconstruction service. Proceedings of the International Conference on Mathematics and Engineering Techniques in Medicine and Biological Sciences (METMBS '02), Las Vegas, Nevada, USA, June 24 – 27, 2002: 680-685 Backfrieder, W., Forster, M., John, P., Engelbrecht, G., Benkner, S., 2003a. Fully 3D Iterative SPECT Reconstruction in A High Performance Computing Environment. J. Nucl. Med. Technol. Vol.31 No.6, pp. 201 Backfrieder, W., Forster, M., Engelbrecht, G., Benkner, S., 2003b. Locally variant VOR in fully 3D SPECT within a service oriented environment. In F. Valafar, H. Valafar (Eds.), Proc. Int. Conf. on Mathematics and Engineering Techniques in Medical and Biological Sciences (METMBS), ISBN 1-932415-04-1, (2003) pp. 216-221 Benkner, S., Engelbrecht, G., Backfrieder,W., Berti, G., Fingberg,J., Kohring,G., Schmidt, J.G., Middleton,S.E., Jones, D., Fenner, J., 2004. Numerical Simulation for eHealth: Grid-enabled Medical Simulation Services. In G.R. Joubert, W.E. Nagel, F.J. Peters, W.V. Walter (Eds.), Software Technology, Algorithms, Architectures and Applications, included in series: Advances in Parallel Computing, Elsevier (2004) Herman, G. T., 2009. Fundamentals of computerized tomography: Image reconstruction from projections, 2nd edition, Springer, 2009 Hudson, H.M., Larkin, R.S., 1994. Accelerated image reconstruction using ordered subsets of projection data. IEEE Trans Med Imaging. 1994;13:601– 609. Shepp, L.A., Vardi, Y., 1982. Maximum likelihood estimation for emission tomography. IEEE Trans Med Imag. 1982;MI-1(2):113–121. AUTHORS BIOGRAPHY Werner Backfrieder received his degree in technical physics at the Vienna University of Technology in 1992. Then he was with the Department of Biomedical Engineering and Physics of the Medical University of Vienna, where he reached a tenure position in 2002. Since 2002 he is with the University of Applied Sciences Upper Austria at the division of Biomedical Informatics. His research focus is on Medical Physics and Medical Image Processing in Nuclear Medicine and Radiology with emphasis to high performance computing. Recently research efforts are laid on virtual

104

STUDY ON THE SERVILIZATION OF SIMULATION CAPABILITY Y L Luo(a), L Zhang(a), F Tao(a),Y Bao(a) ,L Ren(a) (a)

School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191, P. R. China [email protected], [email protected], [email protected], [email protected], [email protected]

however. Because of MC is a complex concept and its correlative research is less, as a result there is no clear definition currently. At present, there are two different understandings on the concept and connotation of MC, some people argue that MC reflects the performance of enterprise from a macro points of view, i.e., Skinner first proposed the MC in 1969, he holds that MC includes many elements such as cost, delivery time, quality, and the relationship between these elements. MC reflects the completion of manufacturing objective, and it is a performance level of the standard which is pre-sat by working organization (Mattias Hallgren. 2006). Guan (2004) commented that MC is the core part of enterprise innovation capability, it is conversion capability of results which meet market demand, design requirements of product and mass-produced. The relationship between MC and enterprise performance is discussed from the perspective of achieving low operating costs and high product quality (Siri Terjesen, Pankanj C. Patel et al. 2011). The other comment that MC is a integration of manufacturing resources based on microscopic pint of view, i.e., Richard (1973) considered that capability includes knowledge, skills, and experience of enterprise. MC reflects the performance of completing setting function based on manufacturing resources in order to support the operation of enterprise activities (Cheng Yun, Yan Junqi. 1996) Keen(2000) commented that MC is the integration of intangible resources and tangible resources, where the tangible resources include labor, capital, facilities and equipment, simultaneously, and the intangible resources include information, procedures, equipment and the organizational system. Khalid(2002) concluded that MC is the effective integration of related resources in the process of achieving expected target task. Cheng (2009) gave the definition that MC is a set of elements involved in the implement process of manufacturing enterprise’s strategy. Zhang Lin (2010) commneted that MC is an intangible and dynamic resource in CMfg model, it is the subjective condition of production –related goals. Combined with above views about MC, several problems are systematically summarized as follows: z The current research about MC have been widely studied from a management point of view, because mostly based on qualitative

ABSTRACT Manufacturing capability (MC) servilization is a key to realize on-demand use, dynamic collaborative work, and circulation of manufacturing resources and capability in the cloud manufacturing (CMfg) system. This paper emphasizes the servilization of simulation capability (SC), which is a very important kind of MC. According to task demands and characteristics of complex product’s simulation process, concepts and state of the art related to MC were systematically analyzed and summarized firstly in this paper, then a conceptual model of SC were presented, A application model of SC service life-cycle in CMfg system is proposed. Then the framework for simulation capability servilization is investigated, as well as several key issues involved in the servilization process such as elements of simulation capability, modeling and decriptiion of simulation capability, and so on. Finally, an application example analysis of SC was presented. Keywords: cloud manufacturing, manufacturing capability, simulation capability, servilization 1. INTRODUCTION Cloud manufacturing (CMfg) is a new service-oriented, highly efficient, lowly consumption knowledge based, and intelligent networked manufacturing model (Bohu Li and Lin Zhang et al. 2010). It is combined with advanced manufacturing and information technologies organically (e.g. cloud computing, the internet of things, semantic web, and information system integration)in order to achieve virtualization and servilization of manufacturing resources and capability, CMfg provides users with application services which are on-demand using, safe and reliable in the whole life-cycle of products through network (Bohu Li, Lin Zhang et al. 2010). CMfg aims to achieve agile, service-oriented, green and intelligent manufacturing, is a new phase of networked manufacturing, and is the materialization of service-oriented manufacturing (Lin Zhang and Yongliang Luo et al. 2011). Therefore, CMfg can provide theoretical and technical supports for the transformation from production-oriented manufacturing to service-oriented manufacturing. Manufacturing capability (MC) servilization is one of the most important innovative points of CMfg,

105

analysis, lack of supports for quantitative description on MC; z Lack of the logical relationship analysis between construction elements of MC In this paper, is considered as a subjective condition, what manufacturing enterprises needed to complete one task or objective. It is a intangible and dynamic resources form. And it is a kind of capability which can be represented in the manufacturing activities. MC including design capability (DC), simulation capability (SC), product capability (PC), and many other capabilities related to life-cycles of complex products. MC is tightly linked to manufacturing activities and manufacturing resources, it can’t be reflected without concrete activity tasks and resources elements. According to task demands and resources characteristics of complex products’ simulation process, this paper studies MC based on simulation. In order to comprehensively understand SC, concepts and state of the art related to MC were systematically analyzed and summarized firstly, then a conceptual model of SC was presented. An application model of SC service lifecycle in CMfg system is proposed, and the key technologies involved in each process were investigated. On this basis, a new simulation capability servilization (SCS) framework is presented, several issues related to SCS are discussed in detail.

simulation equipment and so on. Resources are the foundation of forming SC; it is the subject of SC servilization as well. Resources can be divided into two kinds, one is subject resources, which is the carrier of SC performance, for example, the subject resources of software’s SC is the simulation software. The other is auxiliary resources such as material, which is support for product and related goal. In addition, a SC possibly refer to several subject resources 2) Process dimension(P) SC is a kind of activity process; it contains a knowledge set generated in the implementation process of task and goal realization, such as constraint condition, simulation method, and simulation experience and so on. In addition, knowledge is the effective carrier of procedure representation. 3) Task dimension(T) It contains two aspects information, one part is about the simulation task, and the other part is about the completion of the simulation task target, which include many objective factors and evaluation of user satisfaction factors, elements, i.e., delivery time, cost, quality, innovation, service, et al. This dimension is the most important selection basis for SC users in cloud manufacturing service platform. The relationship among resources dimension, process dimension and task dimension is investigated as follows, resources is the basic of achieving SC. Task dimension shows us the result of SC, it is the most important basis for user optimize selecting in CMfg system. Process dimension is the method of SC’s forming.

2.

THE CONCEPT AND CONNOTATION OF SIMULATION CAPABILITY SC is an important kind of MC in CMfg system, is a simulation process, which reflects a capability of complete a simulation task or experiment supported by related resources and knowledge. Through SCS, it can not only realize the function sharing of resources, but also share the experience and knowledge in the simulation process, such as simulation flow, simulation data, experience of simulated staff, and so on.

3.

SIMULATION CAPABILITY SERVICE LIFE CYCLE MODEL As shown in figure 1, the life cycle of simulation capability service can be divided into the following four parts: (1) Simulation capability publication (SCP) SCP is a servilization process of SC; it is the basic of cloud manufacturing service platform to realize ondemand use and sharing of simulation capability. It combines the characteristics of simulation resources with simulation capability classification in the CMfg model. Elements of SC are extracted and analyzed firstly, and the unified semantic description model of SC is presented. Secondly, in order to achieve the formal description of SC, the existing services description language will be expanded or improved, then SC will be released in the form of service in CMfg system. It will support the trading and distribution of SC for users through network. Some key technologies involved in this process, such as simulation capability classification, simulation capability modeling, manufacturing capability description language expansion and so on.

Task Quality

Flexible

Time 

Cost 

Product

Software  Hardware  Human

Resources

Requirement

Simulation capability

Experience  Method  Flow

Process

Fig.1 The conceptual model of simulation capability

The conceptual model of simulation capability is illustrated in fig1, which primarily consists of three dimensions: 1) Resources dimension(R) SC is the integration of all kinds of simulation resources related to perform some tasks or activities, such as simulation software,

106

then construct the ontology of SC to support semantic search. The ontology can improve searching accuracy. At last SCD can support the sharing of SC through the network. The process of SCD includes several key technologies, such as domain ontology construction, semantic matching and dynamic composition of SC, services of SC sharing and so on.

(2) Simulation capability discovery (SCD) SCD is responsible for achieving semantic searching and dynamic composition of SC services in CMfg system. According to the characteristics of SC, such as relative, complexity, dynamic and so on, in order to reflects multi-dimensional attributes of SC fully and clearly, all kinds of SC description information should be classified, fused and normalized firstly, and

Capability publication

Capability Capability modeling Capability classification description

...

Evolution mechanism

Simulation capability

Evolution driver

...

data acquisition

Assessment rule

...

Evolution model

...

Capability evolution

Ontology construction Semantic matching

Capability discovery

Capability using Assessment system

Capability assessment Fig.2 The life cycle of simulation capability service (3) Simulation capability assessment (SCA) changing external environment. On the basis of Combined user feedback with operation of the SC, enterprise evolution theory and dynamic capability SCA will realize the comprehensive utility evaluation of theory, driver attributes of simulation capability simulation capability. Due to the complexity and evolution is systematically analyzed in the CMfg mode dynamic of SC, how to measure SC is an important firstly, combined with assessment index system of SC, issue of SCS, but in order to achieve the assessment of form, process and mechanism of simulation capability SC, we need to provide comprehensive assessment evolution are deeply discussed from the qualitative system establishment and suitable assessment methods point of view, and then an empirical analysis on the based on description system and evaluation factors of process of SCE is done by mapping the qualitative to simulation capability, where the assessment system quantitative. SCE will provide support for dynamic should be level and systemic, it can reflect the overall maintenance and intelligent update of SC to CMfg characteristic of simulation capability , and all level and system. Key technologies of SCE include evolution dimensions attributes which contains quantitative and driving factors, evolution mechanism and method of qualitative. The assessment methods of SC should take simulation capability, evolution model construction, full account of dynamic changes during the process of evolution procedural knowledge representation and so the SC’s using. Some key technologies involved in this on. process, such as capability assessment system construction, SC assessment of data acquisition, SC 4. THE FRAMEWORK OF SIMULATION assessment method, operation monitoring of SC and so CAPABILITY SERVILIZATION on. SCS plays an important role in the process of achieving (4) Simulation capability evolution (SCE) on-demand use of SC. The process of SCS is shown in It is a response and adjustment process of fig3, it can be divided into the following five parts: manufacturing enterprises or systems in the face of

107

The center of simulation capability service

Simulation capability service application simulation capability registry

capability Semantic matching

simulation capability service using

simulation capability evolution

algorithm library

Simulation capability description description language

grammar detection static attribute

modeling base

logical reasoning dynamic attribute

Simulation capability modeling information extraction

information classification

information fusion

procedural representatin

model foundation

template setting

Simulation capability elements Process

Task

Software



Process

Task



Simulation capability

ontology library

Simulation capability

Participant

Hardware



Process

Task Simulation capability

 Participant

Model



Participant

Fig.3 The framework of simulation capability servilization to users should be analyzed and classified firstly. Then Simulation capability elements Most of the current researches about capability according to the actual requirements related to product elements refer to the performance of manufacturing and goals, elements of SC will be summarized. For capability. SC elements mainly contain various example, SC elements can be divided into six parts ,as resources and related assessment factors in the shown in fig4: major resources, product and business, construction process of SC, it is a comprehensive and participator, process knowledge, SC assessment integrated reflection of SC, and will provide support for information, enterprise integrated information. And construction and formal description of SC model. In each part is also consisted by many related elements in order to achieve on-demand use and dynamic detail, e.g., process knowledge contains design model, collaborative of SC, the display content of SC oriented experience knowledge, simulation method and so on.

¾

organizational culture

Simulation Capability (SC)

enterprise comprehensive information

creative ability  environment effect

major resources

simulation simulation software  equipment

computing resourcess

product &business

simulation  business

participator

simulation product

knowledge  business and skill experiences

process knowlege

design model

 experience knowledge

simulation method

Fig.4 Elements of simulation capability representation, SC description template construction Simulation capability modeling According to the characteristics and connotation of and so on. SC, combined with the above introduction about SC The model can be divided into two parts as follows: elements, description model of simulation capability DMSC=(T, R, P, K,E˅+ fun(T, P, R, K,E) (DMSC) is abstractly represented. The transformation In above description model, the first part is a fourfrom qualitative to quantitative elements of SC is quad, where T is the simulation task and the objective realized. And related key technologies involved in the needed to realize a simulation activity. R is simulation model as shown in fig4, such as information extraction, resources elements involved in the simulation process, information classification, information fusion, process i.e., hard-resource (including simulation equipments,

¾

108

materials) software, et al). P is the participants involved in the simulation process, i.e., the human resource in the simulation resources elements including personal and organization. K mainly includes all kinds of knowledge possessed by resources elements and experiences accumulated in simulation process, i.e., production flow, burden scheme, et al. E is reflects the comprehensive information of enterprise, such as organizational culture, creative ability and so on. The second is SC evolution function - fun (T, P, R, K, E), it expresses the logical relationship between each elements of SC, for instance, the completion of task will be affected by enterprise reputation and organizational culture. ¾ Simulation capability description Based on the above simulation capabilities model, in order to achieve SC servilizaiton, we need to select a proper way to realize the formal description of SC, as so far the existing service description language include web ontology language (OWL), OWL for service (OWL-S), simple HTML ontology extension (SHOE), can’t fulfill the actual requirements. According to the above problems and resources characteristics, many extended service description languages are proposed, for example, OWL-SP is presented based on OWL-S by adding dynamic logical operator. Due to the complexity, uncertainty and knowledge of SC, existing service description languages are not able to meet the requirement of SC servilizaiton. Based on the grammar and lexical of existing service description language, characteristics and servilizaiton requirement of SC were systematically analyzed and summarized, then achieved the expansion of service description language, expanded content main includes the description of the formation of SC, the logical relationship between the elements and related reasoning. ¾ Simulation capability application It is responsible for the application of SC servilizaiton. On the basis of above proposed methods and related technologies, a prototype of SCs system will be developed, its main functions include SC servilizaiton publication, capability service semantic matching, SC transaction and assessment, SC evolution and so on. Then SC can be provided to user in term of cloud services which are stored in CMfg system through the network. In addition, corresponding methods are taken according to different simulation resources in order to achieve resources intelligent accessing to CMfg system. For example, virtualization technology is adapted to simulation software access, but to simulation equipment, related technologies of the internet of things will be used for that.

support for user achieving on –demand use of simulation capability service.

5. CONCLUSIONS MC servilization is a core of CMfg philosophy, and is the key to achieve on-demand use and dynamic collaborative of manufacturing resources and capability. This paper discusses the application process of MC services’ life cycle in the complex product’s simulation stage, and then elaborates the simulation capability servilization in detail. Simulation capability servilization is helpful to provide user with simulation capability service by network. In the future, a prototype of cloud service platform for complex product simulation will be developed according to above proposed methods and related technologies.

ACKNOWLEDGMENTS This paper is partly supported by the NSFC (National Science Foundation of China) Projects (No.61074144 and No.51005012), the Doctoral Fund of Ministry of Education of in China (20101102110009), and the Fundamental Research Funds for the Central Universities in China. REFERENCES Skinner W, 1969. Manufacturing –missing link in corporate strategy. Harvard business review. 6-9, 136-145 Richard G B, 1972. The organization of industry. Economic Journal. 82, 883-896 Skinner W. 1974. The focused factory. Harvard business review, 5-6, 113-121 Hayes R, Wheelwright S. 1984. Restoring our competitive edge: Competing through manufacturing. John Wiley & Son. New York, 1721 P T Ward, R Duray, G K Leozzg, C sum. 1995. Business environment, operations strategy and performance: an empirical study of Singapore manufacturers. Journal of operations management. 13, 99-115 Morgan Swink, W. Harvey Hegarty. 1998. Core manufacturing capabilities and their links to product differentiation. International journal of operations & production management, 18(4), 374396 Keen P, Mcdonald M, 2000. The process edge: creating customer value and business wealth in the internet era. McGraw –Hill. New York Toni M. Somers, Klara G. Nelson, 2003. The Impact of strategy and integration mechanisms on enterprise system value: empirical evidence from manufacturing firms. European Journal of operational research, 146,315-338 Guo Haifeng, Tian Yezhuang, Liang Zhandong. 2006. An empirical analysis on relationships of manufacturing practices and manufacturing

¾

The center of simulation capability service The center of simulation capability service is the foundation of application related to SCS, it is responsible for field ontology library, modeling base and algorithm library. Due to knowledge is the basis of SC formation, so how to store the formal knowledge, is an important issue to SC servilization. Furthermore, description information of SC servilizaiton need to be stored and classified with related rules, it can provide

109

capabilities. IEEE International conference of management of innovation and technology, pp: 969-972 Rajiv D. Banker. 2006. Plant information systems, manufacturing capabilities & plant performance. MIS Quarterly, 30(2), 315-337 Robert Sarmiento, Mike Byrne, Nick Rich. 2007. Delivery reliability, manufacturing capabilities and new models of manufacturing efficiency. Journal of manufacturing technology management, 18(4), 367-386 LI Bohu, ZHANG Lin, WANG Shilong, TAO Fei, et al. 2010. Cloud manufacturing: a new serviceoriented manufacturing model. Computer Integrated Manufacturing Systems, 6(1), 1-7. ZHANG Lin, LUO Yongliang, TAO Fei, REN Lei, et al. 2010. Study on the key technologies for the construction of manufacturing cloud. Computer Integrated Manufacturing Systems, 16(11), 25102520 ZHANG Lin, LUO Yongliang, FAN Wenhui, TAO Fei, REN Lei, 2011. Analyses of cloud manufacturing and related advanced manufacturing models. Computer Integrated Manufacturing Systems, 17(3), 0458-0468. TAO Fei, CHENG Ye, ZHANG Lin, et al, 2011. Cloud manufacturing. Advanced materials Research, 201-203, 672-676 LUO Yongliang, ZHANG Lin, HE Dongjing, REN Lei, TAO Fei. 2011. Study on multi-view model for cloud manufacturing. Advanced materials Research, 201-203, 685-688 Siri Terjesen, Pankanj C. Patel, Jeffrey G. Covin. 2011. Alliance diversity, environmental context and the value of manufacturing capabilities among new high technology ventures. Journal of operations management, 29,105-115 CHENG Yun, YAN Junqi, FAN Minlun. 1996. Manufacturing environment modeling based on object-oriented and step technology. Chinese journal of mechanical engineering, 32(4),5-10 Khalid Hafeez, YanBing Zhang, Naila Malak. 2002. Determining key capabilities of a firm using analytic hierarchy process. International journal of production economics, 76, 39-51 Andreas Grobler. 2006. An empirical model of the relationships between manufacturing capabilities. International journal of operations & production management, 26(5),458-485

Lin Zhang received the B.S. degree in 1986 from the Department of Computer and System Science at Nankai University, China. He received the M.S. degree and the Ph.D. degree in 1989 and 1992 from the Department of Automation at Tsinghua University, China, where he worked as an associate professor from 1994. He served as the director of CIMS Office, National 863 Program, China Ministry of Science and Technology, from December 1997 to August 2001. From 2002 to 2005 he worked at the US Naval Postgraduate School as a senior research associate of the US National Research Council. Now he is a full professor in Beijing University of Aeronautics and Astronautics. He is an Editor of “International Journal of Modeling, Simulation, and Scientific Computing”, and “Simulation in Research and Development”. His research interests include integrated manufacturing systems, system modeling and simulation, and software engineering. Fei Tao is currently an associate professor at Beihang Universigy since April 2009. His research interests include service-oriented manufacturing, intelligent optimizaiton theroy and algorithm, resource servcie management. He is author of one monograph and over 20 journal articles of these subjects. Dr Tao was awarded the Excellent Doctoral Dissertation of Hubei Province, China and was elected to be a research affiliate of CIRP in 2009. He is the founder and editore-in-chief of Internatioanl Journal of Servcie and Computing-Oriented Manufacturing (IJSCOM). Yuan Bao received the B.S. degree in 2010 from the Department of Information Engineering at Tianjin University of Commerce, China. He is currently working for the M.S. degree in Modeling simulation theory and technology at Beihang University. His research interests include service-oriented manufacturing and integrated manufacturing systems.

AUTHORS BIOGRAPHY Yongliang Luo received the B.S. degree and the M.S. degree in 2006 and 2009 from the Department of Computer Science at Shandong University of Science and Technology, China. He is currently working for the Ph.D. degree in Modeling simulation theory and technology at Beihang University. His research interests include service-oriented manufacturing and integrated manufacturing systems.

110

GYRUS AND SULCUS MODELLING UTILIZING A GENERIC TOPOGRAPHY ANALYSIS STRATEGY FOR PROCESSING ARBITRARILY ORIENTED 3D SURFACES Gerald Zwettler(a), Werner Backfrieder(a,b) (a)

Bio- and Medical Informatics, Research and Development Department, Upper Austria University of Applied Sciences, Austria (b) School of Informatics/Communications/Media, Upper Austria University of Applied Sciences, Austria (a)

[email protected], (b)[email protected]

processing. Utilizing T1-weighted brain MRI data, segmentation of gray and white matter can be achieved, using k-means clustering (Kanungo et al. 2002; Ibanez et al. 2005) for determination of the tissue types to discriminate and region growing for ensuring connectedness. Based on a binary segmentation of the brain surface, several strategies for sulcus and gyrus classification have been presented and published in the past. A morphologic closing operation, i.e. dilation followed by erosion, with subsequent subtraction of the original MRI data allows processing of the sulcus volume via 3D skeletonization for extraction of the sulcus and gyrus folds (Lohmann 1998). In contrast to applying Euclidean distances, the use of a geodesic depth profile accounts for complexity and partial occlusion of the sulcus folds (Kao et al. 2006). Besides these morphologic concepts, curvature analysis of a surface mesh calculated from gray and white matter can be utilized for detection of the gyrus and sulcus course (Vivodtzev et al. 2003) with respect to convexity and concavity. In this paper we present a generic strategy for topographic analysis of arbitrary shapes and transformation of the depth profiles into cyclic graph representations. Thereby we account for imbalances in the local depth profiles due to asymmetries and deformations of the brain. The minimum graph connecting all local maxima and minima respectively is calculated, normalizing the local depth levels similar to the watershed segmentation concept. Our strategy is perfectly applicable for the task of gyrus and sulcus modelling as concave and convex paths can be identified. Based on the graph representations of the sulcus and gyrus courses, modelling of the brain surface can be easily achieved via distance-based classification utilizing morphologic operators as presented and discussed in the following sections.

ABSTRACT Accurate and robust identification of the gyri and sulci of the human brain is a pre-requisite of high importance for modelling the brain surface and thus to facilitate quantitative measurements and novel classification concepts. In this work we introduce a watershedinspired image processing strategy for topographical analysis of arbitrary surfaces in 3D. Thereby the object’s topographical structure represented as depth profile is iteratively transformed into cyclic graph representations of both, the lowest and the highest characteristics of the particular shape. For graph analysis, the surface elements are partitioned according to their depth value. Neighbouring regions at different depth levels are iteratively merged. For region merging, the shape defining medial axes of the involved regions have to be connected by the optimum path with respect to a fitness function balancing shortness and minimal depth level changes of the solution. Keywords: topographical surface analysis, cyclic graph representation, sulcus and gyrus classification 1. INTRODUCTION The accurate quantification of metabolic processes from functional emission tomography imaging modalities like positron emission tomography (PET) and single photon emission computed tomography (SPECT) for diagnosis of neurodegenerative diseases necessitates a precise and patient-specific segmentation and classification of the brain. For segmentation and classification tasks, morphological image modalities as magnetic resonance imaging (MRI) have to be fused with the data acquired by functional emission imaging. Thus, the segmentations and classifications evaluated based on the anatomically-precise imaging modalities can be applied to the emission data, facilitating quantitative analysis of the metabolic activity with respect to pre-classified anatomical regions. The classification concept addressed in this work is the partitioning of gray and white brain matter according to the gyrus and sulcus characteristics. Any computer-based functional or anatomical classification requires binary segmentation as pre-

2. MEDICAL BASICS Classification of the human brain can be accomplished at different levels of granularity. At a top level, the main anatomical structures, like cerebrum, cerebellum and the brain stem can be identified. The cerebrum is

111

subdivided into two hemispheres and the main anatomical components, like white matter, gray matter, cerebro-spinal fluid (CSF), ventricle, fat, bones and the arterial and venous vessel systems are demarcated. The brain tissue composed of white and gray matter is subclassified into frontal lobe, parietal lobe, occipital lobe and some more, all specific areas responsible for diverse neurological functions of the body (Pschyrembel 2002). Each lobe comprehends several gyrus and sulcus areas, forming the brain surface. Thereby the gyri refer to the convex bulgs on the brain surface that are delimited by convex trenches, the so called sulci. The notable main sulci and gyri are named, listed and charted in anatomical atlases (Ono et al. 1990). The topography of gyrus and sulcus characteristics is highly applicable for registration tasks in case of multi-modal image processing or follow-up examinations. Furthermore, modelling of the gyrus segments facilitates the quantitative analysis of metabolic activities with respect to defined anatomical structures.

erosion can be utilized for smoothing the surface and calculation of the reference shape, see Fig. 2 (c).

(a)

(b)

(c)

(d)

Figure 1: The stack of 2D MRI slices (a) assembles a 3D volume of the brain (b). After binary segmentation (c), the calculated convex hull (d) is the reference shape for depth profile calculation.

3. DATA For testing of the gyrus and sulcus modelling concept, n=20 T1-weighted MRI datasets of simulated brainweb database (Cocosco et al. 1997; Kwan, Evans, and Pike 1999) and associated reference segmentations are used. Further test runs and validations will be performed utilizing n=12 anonymous multi-modal patient studies comprising morphologic image acquisitions (T1, T2, PD, …) as well as functional images (SPECT, PET).

(a)

(b)

(c)

Figure 2: The precisely calculated surface model (a) can be smoothed for use as reference shape via rough isosurface calculation (b). As an alternative reference shape calculation strategy, a morphological closing operation with dilation kernel size 7x7x7 followed by erosion 6x6x6 can be applied (c).

4. METHODOLOGY Prior to performing the analysis process, a binary representation of the targeting object’s surface, not addressed in this work, and a 3D depth profile must be pre-processed. 4.1. Estimation of the Reference Shape For calculation of the depth profile of an arbitrarily shaped object, the reference shape, i.e. the smoothed shape without the vales and ridges, must be estimated. Processing a solid body with an approximately spherical shape, like the human brain, calculation of the 3D convex hull (Barber, Dobkin and Huhdanpaa 1996; Sonka, Hlavac and Boyle 2007) as reference shape is highly feasible, see Fig. 1. For other more complex shapes, where a spherical approximation would be too imprecise, an alternative calculation of the reference shape is feasible. When calculating a winged-edge isosurface of a binary 3D body (Baumgart 1972; Baumgart 1975; Ritter 2007; MeVis 2011), utilizing the quality factor, allows steering of the smoothing effect by polygonal reduction, i.e. up to which level, vales and ridges should influence the depth profile calculation, see Fig. 2 (b). The resulting isosurface is projected back to regular 3D voxel grid for further processing. Furthermore, morphologic closing operations as dilation followed by

4.2. Calculation of the Depth Profile The depth profile is calculated as the minimum Euclidean distance between the surface of the object and the reference shape, see Fig. 3 as illustration of 2D depth profile calculation. A distance map calculation is used to represent the depth profile in 3D with the Euclidean neighbourhood weights for the adjacency constellations N6, N12 and N8 in 3x3x3 neighbourhood according to the distance from the hot spot as w N 6 1.0 ; w N12

2 ; w N8

3.

(1)

Starting at the convex hull, the distance weights are propagated to the particular neighbours to set or update their values. Whenever the depth value of a neighbour gets adapted, the change is recursively propagated to all adjoined neighbours. Calculation of the depth profile is finished, when the depth value of each voxel refers to the minimal Euclidean distance from the convex hull.

112

(a)

(a)

(b)

(c)

(b)

(d)

(e)

Figure 4: Planar representation of binary brain (a), convex hull (b) and depth profile (c) of mid slice transversal (axial) view. The 3D depth profile (d) as 2D map via Mollweide projection (Snyder 1987) is shown in (e).

(c)

The depth profile is processed at each particular depth level, starting at the maximum depth values and ending with the lowest depth values. At each depth level, voxels adjoined in N26 neighbourhood having matching depth values, are assembled together to recursively build up larger connected regions. For each of these constructed regions, the adjoined neighbourhood regions must be identified, differentiating between the following three constellations for the neighbourhood count N:

(d)

Figure 3: Calculation of the 2D depth profile for cavities between convex hull and the object’s surface (a). In 2D case the neighbour weights are defined as w(N4)=1.0 and w(ND)=1.41. The depth profile is iteratively propagated until all depth values are calculated and convergence is reached (b-c). A colourencoded representation of the final depth profile is plotted in (d).

1.

The depth profile is calculated utilizing an Euclidean distance transform (Sonka, Hlavac and Boyle 2007). For calculation of the distance metrics, morphologic propagation of the outer surfaces, similar to the concept of grassfire transform (Blum 1967) is applied for fast approximate calculation of the Euclidean distance map from the object’s borders. Results of the depth profile calculation are presented in Fig. 4. For the depth profile only the outer depth values below a threshold t are of relevance for gyrus and sulcus modelling. Depth profile values in the inner ventricular area are to exclude.

2.

3.

4.3. Watershed-Inspired Topography Analysis Based on the calculated depth profile, the shortest graph interconnecting all local depth minimums is calculated, as well as a graph for connection of all local depth maxima. In the following delineation of the method, only the graph creation for the local maxima is addressed. The graph modelling for the local minimums can be derived by changing the leading signs and processing order. The iterative process of graph construction is outlined in pseudo-code listing 1 and explained in detail in the following paragraphs.

N=0: there are only background voxels or voxels at lower depth level not yet processed in the neighbourhood Î local maximum detected. N=1: region is adjoined to one neighbouring region processed before with a higher depth value Î current region will be merged with the existing one. N>2: there are several adjoined regions with higher depth values. The region to process is merged with the first neighbour region. Then the other neighbours are iteratively merged to one remaining cumulated region. Thereby, the skeleton sub-graphs must be interconnected.

For the autonomous regions to be interpreted as local maxima (condition 1), the first part of the graph is calculated via skeletonization. Below a region count of R=50 voxels, the region element closest to the centre of mass is taken as starting sub-branch of the graph, whereas for larger regions the medial axis, precisely extracted via skeletonization (Jonker 2002; Zwettler et al. 2009), is applied as starting sub-branch of the graph. Regions with only one adjoined neighbouring region (condition 2), necessitate no discrete skeleton calculation. Instead, the region elements are just merged

113

with the neighbouring region, already defining a subbranch. If the current region is surrounded by more than one neighbour region at higher depth profile values (condition 3), besides a merge with the first neighbour region, all involved adjoined regions need to be cumulated. As all of the involved neighbouring regions have an already defined sub-part of the graph, these segments must be iteratively linked together. This link operation, described in the following sub-chapter, is a crucial task as it significantly influences the resulting graph after processing all depth levels from the deepest to the lowest profile values, see Fig. 5 for illustration of described iterative topography analysis. Starting at the deepest values with autonomous regions and the first skeleton parts Fig. 5 (c), the regions are enlarged whenever adjoined new regions at lower depth profile values are reached Fig. 5 (d-e). In case of reaching regions with already defined skeletons, the optimum connective path must be found Fig. 5 (f-g) utilizing detection of the optimal path, see Fig. 6. That way the topography describing path can be iteratively constructed until one final region remains Fig. 5 (h). regions; for(depth=maxVal; depth>=minVal; depth--) currRegions; for(xIdx=0; xIdx < sizeX; xIdx++) for(yIdx=0; yIdx < sizeY; yIdx++) for(zIdx=0; zIdx < sizeZ; zIdx++) if((distanceMap[xIdx][yIdx][zIdx] == depth)&& (!classified(xIdx,yIdx,zIdx))) currRegions.Add( getRegionAt(xIdx,yIdx,zIdx)); for(region : currRegions) neighbourRegions = getNeighbours(region); if(neighbourRegions.size == 0) region.CalculateSkeleton(); regions.Add(region); else if(neighbourRegions.size == 1) merge(neighboruRegions.first, region) else merge(neighbourRegions.first, region) for (nRegion : neighbourRegions) if(nRegion != neighbourRegions.first) link_merge( neighbourRegions.first, nRegion)

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

Figure 5: Illustration of the topography analysis on an cross-shaped 2D terrain. 4.3.1. Assembling Graph from Sub-Segments When assembling two regions r1 and r2 with precalculated skeletons, i.e. graph sub-segments, the optimum path between the two skeletons skel1 and skel2 must be found. The search for the optimal connection path pmin is defined as minimization problem of a fitness function F, accounting for the Euclidean distance dist() of the path and depth gradients of the path as depth() with respect to the maximum profile depth depthmax, as

Code Listing 1: Illustration of the topography analysis algorithm implemented in pseudo code. Regions at the same depth level are grouped together via region growing and stored in currRegions. Then for each region in currRegions, the neighbouring regions are identified. In case of neighbourhood condition 1 with N=0, a seed region has been detected and the first skeleton is calculated. Seed regions are added to the global region container regions. If there is exactly one neighbouring region at higher depth level (condition 2), a region merging is performed. In case of additional neighbours (condition 3), the shortest skeleton path linking the involved regions is calculated prior to performing the merge operation.

length ( p )

F ( p)

¦ dist ( p , i  1, i ) ˜ 1  depth

max

 depth ( i ) . (2)

i 2

Thereby the target path pmin is that path of all possible non-cyclic connections between skel1 and skel2 with a minimal value for F. For calculation of the optimal path, the metric value F of the current sub-path is propagated to its neighbours, starting at the skeleton elements skel1 of

114

region r1. The sub-path fitness value is recursively propagated to the particular neighbours and updated for each added path element. Whenever the first element of skel2 is reached, an upper border for the optimal path fitness value is given. That allows reduction in the calculation complexity as all sub-paths exceeding this upper border fitness value can be aborted. After convergence of the fitness values is reached, the optimum path can be traced back starting at the element e2 of skeleton skel2 with an adjoined neighbour showing the lowest fitness. Starting at e2 the way to skeleton skel1 is traced back by selecting the element with lowest fitness in each particular neighbourhood. The optimal path is finished, when the first neighbour of skel1 has been reached. The described algorithm is illustrated in Fig. 6.

(a)

(c)

(a)

(b)

(c)

(d)

(b)

(e)

(f)

(g)

(h)

(d)

Figure 6: Detection of the optimal path between the two regions in Fig. 5 (f). Starting at the skeleton of the first region, all neighbours in N8 are set or updated with the minimal path value according to the metrics defined in Eq. 2. Whenever a path value gets changed, all adjoined neighbours must be checked for propagation of the new value (b). Finally the path connecting the skeletons skel1 and skel2 can be traced back by picking each neighbour with the lowest value until skeleton of r1 is reached (d).

Figure 7: Visualization of sulcus analysis results in sagittal view (a-d) and axial view (e-h). The resulting tree model for the sulcus folds (b)(f) is amplified via morphologic dilation operation utilizing a 3x3x3 structuring element for better visibility (c)(g). Correlation of original brain surface (a)(e) and the sulcus models is presented in (d)(h) via overlay.

5. RESULTS In this chapter first results of the discussed method are presented and discussed. 5.1. Results of Sulcus Classification The graph model resulting from sulcus classification highly correlates with the main anatomical folds, see Fig. 7. Some pruning and smoothing operations on the graph can be utilized in future to remove dispensable bifurcations and redundant node elements.

As our presented algorithm only interconnects adjoined region skeletons bridging shallow sections, foreshortening could arise if the depth profile is strictly monotonic increasing. That problem can only be exploited by processing artificial testing data. As the first results show, even minimal deflection from strict monotonic behaviour prevents any foreshortening of the particular sub-branch, processing real-world medical image data.

115

Ibanez, L., Schroeder, W., Ng., L., Cates, J., 2005. The ITK Software Guide. Kitware Inc. Jonker, P.P., 2002. Skeletons in N dimensions using shape primitives. Pattern Recognition Letters, 23:677-686. Kanungo, T., Mount, D.M., Netanyahu, N., Piatko, C., Silverman, R., Wu, Y.A., 2002. An efficient kmeans clustering algorithm – Analysis and implementation. In: IEEE Transactions on Pattern Analysis and Machine Intelligence, 24:881-892. Kao, C.-Y., Hofer, M., Sapiro, G., Stern, J., Rottenberg, D., 2006. A Geometric Method for Automatic Extraction of Sulcal Fundi. In: IEEE International Symposium Biomedical Imaging, 1168-1171. Kwan, R.K.-S., Evans, A.C., Pike, G.B., 1999. MRI simulation-based evaluation of image-processing and classification methods. IEEE Transactions on Medical Imaging, 18(11):1085-1097. Lohmann, G., 1998. Extracting Line Representations of Sulcal and Gyral Patterns in MR Images of the Human Brain. In: IEEE Transactions on Medical Imaging 17(6):1040-1048. MeVis, 2011. MeVisLab – medical image processing and visualization, MeVis Medical Solutions, Bremen, Germany. Available from http://www.mevislab.de/developer/documentation/ [04/2011]. Ono, M., Kubick, S., Abernathey, C.D., 1990. Atlas of the cerebral sulci. Thieme publisher Pschyrembel, W., 2002. Pschyrembel klinisches Wörterbuch. De Gruyter, Berlin, Germany. Ritter, F., 2007. Visual Programming for Prototyping of Medical Applications. In IEEE Visualization 2007 workshop. Snyder, J.P., 1987. Map Projections – A Working Manual. In: U.S: Geological Survey Professional Paper 1395, Washington DC., U.S. Government Printing Office. Sonka, M., Hlavac, V., Boyle, R., 2007. Image Processing, Analysis, and Machine Vision. Cengage Learning, 3rd edition. Vivodtzev, F., Linsen, L., Bonneau, G.-P., Hamann, B., Joy, K.I., Olshausen, B.A.,2003. Hierarchical Isosurface Segmentation Based on Discrete Curvature. In: Joint EUROGRAPHICS – IEE TCVG Symposium on Visualization, 249-258. Zwettler, G.A., Swoboda, R., Pfeifer, F., Backfrieder, W., 2009. Fast Medial Axis Extraction Algorithm on Tubular Large 3D Data by Randomized Erosion. In: Ranchordas, A.K., Araujo, H.J., Pereira, J.M., Braz, J., eds. Computer Vision and Computer Graphics – Communications in Computer and Information Science. Springer Publisher, 24:97-108. Zwettler, G., Backfrieder, W., Pichler, R., 2011. Diagnosis of Neurodegenerative Diseases based on Multi-modal Hemodynamic Classification of the Brain. In Proceedings of the International Conference on Computer Aided Systems Theory EUROCAST 2011, 363-365.

6. DISCUSSION AND CONCLUSIONS In this paper we introduce a novel concept for classification of the gyrus and sulcus folds and representation as graph model. The presented iterative topography analysis with step-wise assembling and interconnecting the final graph allows handling of local imbalances of the depth profile, e.g. due to deformation or asymmetries of the brain. Besides roughly estimated upper threshold value for the profile depth to process with respect to imaging resolution, no further parameterization is required. Linking the final graph representations and the depth profile allows later restrictions on the granularity of sulcus and gyrus representation. Utilizing the presented sulcus and gyrus graph models, classification of the brain areas can be easily achieved by iterative assigning the gray matter and white matter voxels to the closest neighbouring subtrees, similar to vascularization-based anatomy classification (Zwettler, Backfrieder and Pichler 2011). Thereby the gyrus course can be used for iterative classification of the surrounding tissue and the sulcus lines are incorporated as additional barriers to prevent invalid merging of adjoined brain sections. ACKNOWLEDGMENTS Thanks to our medical partners from the WagnerJauregg state mental hospital, Linz, Upper Austria, at the institute for neuro-nuclear medicine headed by MD Robert Pichler for providing medical image data and valuable discussion. This research is part of the INVERSIA project (http://inversia.fh-linz.at) which was funded by the European Regional Development Fund (ERDF) in cooperation with the Upper Austrian state government (REGIO 13).

REFERENCES Barber, C.B., Dobkin, D.P., Huhdanpaa, H., 1996. The Quickhull Algorithm for Convex Hulls. In ACM Transactions on Mathematical Software (TOMS). Baumgart, B.G., 1972. Winged-Edge Polyhedron Representation, Stanford University Artificial Intelligence Report No. CS-320. Baumgart, B.G., 1975. Polyhedral Representation for Computer Vision. In Proc. of the National Computer Conference, 589-596. Blum, H., 1967. A Transformation for Extracting New Descriptors of Shape. In Wathen-Dunn, W., eds. Models for the Perception of Speech and Visual Form. MIT Press, Cambridge, 362-380. Cocosco, C.A., Kollokian, V., Kwan, R.K.-S., Evans, A.C., 1997. BrainWeb: Online Interface to a 3D MRI Simulated Brain Database. Proceedings of the 3-rd International Conference on Functional Mapping of the Human Brain, 5(4):425.

116

AUTHORS BIOGRAPHY Gerald A. Zwettler was born in Wels, Austria and attended the Upper Austrian University of Applied Sciences, Campus Hagenberg where he studied software engineering for medicine and graduated Dipl.Ing.(FH) in 2005 and the follow up master studies in software engineering in 2009. In 2010 he has started his PhD studies at the University of Vienna at the Institute of Scientific Computing. Since 2005 he is working as research and teaching assistant at the Upper Austrian University of Applied Sciences at the school of informatics, communications and media at the Campus Hagenberg in the field of medical image analysis and software engineering with focus on computer-based diagnostics support and medical applications. His email address is [email protected] and the research web page of the Research & Development department at campus Hagenberg, he is employed at, can be found under the link http://www.fhooe.at/fe/forschung. Werner Backfrieder received his degree in technical physics at the Vienna University of Technology in 1992. Then he was with the Department of Biomedical Engineering and Physics of the Medical University of Vienna, where he reached a tenure position in 2002. Since 2002 he is with the University of Applied Sciences Upper Austria at the division of Biomedical Informatics. His research focus is on Medical Physics and Medical Image Processing in Nuclear Medicine and Radiology with emphasis to high performance computing. Recently research efforts were laid on virtual reality techniques in the context of surgical planning and navigation.

117

FAST MARCHING METHOD BASED PATH PLANNING FOR WHEELED MOBILE ROBOTS Gregor Klanˇcar (a) , Gaˇsper Muˇsiˇc(a) (a)

University of Ljubljana Faculty of Electrical Engineering Trˇzaˇska 25, Ljubljana, Slovenia (a)

[email protected], [email protected]

rithms may be used in combination with different representations of the environment, although certain representations are more suitable for some algorithms. Most generally used algorithms include standard A* and its derivatives D*, D* Lite, E*, algorithms based on fast marching method and others. The literature review indicates environment segmentations based on Delaunay triangulation (dual to Voronoi diagram) (Hongyang, et al. 2008, Kallmann 2005) and framed quadtrees (Davis 2000) among the most promising approaches, and A* based path search algorithms as a suitable way of path optimization within the segmented environment. The computational demand of planning in complex environments is reduced by two or three level planning and exploitation of a-priori knowledge or heuristics (Botea, M¨uller and Schaeffer (2004)). An approach to improve path planning computational complexity by the use of interpolation and D* search algorithm is proposed in (Ferguson and Stentz 2006). The incremental search method that finds shortest paths for similar path -planning problems faster than uninformed search methods is presented in (Koenig 2004). Alternative approaches, e.g. bug algorithms, are also investigated but their application is questionable as the results are often not predictable, calculated path may be far from optimal or the target is not reached. A good overview of path planning approches can be found in (LaValle 2006). In the presented work the focus is given to the efficiency of path planning algorithm which is applicable to static or slowly changing dynamic environments. An A* path planning based algorithm is introduced with the quadtree and triangular representations of the search space. For fast changing environments the approach could be upgraded using D* algorithm as proposed in (Ferguson and Stentz 2006). As an alternative, the Fast Marching Method (FMM) based path planning was also investigated. A novel combination of path planning algorithms has been proposed with a suboptimal but efficient corridor calculation and an advanced FMM based path optimization within the corridor. The main goal of the presented approach is to obtain shortest smooth path between the obstacles, which complies to given kinematic constraints and can be calculated at acceptable computational complexity.

ABSTRACT The paper presents a path planning approach for wheeled mobile robots in obstructed environments. The trajectories of moving objects have to be carefully planned in order to obtain a near-shortest smooth path at still acceptable computational complexity. The combined approach is therefore proposed which utilizes search algorithm A* as well as methods of numerical solving of a particular form of partial differential equation - an eikonal equation. The use of related fast marching method enables to derive smooth trajectories within the shortest path corridor identified by the heuristic search algorithm while keeping the on-line computational burden relatively low. To illustrate the basic idea our investigation is limited to situation with static obstacles, e.g. buildings in the area which is crossed by autonomous vehicles. The proposed approach operation is validated by experimental results on a differential mobile robot. Keywords: mobile robots, path planning, quadtrees, triangulation, fast marching method 1. INTRODUCTION In the obstructed environments autonomous moving vehicles need to plan collision safe paths. With the given map of the environment and the target location the path planning aims to determine a trajectory, which will lead the object from the starting position to the target position. In general the planning involves two stages: (i) a suitable representation of the environment where the path has to be planned, and (ii) a search algorithm, that is capable of finding (sub)optimal path from the initial to the target position. The planning can involve a third stage where the path is optimized taking into account dynamic constraints of a moving object. This often leads to a requirement for smooth moving paths free of sharp turns. The path planning methods initially transform the environment where the object of interest resides into a structure adapted to the requirements of path planning. Such representations include generalized Voronoi diagrams, various methods of space triangulation, regular grids and quadtrees, among others. Particular path planning algo-

118

1.8

2. SEARCH SPACE SEGMENTATION BASED PATH PLANNING METHODS Two space segmentation methods were considered in combination with an A* based path optimizer: quadtree segmentation and Delaunay triangulation.

1.6

1.4

1.2

1 y [m]

2.1. Quadtrees space segmentation Quadtrees (QT) enable a decomposition of the space map into quadratic cells (with the varying dimension) which are either free or occupied by obstacles. An essential step in quadtree generation is the cell occupancy test which should be carefully designed to retain computational efficiency. QT segmentation results in compact environment map presentation and enables efficient query about occupancy of some position or area in the environment (Botea, M¨uller and Schaeffer (2004), Davis 2000). To shorten the computational time of A* path finding algorithm the obtained quadtree is extended with visibility graph which indicates possible paths among free cells. To each free leaf cell an array of its free neighbour indices is adjoined. This enables easier and computationally efficient moves of A* algorithm among the free quadtree leafs. The visibility graph is computed as follows:

0.8

0.6

0.4

0.2

0 0

0.5

1 x [m]

1.5

2

Figure 1: A* path planning based on QT segmentation for the given example. When certain edges are fixed in advance and the rest of the edges are determined in the DT way, the triangulation is called Constrained Delaunay Triangulation (CDT). The pre-defined edges are constraints during triangulation. CDT is useful in cases where there are objects in the plane which should be taken into account during plane segmentation. Such is the case of path planning in a plane with obstacles. The triangulated plane can be used for path planning in a similar manner as the quadtree based segmentation. A connectivity graph is built which contains information about the edges that can be crossed, i.e. the edges that do not belong to obstacles. Based on the triangulation and the connectivity graph, a path can be searched for by an A* type algorithm. A corresponding example for the same environment configuration as in Figure 1 is shown in Figure 2. The additional thin edges drawn within the triangles show the alternative paths explored by the algorithm. It can be observed that the built-in heuristic helps the algorithm to search only a part of the whole space.

• Query for the area that is a little bigger (1/2 size of the prescribed minimal cell size) than current leaf size. • Find leafs that are visible (accessible) from current leafs. Store their indices to the current leaf visibility array. • Calculate distances to the visible neighbor leafs and store them in an array. By this stage the obtained quadtree structure is prepared for A* path finding algorithm. The pathfinding algorithm based on A* search strategy guarantees the shortest path between start and end point if an optimistic heuristic is used. This means that predicted path cost (length) must always be smaller (shorter) or at lest equal to the real path cost (length). Choosing line of sight for the predicted path length always fulfils this condition. A corresponding example of the space segmentation and a calculated sample path are shown later in the section with experimental results (Fig. 1).

2

2.2. Space segmentation by constrained Delaunay triangulation Triangulation is an important tool for representation of planar regions and is used in several application. The planar region is divided into a number of sub-regions of triangular shape. Commonly a set of points in the plane is given and the vertices of sub-regions must match the given set. This can be achieved in several ways, one of the possible triangulations is the Delaunay triangulation (DT). The Delaunay triangulation assures a minimal number of narrow triangles (with small internal angles) which is a required property in many applications. The algorithms of Delaunay triangulation are also well explored and efficient algorithms yield computational complexity of n log n where n is the number of given points.

1.5

1

0.5

0 0

0.5

1

1.5

2

Figure 2: Example of path planning based on triangular segmentation

119

2.3. Path smoothing and shortening As can be observed from Figures 2 and 1 the obtained path passes through centers of the cells in quadtree space representation or crosses midpoints of the triangle edges in triangulation. None of the two approaches gives a smooth or straight path. The path is cornered even in the areas with no obstacles, which is not desired from the practical point of view, since the movement of an autonomous object within the environment would be unnatural in such cases. This can be improved by considering the fact that the array of cells which are crossed by the calculated path actually defines a channel between the starting and the target point. This channel will be termed a path corridor and consists of a set of neigbouring square shaped cells when using quadtrees. When using CDT the corridor consists of a set of neighbouring triangular cells. In both cases, the path within the corridor can be optimized by applying the funnel algorithm (Kallmann 2005), which finds the shortest path within the corridor. The basic steps of the algorithm are briefly described as follows:

1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 -0.2

0

0.5

1

1.5

2

Figure 3: Example of the optimal path (thick dark line) determination inside the corridor based on the triangular representation.

• The corridor is defined by two sets of points: a set of points defining the upper corridor border and a set of points defining the lower corridor border. In case of triangulation the corridor border follows the edges of participating triangles (Fig. 3). In case of quadtree representation, the size of the neigbouring cells may differ, and in such a case only the overlapping part of the edges between the two cells is considered when defining lower and upper corridor border (see Fig. 9).

Real-time in this case means a time, which is short enough for the robot to plan a journey without a considerable delay. E.g., delays up to 0.5 s may be tolerated. Furthermore, we limited our investigation to situation with static obstacles, e.g. buildings in the area which is crossed by autonomous vehicles. All mentioned algorithms were implemented in Matlab m-code and both segmentation methods were tested in combination with environments of different complexities. Tables 1 and 2 show results of a set of experiments with varying environment size and varying number of obstacles within the environment on a 2.4 GHz PC. Within the tables, the first column shows the number of obstacles, next is the min/max size of segmentation cells, Ncells is the number of cells after the segmentation, tQT and tDT are the segmentation computing times, and the remaining columns show computing times and corresponding standard deviation of path search algorithm, corridor boundaries calculation, and the funnel algorithm, respectively. The computing times were obtained by averaging eight consequent runs of the path planning algorithm with different start and target points. The obtained results show that both methods can be used for path planning in real-time for moderately complex

• The start and the target point are linked to the upper and lower corridor border by additional edges. • Let p be the starting point and let u and v be the points on the upper and lower border of the corridor, respectively. The shortest path from p to u and from p to v (not leaving the corridor) may overlap up to some point a. At a the paths diverge and are concave until they reach u and v. The a is called apex and the region delimited by path segments from a to u, a to v and uv segment is the funnel. • The algorithm iteratively adds points on the corridor borders narrowing the funnel and discarding the points falling out of the narrowed funnel. When the top of the funnel shrinks down to a line, the shrunk part of the funnel defines a new segment of the shortest path and the new apex is set at the end of the new segment. The procedure stops when the target is reached. For more details, see (Kallmann 2005).

Table 1: Computation times - quadtrees Environment size 10 x 10 m, maximal obstacle size 1 x 1 m tCO tFU tA* Nobst MinDim Ncells tQT ? A* ?CO ?FU [s] [ms] [ms] [ms] [ms] [ms] [ms] [m] 10 10/32 273 0.2 9 6 4 2 6 3 10/64 577 0.5 11 6 2 1 4 2 10/128 1301 1.0 12 11 2 1 4 2 20 10/32 405 0.4 12 7 4 2 6 3 10/64 877 0.8 11 8 3 1 5 2 10/128 2189 1.8 14 12 3 1 5 2 50 10/32 797 1.0 13 9 4 1 6 2 10/64 1921 2.6 23 23 3 2 5 2 10/128 4817 6.2 49 56 4 3 7 4 100 10/64 3345 6.0 47 37 5 3 9 5 10/128 9577 16.7 83 114 4 3 7 4 Environment size 100 x 100 m, maximal obstacle size 5 x 5 m 100 100/256 12265 25 196 192 5 2 8 4

As an example, the planned path from Fig. 2 for the triangular representation is improved by funnel algorithm as shown in Fig. 3. 2.4. Experimental results The initial requirement of this investigation was to determine a path planning method which would enable fast, real-time path planning for moving objects within the obstructed environments. The primary application area are wheeled mobile robots moving at relatively slow speeds.

120

where

Table 2: Computation times - triangulation Environment size 10 x 10 m, maximal obstacle size 1 x 1 m tA* t CO t FU Nobst MaxDim Ncells tDT s A* sCO sFU [s] [ms] [ms] [ms] [ms] [ms] [ms] [m] 10 / 82 0.2 35 22 / / 4 7 1.25 244 0.4 63 54 / / 0 0 0.67 594 1.3 147 169 / / 0 0 0.4 1434 6.6 369 350 / / 14 6 20 / 168 0.3 53 36 / / 6 8 1.25 328 0.6 96 67 / / 0 0 0.67 678 1.7 160 145 / / 6 8 0.4 1518 6.0 465 447 / / 10 8 50 / 432 0.9 107 36 / / 6 8 1.25 606 1.6 133 91 / / 4 7 0.67 956 3.3 348 280 / / 8 8 0.4 1796 10.1 797 539 / / 12 7 100 / 958 3.5 199 88 / / 6 8 1.25 1076 4.0 217 136 / / 8 8 0.67 1426 6.2 348 188 / / 6 8 0.4 2266 13.8 687 513 / / 10 8 Environment size 100 x 100 m, maximal obstacle size 5 x 5 m 100 / 838 2.8 352 220 / / 10 8

−x T = Dij

(3) Ti,j − Ti,j−1 −y Dij T = Δy and +x T = Dij

−x +x T, −Dij T, 0)2 + max(Dij −y +y T, −Dij T, 0)2 max(Dij

1 Fij2

T T1 T2

= Ti,j = min(Ti−1,j , Ti+1,j ) = min(Ti,j−1 , Ti,j+1 )

(7)

the equation becomes  max

2 2  T − T2 1 T − T1 , 0 + max ,0 = 2 Δx Δy Fij

(8)

Assuming F is always positive, T is monotonically increasing. The solution in a given point is only influenced by solution values in those points where the solution value is smaller. The fast marching method is based on the information propagation in one direction, from smaller values of T to larger values (Baerentzen 2001, Farag and Hassouna 2005). T(i,j+1)

T(i-1,j)

T(i,j)

T(i+1,j)

T(i,j-1)

1 = 2 Fij

(5)

By setting new labels

(1)



=

Considering (3) and (4) the last equation can be modified to 2  Ti,j − min(Ti−1,j , Ti+1,j ) ,0 + max Δx 2  1 Ti,j − min(Ti,j−1 , Ti,j+1 ) ,0 = 2 (6) max Δy Fij

at given border conditions, i.e. at a condition of zero value of T in the starting point. If F depends only on space coordinates, the above equation is an Eikonal equation. The numerical solution of the equation is based on a space grid, approximation of the gradient by the values in the neighbouring points and an efficient strategy of point calculation order. Figure 4 shows a point in a 2D grid and its neihbouring points. In 2 dimensions the gradient |∇T (x, y)| = |∇T (i, j)|, where x = iΔx in y = jΔy can be substituted by an approximation (2) −x +x T, 0)2 + min(Dij T, 0)2 + max(Dij −y +y 2 max(Dij T, 0) + min(Dij T, 0)2

Ti,j+1 − Ti,j = Δy

In (Sethian 1999) is proposed that the gradient is substituted by a simpler, less accurate approximation

3. FAST MARCHING METHOD BASED PATH PLANNING Fast marching method (FMM) is based on numerical analysis of viscous matter and is a method of numerical solving a particular form of nonlinear partial equation, i.e. an Eikonal equation. Simplified, the method gives a description of wavefront propagation through nonhomogeneous medium, where the propagation is represented by a wavefront arrival time for every point in the space (Sethian 1999). When the propagation velocity for a point in the space is defined by F (which is always non-negative), the arrival time function T is determined by the solution of equation



Ti+1,j − Ti,j Δx (4)

+y T Dij

environments provided the space segmentation is done in advance. So only the path search, corridor calculation, and funnel algorithm need to be computed in real-time. The main drawbacks of the two segmentation approaches are non-smooth paths which consists of straight line segments. Funnel method yield smooth paths, but tend to follow the edges of the obstacles. Therefore an alternative way of path planning was considered, which is described in the next section.

|∇T |F = 1

Ti,j − Ti−1,j Δx

(2)

Figure 4: A point in the grid and its neighbours

121

3.1. FMM method and path planning The method can be easily applied to shortest path planning as the time of arrival into a point in the space is always the earliest possible time, and the known obstacles are simply taken into account during the calculation of the wavefront propagation. It suffices to set the propagation velocity to zero for any point inside the obstacle, which prevents wavefront from entering. Once the arrival time function is calculated, the shortest path can be reconstructed by following the largest gradient. This can be done by simple Euler’s method or by more precise Heun’s (modified Euler’s) method. The path can be calculated in both directions, i.e. from the wavefront starting point into any target point in the space or reversed, from a set of starting points to a fixed target point, which is the wavefront starting point. Such an example for an environment of 2.2 m × 1.8 m with 10 A4 size obstacles is shown in Figure 5. The time arrival function and the calculated paths from 5 points are shown in 3D view first, and then the projection to the x-y plane is shown. Note that the time arrival function value is not defined for the interior of the obstacles, but was fixed at 300 for visualization purposes.

Figure 6: Example of smooth path planning by FMM method and a fixed target point shown above the size of environment was 2.2 × 1.8 m, grid points were 0.01 m apart, the EVT calculation took approx. 12 s, but fortunately needs to be performed only once for a fixed configuration of obstacles. FMM algorithm takes around 3.5 s and then backtracking by Heun’s method another 0.01 to 0.1 s, depending on the step size. All times were obtained on a 2.4 GHz PC by implementation of algorithms in Matlab m-code. The code is not optimal, nevertheless, the listed computation times indicate that the use of the method in real-time is not feasible for any complex environment. The computational time for the calculation of the arrival time function strongly depends on the length of the propagating wavefront. It is relatively short when the wavefront can only advance in a narrow corridor. This implies the feasibility of a combined approach, where segmentation of the space is first applied to coarse initial path planning, then a suitable corridor is defined within which a finer path planning is performed by the FMM method. The corridor used for FMM can be exactly the same that is used by the funnel algorithm. Given a set of obstacles, the EVT transformation is computed first. This could be done for points within the corridor only, but also the obstacles away from the corridor should be taken into account. Due to large computing time of the EVT it is more convenient to compute it for all the points in advance and then mask the points outside of the corridor by setting their EVT value to zero. More precisely, the EVT transformation operates on a bitmap image and calculates the Euclidean distance from every pixel to the nearest pixel carrying a specific property, in our case a pixel that belongs to the obstacle. This is a concept related to Voronoi graph, with the difference that the information is attached to every picture element and not only the points on the edges of the Voronoi cells. A number of EVT algorithms calculate the Voronoi diagram as an intermediate step. By using the EVT information during the FMM based path planning the planned path is automatically pushed away from obstacles and smoothed at the same time because the path does not follow the obstacle edges as it may happen with funnel algorithm. In (Garrido, Moreno and Blanco 2009) the authors suggest to include the distance to the obstacles as a velocity parameter when calculating the wavefront propagation. This makes analogy to propagation of light ray through the medium with non-homogeneous refractive index. The light in such a medium is not refracted but bent smoothly. For

Figure 5: Example of shortest path planning by FMM method and a fixed target point

3.2. Smooth path planning and FMM method The arrival time function calculated by FMM method serves as a potential field which directs the moving object toward the target point. If additional information is included into this field, this can influence the calculated path. Such an information is the information about the distance to the obstacles. This way the path may be pushed away from the borders of the obstacles. The distance to the obstacles can be included by the use of Extended Voronoi Transformation (EVT), which is used in digital image processing (and called Distance Transform therein). If EVT information is included into FMM based path planning, the obtained paths are driven away from the obstacles and smooth at the same time (Garrido, Moreno and Blanco 2009). An example is shown in Figure 6. 4. THE COMBINED APPROACH While in general the FMM method gives better results in terms of optimal lengths of the planned paths and can also be adapted to smooth down the paths as described above, its limitation in the substantial computational burden compared to segmentation based methods. E.g. in the example

122

the path planning the distance is not used directly, but is transformed analogously to electric potential by a logarithmic function, e.g. (Garrido, Moreno and Blanco 2009): Fi,j = c1 log(Ri,j ) + c2

(9)

where Ri,j stands for the distance of point (i, j) to the nearest obstacle. For the purpose of the presented study this function was further modified to  Ri,j Ri,j + c3 ) + c2 , log( Δx + c3 ) > 0 c1 log( Δx Fi,j = Ri,j 0, log( Δx + c3 ) ≤ 0 (10) Equality Δx = Δy is assumed, and the proposed function enables balanced results with various grid sizes. Weights ci define the shape of the velocity field and consequently influence the shape of the derived path. In particular, c1 is related to the velocity of an object moving along the path, c2 influences the curvature of the path, and c3 can be used to define a safety margin around obstacles. At the same time the function is consistent with the approach used by FMM, where the interior of obstacles is indicated by adjoining the corresponding grid elements by zero velocity of wavefront propagation. k may be added, e.g. Additional distance terms Fi,j measuring the distance to some predefined environment borders. These are calculated in a similar manner as (10). The velocity field given by EVT is then calculated as a sum of partial distance functions. The corresponding weights ckl may be used as additional design parameters. With the described calculation of EVT the proposed path planning method can be summarized in the following steps:

Figure 7: Example of the path smoothing in triangular space segmentation – arrival time function within the corridor and resulting path obtained by the FMM algorithm. parameters of (10) can be used to trim the shape of the calculated trajectory. Some preliminary results regarding computational complexity of the proposed method are shown in Tables 3 and 4. The first table shows the results obtained by the combination of quadtrees and FMM method while the second table shown the results obtained by the combination of constrained Delaunay triangulation and FMM method. The number of obstacles Nobst is shown, the min/max size of segmentation cells, raster size used by FMM method, computing time of EVT transformation, and the remaining columns show the computation time needed for adjusting EVT to the given corridor (tEC ), for calculation of arrival time function within the corridor by FMM method (tF M M ), and for calculation of the path by following the largest gradient by Heun’s method (tH ). By the last part, the integration step size was chosen according to the raster size: step = 0.1/raster. Similarly as before, the last three computation times refer to the average of eight consequent runs of the path planning algorithm with different start and target points.

• The QT or CDT segmentation of the environment with known obstacle positions is calculated. • The environment is covered by a grid of points and the EVT transformation is calculated interpreting every point of the grid as a pixel of the image. • Start and target points are chosen and a corresponding path is planned by an A* type search algorithm. As a result, a path and a corresponding set of segmentation cells are obtained, the cells defining a corridor surrounding the calculated path.

Table 3: Combined method computation times - quadtrees Environment size 10 x 10 m, maximal obstacle size 1 x 1 m Nobst MinDim raster t EVT [s] tEC [s] tFMM tH [m] [m] [s] [ms] 100 10/128 0.02 612 9.4 1.0 21.5

• The information about the corridor boundaries is used to adjust EVT transformation by setting the calculated values of the EVT to zero for any point outside the corridor. • The newly obtained EVT is used to parametrize the FMM based path search algorithm to obtain a smooth suboptimal path within the corridor.

Table 4: Combined method computation - triangulation Environment size 10 x 10 m, maximal obstacle size 1 x 1 m Nobst MaxDim raster t EVT [s] tEC [s] tFMM tH [m] [m] [s] [ms] 0.05 96 2.5 0.42 4.1 100 1 0.02 612 12.1 2.6 22.5

If new path has to be calculated for another set of points, it suffices to repeat only the last three steps of the proposed procedure. An example of the path planning results for the CDTA*-FMM based method is shown in Fig. 7. Configuration of the obstacles as well as starting and target points are the same as in Fig. 2 and Fig. 3. Three paths obtained by varying weight c2 are shown, which demonstrates how

4.1. Experimental results The proposed approach operation was also validated experimentally. The experiments were performed on the small,

123

two-wheeled, differentially driven mobile robot in the environment 2.2 × 1.8 m where obstacles are implemented by A4 paper sheets which are randomly placed in the environment as shown in Fig. 8. The robot measures 7.5 × 7.5 × 7.5 cm and weighs 0.6 kg. It contains a C167 microcontroller running at a 20MHz clock, a 12-V battery supply, two powerful DC motors equipped with incremental encoders (512 pulses per revolution), and a gear reduction head. The reference path for the robot is calculated using the presented algorithm with A* path optimization on QT space representation (Fig. 1), corridor determination and funnel algorithm or FMM inside the corridor. The path planning part of the algorithm is implemented in Matlab environment while the path-tracking control part is implemented in C++ environment. robot current pose is obtained by the overhead camera with 33 ms sampling rate. The determined corridor and the shortest path within the corridor obtained by the funnel algorithm are given in Fig. 9. The path is piecewise linear, which enables good tracking within the segments but could result in tracking errors at the junctions of the linear segments. The overall tracking performance can be improved by the proposed combined planning approach. Therefore the path within the corridor is determined by FMM taking into account distance to the obstacles as well as distance to the borders of the corridor. The feasibility of the planned path is analysed by calculation of the linear and angular velocities as well as tangential and radial accelerations. For the used mobile robot the maximal velocity is constrained at 1 m/s while the analysis in (Lepetic, et al. 2003) shows that the path can be tracked when the maximal tangential acceleration remains below 2m/s2 and maximal radial acceleration is below 2m/s2 . The velocity profile is calculated directly from the points obtained by Heun’s method. The obtained points are interpreted as points on the path sampled at equidistant sample times. The linear and angular velocities are then calculated by backward difference approximation of

1.8

1.6

1.4

1.2

y [m]

1

0.8

0.6

0.4

0.2

0 0

0.5

1 x [m]

1.5

2

Figure 9: Example of the optimal path (thick line) determination inside the corridor (thin lines) obtained by A* path finding algorithm and QT space representation. 0.65 0.6

velocity (m/s)

0.55 0.5 0.45 0.4 0.35 0.3 0.25

0

1

2

3

4 time(s)

5

6

7

8

Figure 10: Velocity profile of the mobile robot for the planned trajectory. expressions  x˙ 2r (t) + y˙ r2 (t)

(11)

x˙ r (t)¨ yr (t) − y˙ r (t)¨ xr (t) x˙ 2r (t) + y˙ r2 (t)

(12)

vf f (t) = ωf f (t) =

The velocity profile obtained this way is shown in Fig. 10. The velocity remains below the limit of 1 m/s. Similarly, the tangential and radial acceleration are calculated by approximation of at (t) =

dvf f (t) dt

ar (t) = vf f (t)ωf f (t)

(13) (14)

Fig. 11 shows that the calculated accelerations remain within the limits given by (Lepetic, et al. 2003), which proves the feasibility of the planned path. To further validate the feasibility of the approach the tracking of the planned paths was also tested experimentally. To drive the robot with differential kinematic on the

Figure 8: Experiment setup. The robot needs to travel between the obstacles (A4 paper sheets) from right corner to the left corner.

124

1.8

2 1.6

1.5

1.2

0.5 1

0

y [m]

tangential acceleration (m/s2)

1.4

1

0.8

í0.5 0.6

í1 0.4

í1.5 0.2

í2 í4

í3

í2

í1 0 1 radial acceleration (m/s2)

2

3

4

0 0

Figure 11: Radial and tangential acceleration of the mobile robot for the planned trajectory.

0.5

1 x [m]

1.5

2

Figure 12: Mobile robot trajectory tracking experiment (thick line). The reference path (thick dotted line) is obtained by Funnel algorithm inside the determined corridor (thin lines).

planned path the control law (Oriolo, Luca and Vandittelli 2002) is as follows

1.8

v(t) = vf f (t) cos eθ (t) + kx ex (t) θ (t) ey (t) + kθ eθ (t) ω(t) = ωf f (t) + ky vf f (t) sineθe(t) (15) where ky is a positive constant, while kx (·) and kθ (·) are continuous positive bounded functions. The trajectory tracking errors e(t) = [ex (t), ey (t), eθ (t)] = R(qr (t) − q(t)) are expressed in robot local coordinates (fixed to the robot center and x axis is in the robot forward direction) where R is the rotation matrix between global and robot coordinate frame, qr (t) is the reference robot pose on the planned path and q(t) is the current robot pose. The feedforward inputs are calculated from the planed reference trajectory by (11) and (12). In Fig. 12 the robot was controlled to follow the optimal planned (sequentially linear) path obtained using the funnel algorithm inside the corridor. Robot stats in the upper right corner (x = 2.06m, y = 1.69m) and ends in the lover left corner (x = 0.14m, y = 0.11m) which coincides with situation in Fig. 1 . Robot feedforward inputs were selected as vf f (t) = 0.4 m/s and ωf f = 0 1/s and gains in Eq. 15 are kx = 4, ky = 30 and kθ = 4. It can be seen that robot has bigger tracking error when the path changes discontinuously. At higher tracking errors it become possible for the robot to leave the safe corridor and can even hit some obstacle. Better tracking performance is obtained in Fig. 13 where the smooth reference path is obtained using FMM inside the corridor. The feedforward inputs vf f (t) and ωf f (t) are calculated from the smooth reference path using equations (11) and (12). The robot can follow the reference path with much smaller tracking error therefore the possibility that the robot escapes the safety corridor and hit some obstacle becomes much smaller. In the presented experiments the environment is partitioned in 813 cells where the smallest cell is approximately the size of the robot. On 2.4 GHz PC the QT algorithm takes approximately 0.4 s, optimization with A* algorithm takes around 15 ms, funnel algorithm inside the corridor

1.6

1.4

1.2

y [m]

1

0.8

0.6

0.4

0.2

0 0

0.5

1 x [m]

1.5

2

Figure 13: Mobile robot trajectory tracking experiment (thick line). The reference path (thick dotted line) is obtained by FMM inside the determined corridor (thin lines). takes 7 ms while the FMM (with raster Δx = Δy = 1 cm) takes some 3.5 s for the whole image and approximately 350 ms for the calculation inside the corridor only. 5. CONCLUSIONS AND FUTURE WORKS The results show the feasibility of the proposed path planning method and confirm the significant reduction in computation times of FMM in the corridor compared by the FMM for the whole search space. Due to rather small size of the minimal cells of the QT a small FMM raster size must be chosen in this case to avoid numeric problems when following the gradient in the narrow parts of the corridor. This indicates that smaller granulation of QT is not necessary advantageous. Besides, the larger granulation yields more room in the corridor for FMM to smooth down the planned path. On the other hand, larger granulation prevents the path planning algorithm to draw the

125

planned path over the narrow passages among obstacles. The resulting paths may therefore not be optimal in terms of their lengths. Presented path planning approach is evaluated by experiments on differential mobile robot where the proposed corridor constrained FMM approach results in smooth planned paths which enable good trajectory tracking results. Future work will try to increase the robustness of the algorithm in exceptional cases, and further investigate the parametrization of the velocity field with respect to kinematic constraints. The study how to upgrade the proposed approach for dynamic environments where A* is used only for initial planning and D* for each path replanning will also be performed.

AUTHOR BIOGRAPHY ˇ GREGOR KLANCAR received B.Sc. and Ph.D. degrees in electrical engineering from the University of Ljubljana, Slovenia in 1999, and 2004, respectively. His research interests are in the area of fault diagnosis methods, multiple vehicle coordinated control and robot soccer related problems. His Web page can be found at http://msc.fe.uni-lj.si/Staff.asp. ˇ ˇ C ˇ received B.Sc., M.Sc. and Ph.D. deGASPER MUSI grees in electrical engineering from the University of Ljubljana, Slovenia in 1992, 1995, and 1998, respectively. He is Associate Professor at the Faculty of Electrical Engineering, University of Ljubljana. His research interests are in discrete event and hybrid dynamical systems, supervisory control, planning, scheduling, and industrial production control. His Web page can be found at http://msc.fe.uni-lj.si/Staff.asp.

REFERENCES Baerentzen, J.A., 2001. On the Implementation of Fast marching Methods for 3D Lattices, Technical Report IMM-REP-2001-13, Technical University of Denmark. Botea, A., M¨uller, M., Schaeffer, J., 2004. Near optimal hierarchical path-finding, Journal of Game Development, 1, 7-28. Davis, I., 2000. Warp speed: Path planning for star trek: Armada, AAAI Spring Symposium on AI and Interactive Entertainment, AAAI Press, Menlo Park, CA. Farag, A.A., and M.S. Hassouna, 2005. Theoretical Foundations of Tracking Monotonically Advancing Fronts Using Fast Marching Level Set Method, Technical report, University of Louisville. Garrido, S., L. Moreno and D. Blanco, 2009. Exploration of 2D and 3D Environments using Voronoi Transform and Fast Marching Method, Journal of Intelligent and Robotic Systems, 55, 55–80. Ferguson, D. and A. Stentz, 2006. Using Interpolation to Improve Path Planning: The Field D* Algorithm, Journal of Field Robotic Systems, 23(2), 79–101. Hongyang, Y, H. Wang, Y. Chen, D. Dai, 2008. Path Planning Based on Constrained Delaunay Triangulation, Proceedings of the 7th World Congress on Intelligent Control and Automation, June 25 - 27, Chongqing, China. Kallmann, M., 2005. Path Planning in Triangulations, Proceedings of the Workshop on Reasoning, Representation, and Learning in Computer Games, International Joint Conference on Artificial Intelligence (IJCAI), Edinburgh, Scotland, July 31, pp. 49-54. Koenig, S., M. Likhachev and D. Furcy, 2004. Lifelong Planning A*, Artificial Intelligence, 155(1-2), 93–146. Sethian, J.A., 1999. Fast Marching Methods, SIAM Review, 41(2), 199–235. LaValle, S.M., 2006. Planning Algorithms, Cambridge University Press. G. Oriolo, A. Luca and M. Vandittelli, 2002. WMR Control Via Dynamic Feed-back Linearization: Design, Implementation, and Experimental Validation, IEEE Transactions on Control Systems Technology, 10(6), 835–852. M. Lepetic, G. Klancar, I. Skrjanc, D. Matko, B. Potocnik, 2003. Time optimal path planning considering acceleration limits, Robotics and Autonomous Systems, 45, 199–210.

126

MODELING AND SIMULATION ARCHITECTURE FOR CLOUD COMPUTING AND INTERNET OF THINGS (IOT) BASED DISTRIBUTED CYBER-PHYSICAL SYSTEMS (DCPS) Xie Lulu (a), Wang Zhongjie (b) (a) (b)

College of Electronics and Information Engineering. Tongji University. Shanghai. China. 201804 College of Electronics and Information Engineering. Tongji University. Shanghai. China. 201804 (a)

[email protected] , (b) [email protected]

Comparing with these technologies, CPS is much more complicated and with stricter system properties. CPS focuses on its real-time and resource-saved performances, being interactive, accurate, coordinate, intelligent, secure, robust, efficient and autonomous. Thus makes some existing technologies in traditional embedded systems; intelligent systems and hybrid systems can be optimized and employed into the modeling and simulation of DCPS. Being an auto-controlled real-time intelligent system with excellent performances, CPS technologies are intended to improve the quality of human life, promoting a harmonious living environment. The idea of CPS can be widely applied to various research areas, such as: smart vehicles in the area of transportation; smart grids in the area of infrastructures and energy; various smart devices and robots in medical treatment, biology, industries, agricultures. Thus constructed a distributed heterogeneous DCPS Environment, consists of either same kinds or different kinds of CPS subsystems and resources, with physical entities being monitored, coordinated, controlled and integrated by the computing and communication core, and the coupling among system components being manifested from the nano-world to distributed wide area systems, at multiple spatio-temporal scales.

ABSTRACT Distributed Cyber-Physical Systems (DCPS) consists of many spatio-temporal heterogeneous CPS subsystems and components, making the modeling, management and control of resources complicated. In this paper, firstly, objectives and research backgrounds are described; opportunities and challenges in modeling and simulation of coordinated and efficient DCPS are discussed. Secondly, abstraction and deployment of networked DCPS is presented and analyzed: 1) an Internet of Things (IoT) framework for ubiquitous and self-managed environment of DCPS has been proposed. Which can bridging autonomous DCPS nodes with the future internet; 2) an optimized Cloud-based open and reusable modeling and simulation architecture for the management, scheduling and control of large-scaled dynamic and heterogeneous resources and services is designed. Thirdly, a multi-model based hierarchical architecture for the modeling and simulation of an open and reusable DCPS platform is then proposed, with an experimental technology framework towards cooperative and robust DCPS being constructed and discussed in the last. Keywords: distributed cyber-physical systems, internet of things, cloud-based resource management, model based architecture

2. 1. INTRODUCTION Cyber-Physical System (CPS) integrates computing, communication and storage capabilities with monitoring and control of entities in the physical world dependably, safely, securely, efficiently and in real-time, (Cardenas, Amin and Sastry 2008). It connects the virtual information world with the physical world through the integration and interactions of Cyber and Physical components, Lee (2008). It requires close interaction between the cyber and physical worlds both in time and space, and the interaction among components are autonomous managed. Some basic theories of CPS are derived from the integration of Distributed Real-time Embedded Systems; Wireless Sensor Networks; Networked Control Systems and Decision Support Systems, Bujorianu (2009).

BACKGROUNDS

2.1. United States Concepts of CPS are first proposed in USA, 2002. With attractive performances and gradually emerged importance, this technology has gained a lot of attention from both the governments and research institutions around the world since 2005. Many basic researches and trial applications for the single CPS have been aroused, mainly based on the architecture of embedded autonomous systems and hybrid systems. There are some novel research intentions and meaningful applications. Campbell and Garnett (2006), proposed the ideas of CPS Environments and CPSs Sensor Grids, intending to sense the physical environment in different granularities, and with dynamic and various disturbance elements, for

127

the evaluation and simulation of systems’ overall abilities and performances. Edward (2009), combining the distributed embedded software with dynamic physical models. An efficient programmable temporal semantics abstraction based synergy language has been proposed, for hardreal time control of DCPS. (Ilic and Xie 2008; Zhang and Ilic 2009), study the modeling and control methods of Distributed Smart Grids. Support Vector Machine (SVM) and MarkovState based control and prediction model have been proposed for the modeling and control of Smart Grids, with the optimized scheduling and dispatching among distributed electric power stations. Thiagarajan and Ravindranath (2009), aim to providing a Trust-based intelligent vehicle navigation system: VTrack, which also is a typical application of DCPS. They have done a lot of work in the coordination and stability control between cars and between car and the traffic signals. Correll and Bolger (2010), take part in the research of Distributed Robot Garden. With which a team of robots can take care of the tomato plants in the garden autonomously, through distributed sensing (each plant is equipped with a sensor node monitoring its status), navigation, manipulation; wireless networking and coordination. Asynchronous mixed-signal modeling and verification methods and tools have been suggested by Thacker and Myers (2010), to study the battery-based DCPS, as for the optimized coordination, scheduling and simulation between DCPS nodes. Chen and Ding (2010), use Grammatical reasoning models, combing with distributed Multi-Agents simulation algorithm and other basic models and symbols, to abstract and achieve synthesis and effective control and learning strategies between DCPS. To encourage more researchers to take part in the researches related to DCPS, the research of selfinteracted, coupled, collaborative and integrated platform of DCPS has been listed into the NSF’s important research agenda in 2011. The platform is expected to realize the intelligent coordination and interaction among different DCPS components, and to improve the overall performance of DCPS,

and modeling and control of complex systems, which are beneficial to the research of DCPS.

2.2. Europe European scholars focus primarily on the structure and theoretical foundations for the design, modeling, and implementation, performance and applications of DCPS. Intelligent modeling algorithms have been proposed for the control and optimization of DCPS, such as the ant colony, immune and hormonal algorithms integrated methods, Rammig (2008). International project called “RoboEarth” attempts to let the robots to share information with each other and to store and update their knowledge in a self-managed manners, (Zweigle, Andrea and Haussermann 2009). Besides, Europeans have done a lot of work in intelligent electronic systems; SCADA systems; integration of multiple components;

2.5. Opportunities and challenges Taking into account the existing studies, there are many meaningful researches in the modeling and control of DCPS applications. While the technologies for the overall management and control of DCPS are currently scarce; most of the existing models and algorithms are applied only to the fixed applications; most of them just focused on some parts of a specific CPS, lacking the analysis and modeling, control or scheduling strategies of the overall DCPS environment, with a large amount of physical equipments, computing devices and communication resources are inactivated most of the time, with low utilization rates.

2.3. Japan, South Korea, Australia In Japan and South Korea, DCPS started to get concerned around 2008, Easwaran and Insup (2008), applications and software frameworks of DCPS have been studied, such as the automated integration of embedded objects and computing equipments under hybrid communication networks. Modeling and control experiments of intelligent robots with CPS properties have been studied. Researchers in Australia have also launched many interesting researches in Smart Grids, and Smart Cars, Lyster (2010). 2.4. China In China, the research of DCPS has been proposed and selected as one of the major development directions by the High-tech Research and Development Program of China and NSFC since 2009. There are some exploratory works. Xia and Ma (2008), have made some progress in the QoS and real-time high performance control of medical DCPS, based on feedback control between medical CPS nodes. Zhang (2010), try to design DCPS with networked clouds, high confidence middleware and information exchange technologies. Zhao (2010), announced a cut based on geometric topology control algorithm for the energy balance in DCPS, using network topology control algorithm to improve the energy efficiency and the robustness of the system, with the optimizing of MAC layer protocol to achieve lower energy consumption, reduce transmission delay and optimize network performance. Xiao and Yu (2010), proposed a series of control methods to improve the reliability of DCPS, using a Petri-nets based Case modeling. (He 2010; Ma 2010), proposed a "sensing-control network" concept, for CPS, focuses on the theories research of a single CPS, involving mathematical modeling, analysis, verification methods and theorybased research of CPS, to address key issues encountered, such as real-time, cross-layer, composability, predictability, dynamic evolution.

128

x

As for the complexity of the environment and resource heterogeneity, the coordination and collaboration of the CPS components will greatly affect the real-time property and the overall performance of DCPS. Compared with the networked control technology, embedded technology and the internet of things, DCPS environment has better coordination and collaboration mechanisms, and is capable of achieving a much efficient and real-time performance. Its ultimate goal is to perceive the environment and resources accurately, monitor and coordinate different components, and make real-time decisions based on the feedback of DCPS performance to implement appropriate behavior and actions without any manual supervision. There is a large amount of researchers like Cardenas and Sastry (2009), who have found that although there are already many useful researches on the communication and computing security of the networked DCPS, researches of robustness are somewhat limited. Especially that DCPS environment hold series of dynamic and heterogeneous coupled complex CPS subsystems and applications, it is difficult to discover the unexpected events, and it will be hard to guarantee a safe and stable environment for the DCPS. It is important to design and construct a stable and efficient architecture to have the abilities to avoid the cascading failures and malicious attacks, and be prepared even in an uncertain and unmanned environment. There are already some useful abstractions and architectures, such as the multi-model based real-time architecture proposed by Lee (2008), and the feedback control based architecture by Xia and Ma (2008), the spatio-temporal event based architecture by Tan and Mehmet (2009), ect. We plan to integrate the features of these architectures using the multi-model based layered management and control architecture. There are several research problems that should be especially considered in the management and control of DCPS: x x

x x

x x

How to realize the ubiquitous hard real-time management and control of different DCPS applications over wired and wireless communication environments covering wide areas and different networks? How to detect and predict potential threats through the verification of the models and simulated results? How to quantize and estimate the variety of the environment, to reduce the errors among the lab simulation results and the actual control and action results under undertrained and dynamic environment?

We suggest an universal management and control architecture for the existed CPSs, with the abilities to support the building and testing of newly constructed CPS models and even new applications. That the related CPSs’ knowledge, models and hardware resources among various application areas can be collected, shared, analyzed and reused, with indexes built according to different properties and events concerned. And all of the DCPS resources can be remotely located and managed, and can be customized dynamically integrated, optimized and updated for the possible future uses. Architecture for DCPS has been proposed and designed in this paper. Based on the modeling and interaction of each technological layer, the architecture aims to construct a cooperative and robust modeling and simulation framework for the sustained management and schedule of DCPS components and resources based on the cloud computing and cloud simulation concept (Li, Chai and Hou 2009). Using Complex Network based topology abstraction and behavior prediction to support the research of stability and robustness of CPS under perturbations or uncertain environment, and consider the autonomous interaction and selfcoordination among CPS components and (or) CPS subsystems through Multi-Agents and Petri-net, with the intelligently optimized simulation and verification results, system models can be updated and to be selfadapted, so that the efficiency of resource utility and system performance can be improved.

How to deal with the huge information and resources in the large and space-time heterogeneous DCPS environment? How to build the self-adapted and intelligent learning or understanding of the optimized modeling methods and algorithms for the modeling agents and unified model interfaces or services under dynamic environment with various requirements and restrictions? How to improve the efficacy and robustness of the schedule and control methods assuring both real-time and low energy consumption? How to locate and schedule system resources intelligently and robustly to achieve a stable and sustained DCPS environment under restricted ability constraints?

3.

ABSTRACTION OF NETWORKED DCPS ENVIRONMENT Based on the concept, structure and composition of DCPS, announced by Lee (2008), Rajhans (2009), Chun (2010). DCPS environment can be considered as a set of distributed decision support systems, which combine real-time embedded methods with the networked coordination and control technologies. Networked DCPS environment (as shown in “Figure 1”) bridges and associates the cyber world of computing, communication, and control with the physical world, Vincenzo (2007). It contains both the physical and computational components, with interaction between physical layer and information layer under the network communication environment to make decisions.

129

enabled objects and RFID tags, living things are not included, and nodes in IoT aren’t auto-controlled; they can’t control each other; and without dynamical selfadaption. 2) Applications of DCPS require bi-directional, close-looped, real-time processing between the cyber world and the physical world with end-to-end QoS determinism and predictability, Dillon and Zhuge (2011). While IoT is not required to be hard real-time restricted. 3) Furthermore, existing solutions do not address the scalability requirements for a future IoT, they provide inappropriate models of governance and fundamentally neglect privacy and security in their design. IoT aims at realizing the seamless integration of heterogeneous IoT technologies into a coherent architecture, which is strongly related to the development of the future internet environment required by the characters of DCPS. With the unified and standard communication protocols and frameworks of IoT, it will be easier to provide a unified CPS application development environment to support and promote a much more rapid and cost-effective CPS application development.

Figure 1: Abstraction of Networked DCPS Environment DCPS environment is a self-managed ubiquitous networked environment; components in DCPS environment can be self-organized, self-adaptive, selfoptimized and self-configurable, self-protected and selfhealed to make an autonomous controlled DCPS environment. In the DCPS environment, one needs to consider not only the micro-level interaction and integration between the cyber and physical components, but also the macro-level interaction and collaboration between the different CPS subsystems. Taking into account the complexity of the environment and resource heterogeneity, effective coordination and collaboration of the CPS components will promote the real-time response and even the overall performance of DCPS. Compared with the networked control technology, embedded technology and the internet of things, DCPS environment has better coordination and collaboration mechanisms, and is capable of achieving a much efficient and real-time performance. Its ultimate goals are to perceive the environment and resources accurately, monitor and coordinate different components, and make real-time decisions based on the feedback of DCPS performance to implement appropriate behavior and actions without any manual supervision. 4.

4.2. Cloud computing Events and demands in DCPS are analyzed and described as semantic information and service oriented structures. The amount of the information is always huge, real-time allocation, management and control of spatio-temporal heterogeneous resources can’t be realized by the traditional scheduling algorithms. Cloud computing embraces cyber-infrastructure, and builds upon decades of research in virtualization, distributed computing, utility computing, and more recently networking, web and software services, Vouk (2008). It delivers infrastructure, platform, and software (applications) as services, which are made available as subscription-based services in a pay-as-you-go model to consumers. These services in industry are respectively referred to as Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS), Calheiros and Ranjian (2009). Five essential characteristics of Cloud computing are: broad network access; measured services; on-demand self-services; rapid elasticity; resource pooling, Cloud computing implies a service-oriented architecture (SOA), reduced information technology overhead for the end-user, great flexibility, reduced total cost of ownership, on-demand services and many other things. Therefore, it can be used in the management and optimization of resources.

RELATED WORKS

4.1. Internet of Things (IoT) Internet of Things (IoT) consists of smart devices which are ubiquitous and will be constantly connected to the public Internet. It can be definite as “Things having identities and virtual personalities operating in smart spaces using intelligent interfaces to connect and communicate within social, environmental, and user contexts”. “Things” refers not only to the uniformed devices, but also the heterogeneous devices with different application areas, functions, editions and core technologies, (Dillon and Zhuge, 2011). All these devices can be included into a common community belongs to the same communication environment to identify each other with seamless integrations. Comparing IoT with DCPS, there are some major differences: 1) DCPS is more complicated in modeling and control. Devices in IoT just include sensors, actuators,

4.3. Combination of IoT and Cloud computing There are many design challenges in the research of IoT: such as limited bandwidth; low memory; low transmission power; low bit rate; low computational power; low throughput; low lifetime and low utility ratio. Most of these problems result from constrains of

130

limited computation and storage abilities and resources of the IoT nodes. Therefore, we consider combining the IoT technology with Cloud computing to promote the utility of the IoT resources. Because Cloud computing and IoT are both service-oriented dynamical technologies, IoT can be transferred into virtual resources, (Gyu and Crespi, 2010). Therefore, combination of Cloud computing and IoT is technically feasible. Besides, seamless integration of the ubiquitous IoT communication environment will make the information access in the clouds more convenient, cloud services can be visited by more people, accelerating the use ratio and the spread scale. Then useful services can be more attractive and valuable, making a higher profit; and useless services can be removed or updated, to improve the systems’ performances. 5.

According to the characteristics and description of IoT and Cloud computing, a hierarchical structure of self-managed DCPS environment based on the IoT communication framework and Cloud Services is presented in “Figure 2”. Detailed deployment of this structure is discussed below. 5.1. Physical Layer Physical layer usually contains a variety of different physical entities; these entities are cloud resources at the system level of IaaS. Resources and can be abstracted into different types of nodes: 1) Physical nodes: Simple physical nodes, such as sensors, actuators, sensor+actuators, infrastructures, and end devices. 2) Computation or communication nodes: Modules or methods with limited computing, communication or storage features. 3) Computation-Physical nodes: Software and hardware middlewares and cloud based services available for cyber-physical interactions. 4) Cyber-Physical nodes: Nodes are autonomously controlled and self-managed. They are intelligent nodes with both the ability of sensing and actuating, including humans. Because of the integration of information components and physical components, these nodes have a certain amount of computing, storage, and reasoning ability. Such as smart cars, smart meters, smart vehicles, robots etc. With different coupling strength and combination methods, these nodes can either be considered as a physical node with computing and decision-making capacity or be some computing nodes which can manage physical entities through communication and control. These nodes are heterogeneous in their distributions. Being mobile and adaptive, they have the ability of perception, memory, reasoning and learning, and have characteristics of living things. All kinds of nodes and resources in DCPS environment can be abstracted; simulated into graphs and network topologies, and be simplified with semantic services supported models of events detecting and behavior perception. Therefore, technology of intelligent agents can be used to model these DCPS nodes.

INTERNET OF THINGS (IOT) AND CLOUD COMPUTING BASED DCPS ENVIRONMENT Application Layer (Sevice and Requirements )

Smart Services (Social computing, Enterprise, Scientific, Ă ) ...

self-organizing self-optimization self-deployment self-decision self-healing self-adapting self-protection Ă

...

Wired and Wireless Networks

G

G

Wireless Sensor Network(WSN) G





Cloud computing and IoT based Self-Managed DCPS Environment

...

...

...

Association Link (Wirelessed Link)

Peer to Peer Link

Cloud Computation and Simulation Layer

Platform as a service (PaaS)

( Cloud Servers )

...

Cyber Association

Internet of Things Communication Layer

Application Layer Protocol

...

Wired Link

Infrastructure as a service Software as a service (IaaS) (SaaS) Virtual Resources ( Devices, Components, Objects, Ă)

Transport Layer Protocol

Coordination and Control Network Layer Protocol

... Hardware or Link Layer Protocol ...

Physical Layer ( Physical and Communication Infrastructures ) ( Local Computational Nodes )

G

Gateway

Link node (coordinator)

Router

ZigBee node

Environment

People ... ... ... End Devices

Figure 2: Structure of IoT and Cloud Computing Based Self-managed DCPS Environment. With the rapid developments of Internet of Things (IoT), cloud computing and semantic services, all of the physical entities can be ubiquitously connected, and can be intelligent and remotely identified, located, tracked, monitored and managed under the future internet. In the DCPS world, entities can see and sense the environment, can do some not very complicated computing, and will be auto-controlled, even with the intelligent to thinking, and can be prepared and self-adapted for instant events. The DCPS environment can be considered as a set of distributed decision support systems, which combine real-time embedded methods with the networked coordination and control technologies. Therefore, technologies of IoT combing with cloud computing, will construct a perfect ubiquitous computing and communication networked physical environment for the research and development of DCPS.

5.2. Internet of Things Communication Layer Communication layer consists mainly of network middleware, access equipments, standards, various communication protocols and routing algorithms. This layer communicates with the User-Level Middleware (SaaS) development in the Cloud programming over future internet, and being associated with the reasoning and calculation models in the computation layer, to realize the transmission of information and the management of the associated DCPS network nodes and ability restricted resources with spatio-temporal synchronization clocks. Communication of DCPS can be either wired or wireless. For wired communication, communication layer may contain high-performance server farms and complex industrial equipments. For wireless communication, communication layer may involve a

131

large number of wireless sensor nodes with constraints of size, cost and energy power. Therefore, it is hard for traditional methods to compute network delay and packet loss rate, and even harder to detect and handle cascade failures or malicious attacks with a timely response. In this paper, a complex network analysis method has been proposed to solve the above problems, in which the DCPS stability is guaranteed through flow based route optimization and key nodes analysis. Due to the diversity of communication, the locations of DCPS resources are distributed. And for the pervasive and flexible management or control of the DCPS environment, we proposed an open and pervasive cloud services based DCPS communication and control architecture, as shown in “Figure 3”.

With these services of adaptive resource storage and allocation; resources indexing and updating; virtual machine management and deployment, remote cloudbased computing, simulation and control may be realized. This architecture has four layers, three basic layers are Private Clouds, Public Clouds, and Public Cloud Firewall; and the additional layer is a coordinator layer, it aims to optimize the computing and simulation efficiency of DCPS resources. Private clouds contain all of the user's private resources; public clouds share and storage authorized and common resources; the cloud firewalls between the user terminals integrated safety standards for different systems and resources, make sure that the users can locate and use remote resources safely. Huge amounts of heterogeneous dynamical requirements and services are difficult to proceed, and are always time consumed. To conquer this problem, we proposed an idea of “Intermediary Cloud”. Intermediary clouds are responsible for the discovering, computing, analyzing, managing and deployment of domain specific services, resources or protocols of both private clouds and public clouds. Advanced computing intelligence algorithms can be used to accelerate the speed of information searching and decision making. Social elements will be considered and evaluated, such as the price of the services; behavior of customer; policies differences; ability of suppliers. Intermediary clouds can be considered as the social coordinators of DCPS. With intelligent coordination and management strategies, accuracy and efficiency of the resources allocation can be improve; scheduling and dispatching can be more efficient; along with the safety and robustness can be considered, and the overall performances of DCPS are guaranteed.

5.3. Cloud Computation and Simulation Layer Public Cloud Cloud Firewall Cloud platform

Intermediary Cloud

Cloud platform User

Key

Public cloud manager

Cloud platform

Cloud platform

Resource discovery

Resource recommendation

Middleware services

Private cloud manager

Private resources

Private Cloud

Resource dispatching Demand analysis

DCPS platform Clients access and management enviorment

Resource evaluation

Resource evaluation Capacity restriction

Intermediary Cloud manager Security Certificate

Service/model decoupling

Resource integration Demand analysis

Capacity restriction Service/model coupling

Management and control

Query engine

Rule engine

Uniform interface Behavior analysis Semantic and ontology

Figure 3: An Efficiency Optimized Cloud Computing and Simulation Architecture for DCPS Resources Computation layer of DCPS environment contains clouds of virtual resources, such as virtual data server, virtual communications server, virtual log server, virtual high-performance computing servers, virtual firewalls and security equipments. Judged by the functions, there are storage clouds, computing clouds and simulation clouds. Actual physical data and information of the devices and objects and people; are being virtualized by the Virtual Machine through its management and deployment. These components store historical data, related models and methods, and monitor the operation of the system in a real-time feedback control loop. Associated with the IoT communication environment, this layer will also improve the autonomic of the systems/components, and is associated with all kinds of social and environmental information and services in the future dynamic networks. Interactions of these two methods at this layer take part in the PaaS stage, covering a series of core middleware services: QoS Negotiation, Admission Control, Pricing, SLA Management, Monitoring, Execution Management, Metering, Accounting, Billing, and Virtual Machine Management and Deployment.

5.4. Application Layer Application layer interacted with user interfaces, and supplies all kind of smart services here. Standard cloud applications services including: Social computing, Enterprise, ISV, Scientific, CDNs, etc. Customers can design and order the services they are interested in. Designing of this layer should consider two main aspects: 1) Construction of industrial equipments and embedded components will cost enormous human and financial resources, so it is an important issue to make an effective and efficient reuse of current software and hardware resources. 2) A real physical environment often covers a number of different applications, so the bottleneck of DCPS application is how to abstract a variety of network topologies from the existing CPS subsystems to satisfy different requirements, and to realize resource sharing and reuse among CPS subsystems. Swarm algorithms and community discovery methods can be employed to analyze the above two problems. Based on the interaction mechanisms of multi-agents, coupling and decomposition methods can be used to improve the system performance. As a self-

132

managed DCPS environment should have the abilities of autonomy, coordination, real-time feedback, low energy consumption, and with high-performances. There are still many factors required be considered, such as heterogeneous composition of the DCPS components, uneven distribution of resources, dynamic or uncertain behavior of the components and systems, and environmental complexities.

Topological abstraction and optimization layer is used to abstract the CPS into a topological network. For this procedure, complex network analysis expert experiences will be employed. Graph theory can be used to optimize the topological network. Information fusion layer focuses on spatiotemporal information analysis and data mining. With demand analysis, environment perception and detection of the events, this layer achieves pre-process of information flow between different levels. Kalman Filter can be used to remove the redundant data. The information similarity extraction method based on semantic or some other computational intelligence models are used to process the data, and translate them into understandable knowledge. Modeling and optimization layer can be used to discover the events; to establish the models required in each layers; and to build services. For the openness and reusability of the platform, modeling should be smart and efficient. Models, methods and components in different layers should be associated and optimized. Model bases for DCPS should be constructed, and may include physical and mathematical models, graphs, computation intelligence and many models of hybrid systems. As for social elements related problems, models of complex networks, context-aware cognitive, agents and similarity simulation methods, wireless sensing network routing and relocating methods, may all be helpful. Optimization is based both on the modeling, simulation and evaluation. Behavior analysis and prediction layer aims at realizing a self-learning and self-adaptive intelligent DCPS environment. This layer is novel and special modular, considering for the dynamic and uncertain social elements in DCPS. The behaviors of CPS nodes and network evolution are predicted. Execution of these events will be monitored, recorded and analyzed. Analysis methods include semantic understanding methods, reasoning methods, network evolution analysis and key nodes discovery methods and flow equilibrium methods. Coordination and control layer receives behaviors of the CPS nodes from the behavior analysis layer, and makes a decentralized control to coordinate different nodes. With this layer, CPS is able to achieve better utilization of resource and performances. The coordination and control methods may include spatiotemporal synchronization, intermittent feedback, intelligent perception, perturbation control, pinning control and adaptive control, which will contribute to the rational and efficient management, schedule, dispatching and utilization of limited resources, fits the research of cloud services, internet of things and green computing. Evaluation layer mainly involves in system performance analysis and model verification. CPS performance includes efficiency, surviving period, accuracy (prediction, classification, sorting, etc.), correlation, security, robustness and vulnerability.

6.

MODELING AND SIMULATION ARCHITECTURE FOR AN OPEN AND REUSABLE DCPS PLATFORM As discussed above, it is necessary to build a generalized DCPS architecture that can support cooperation and coordination among heterogeneous CPS components and resources. Based on the structure of IoT and Cloud computing based DCPS computing, communication, and control environment, considering the key research problems in the research of DCPS, a multi-model based hierarchical architecture for the modeling and simulation of an open and reusable DCPS platform has been proposed and discussed in “Figure 4”.

Figure 4: Multi-model Based Modeling and Simulation Architecture for An Open and Reusable DCPS Platform Information layer consists of many databases, model bases, log bases and plan bases. This layer acquires, pre-processes and storage all kind of resource data and information sensed and collected from the DCPS environment. With all of these databases, there are huge information pools of different resources.

133

Model verification refers to verifiability, reliability, security, robustness, and reusability etc. Monitoring and decision making layer consists of both centralized control and decentralized control. Centralized control mode is mainly for the cooperation and coordination of the nodes and resources. Decentralized control mode mean that every node is autonomous controlled, and so centre node caused sudden failures and cascading failures can be avoided, which is especially useful for emergencies. Therefore, elements such as the energy consumption and life-cycle of DCPS nodes; computation complexity or task importance; economic profit of services; government policies and other factors should be considered to gain a balance or creating a better control mode.

decoupled to achieve network and flow equilibrium assignment, for example, the equilibrium methods, the ordinary differential equation (ODE) models, recursive algorithms, and the complex network coordination and synchronization control methods.

Complex Network Coordination and Synchronization

Control Theory, Evaluation Methods

7. TECHNOLOGY FRAMEWORK This technology framework in “Figure 5” is an experimental architecture particularly designed for a DCPS taking “cooperative” and “robust” as the most important performances. This design is based on the modeling and simulation architecture of the unified DCPS platform. Methods of complex network and multi-agents are used to build and analyze the DCPS’s topology network. Topology abstraction of DCPS is constructed with three main elements: Nodes, Links (relationship between nodes), and Flow (information streams, Cheng and Wang (2010), tasks, and event flow through the nodes and edges). Description of the technique steps are as follows: (1) The environment state is firstly perceived. After the state noise being removed by the data fusion methods such as the Kalman filter, the data will be normalized. Then the data ontology and semantic presentations can be built. (2) For different CPS applications, analyze and quantify the system requirements, detect the possible events, acquire and store the knowledge for the control and modeling. (3) According to the quantified requirements and complex network construction rules, the similarity analysis method can be used to define the nodes and links, and abstract the DCPS environment into a topological space. (4) In order to increase the robustness of DCPS, the “Hub” node in the topology network, Ulieru (2007), will be detected and analyzed with the statistic characteristics of complex network centrality, such as degree distribution, power-law, closeness, betweenness centrality, random walk betweenness centrality. (5) Analyze the components’ clusters and the community composition in DCPS environment; reconstruct the CPS subsystems if necessary. (6) With dynamical collaboration mechanism and flow evolution control methods, analyze and control the network flow and make a macro-view over the whole DCPS environment. (7) Analyze and control the link flow and make a micro-view over the resource distribution and the node utilization. Several methods can be combined or

Figure 5: An Experimental Technology Framework (8) Design and modeling the agents for DCPS with different functions and attributes. The Agents will be autonomous controlled, and can be designed based on a BDI (Belief, Desire, Intention) framework, capable of storing, reasoning, computing and communicating, and should have open and reusable services, middlewares and device interfaces. (9) Build a Petri-nets-based behavior analysis model for multi-agents system by combining and extending Petri-nets. Analyze the cooperation and interaction mechanism between different agent-groups. (10) Try to combine the Petri-nets model with the complex network coordination and synchronization control methods to achieve the coordination of the nodes and the efficient scheduling of the tasks. (11) With the system verification methods, test and analyze the system stability, performance and the operation time. Results can be used to optimize the related models. (12) Construct a networked DCPS supported Intelligent Human-Computer Interaction (HCI), to test the DCPS platform, and to design customizable services.

134

Campbell, R.H., Garnett, G. and McGrath, R.E., CPS Environments. NSF Workshop on Cyber-Physical Systems, October 16-17 Austin, TX. Cardenas, A.A., Amin, S., and Sastry, S., 2008. Secure Control: Towards Survivable Cyber-Physical Systems. Proceedings of ICDCS Workshops, pp. 495-500. June 17-20, Beijing, China. Cardenas, A., Sastry, S. and others., 2009. Challenges for Securing Cyber Physical Systems. Workshop on Future Directions in Cyber-physical Systems Security, July 23, DHS. Carney, D., Cetintemel, U. and others., 2002. Monitoring streams: a new class of data management applications. Proceedings of the 28th international conference on Very Large Data Bases, pp 215-226, August 20-23, China. Chen, Y., Ding, X.C., Stefanescu, A. and Belta, C., 2010. A Formal Approach to Deployment of Robotic Teams in an Urban-Like Environment. Proceedings of 10th International Symposium on DARS, pp 14 . Nov 1-3, Lausanne, Switzerland. Cheng, L., Wang, Z.J., 2010. A Stream-based Communication Framework for Network Control System, Proceedings of the CISP2010-BMEI2010, Vol 1, pp 2828-2833, Oct 16-18, Yantai.China. Correll, N., Bolger, A., Bollini, M. and others., 2010. Buildng a Distributed Robot Garden. Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp 219232 . Oct 11-15, St. Louis.USA. Dillon, T., Zhuge, H., Wu, C. and Singh, J., 2011. Web of things framework for cyber-physical systems. Concurrency and Computation: Practice and Experience, 23 (9) : 905–923. Easwaran, A., Insup, L., 2008. Compositional schedulability analysis for cyber-physical systems. SIGBED Review, 5 (1): 11 -12. Gyu, L., Crespi, N., 2010. Shaping Future Service Environments with the Cloud and Internet of Things: Networking Challenges and Service Evolution, Leveraging Applications of Formal Methods, Verification, and Validation, Lecture Notes in Computer Science, 6415: 399-410, Heidelberg: Springer Berlin. He, J.F., 2010. Cyber-physical Systems. China Computer Federation Comunication, 6 (1): 25-29. (in Chinese version) Ilic, M.D., Xie, L. and Khan, A.U., 2008. Modeling Future Cyber-Physical Energy Systems. Proceedings of IEEE Power Engineering Society General Meeting, pp 1 - 9, July 20-24, Pittsburgh. Ingeol, C., Jeongmin, P. and others., 2010. Autonomic Computing Technologies for Cyber-Physical Systems. Proceedings of International Conference on Advanced Communication Technology, ICACT, pp1009-1014. February 7-10, South Korea. Lee, E.A., 2008. Cyber physical systems: Design challenges. Proceedings of 11th IEEE Symposium on Object Component Service-Oriented Real-Time

8. FURTURE WORK Smart Grids is a typical application of DCPS. The proposed architecture will be deployed into the analysis, management, schedule and optimize of the distributed smart power grids in China. The accordingly modeling and simulation architectures and the open management and control platform will be designed, verified and performed. The task is to optimize the deployment of electric power in the advanced distributed power grids. With the consuming and saving behavior of the electric power be tracked, modeled, simulated and predicted; and the relationships between the distributed energy generation equipments, transmission controllers and stations are achieved, coordinated and optimized. 9. CONCLUSIONS Cyber-Physical Systems (CPS) is an emerging technology with excellent performances, which can be applied into many different areas, thus constructed a large-scale heterogeneous environment of Distributed Cyber-Physical Systems (DCPS). This paper aims to manage and control DCPS resources efficiently and smartly. An Internet of Things (IoT) and Cloud computing based DCPS management and deployment structure is designed and proposed to improve the performances in computing and communication. And a multi-model based hierarchical modeling and simulation architecture for an open and reusable DCPS management and control platform over the IoT and Cloud-based self-managed environment is then presented, with an experimental technology framework towards a cooperative and robust DCPS being constructed and described. Future researches will focus on the simulation, realization and optimization of this modeling and simulation architecture in the application area of Smart Grids under the future internet.

ACKNOWLEDGMENTS This work is supported by the National Natural Science Foundation of China (Grant No. 71071116), the Shanghai Key Program for Basic Research of China (Grant No. 10JC1415300), and the National High-Tech. R&D Program of China.

REFERENCES Bujorianu, M.C., and Bujorianu, M.L., 2009. A Unifying Specification Logic for Cyber-Physical Systems. Proceedings of the 17th Mediterranean Conference on Control and Automation, pp 11661171. June 24-26, Tessaloniki, Greece. Calheiros, R.N., Ranjan, R. and others., 2009. CloudSim: A Novel Framework for Modeling and Simulation of Cloud Computing Infrastructures and Services, Echnical Report, GRIDS-TR-2009-1, Grid Computing and Distributed Systems Laboratory, 9 pages. March 13, The University of Melbourne, Australia.

135

Distributed Computing, pp363-369. May 5-7, Orlando. Li, B., Chai, X. and Hou, B., 2009. Networked Modeling & Simulation Platform Based on Concept of Cloud Computing—Cloud Simulation Platform. Journal of System Simulation, 21 (17): 5292-5299. Liberatore, V., 2007. Networked Cyber-Physical Systems: An Introduction. Proceedings of 14th Annual IEEE International Conference and Workshops on the Engineering of Computer-Based Systems, pp 1-2. January 13, Tucson. Lyster, R., 2010. Smart Grids: Opportunities for Climate Change Mitigation and Adaptation. Monash University Law Review, 36 (1):137-191 Ma, W.F., 2010. CPS: a sense of control from sensor networks to the network. China Computer Newspaper, 7: 25-27. (in Chinese version). Rajhans, A., Cheng, S.W. and others., 2009. An Architectural Approach to the Design and Analysis of Cyber-Physical Systems. Rammig, F.J., 2008. Cyber Biosphere for Future Embedded Systems. Proceedings of the 6th International Workshop on Software Technologies for Embedded and Ubiquitous Systems, Vol 5287, pp.245-255. Oct 1-3, Anacarpi. Italy. Tan, Y., Vuran, C. M., and others., 2009. SpatioTemporal Event Model for Cyber-Physical Systems. Proceedings of 29th IEEE International Conference on Distributed Computing Systems Workshops, pp 44-50. June 22-26, Montreal, Quebec, Canada. Thacker, R.A., Myers, C.J. and Zheng, H., 2010. Automatic abstraction for verification of cyberphysical systems. Proceedings of the 1st ACM/IEEE International Conference on CyberPhysical Systems ACM, pp 12-21. April 13-15, New York, USA. Thiagarajan, A., Ravindranath, L., LaCurts, K. and others., 2009. VTrack: Accurate, Energy-Aware Road Traffic Delay Estimation Using Mobile Phones. Proceedings of 14th ACM SenSys, pp 8598. November 30, Berkeley, CA. Ulieru, M., 2007. e-networks in an increasingly volatile world: Design for resilience of networked critical infrastructures, Proceedings of The Inaugural IEEE International Conference on Digital Ecosystems and Technologies - IEEE-DEST, pp 540-545. Feb 21-23, Cairns, Australia. Vouk, M.A., 2008. Cloud computing — Issues, research and implementations. Journal of Computing and Information Technology, Cloud Computing – Issues, Research and Implementations. – CIT, 16 (4):235–246. Xia, F., Ma, L.H. and others., 2008. Network QoS Management in Cyber-Physical Systems. Proceedings of International Conference on Embedded Software and Systems, pp 302-307. July 29-31, Chengdu.

Xiao, X., Yu, H.Q. and Fan, G.S., 2010. A Petri NetBased Approach to Aspect-Oriented Use Case Modeling. East China University of Science and Technology: Natural Science, 35 (2): 248-254. (in Chinese version) Zhang, L., Leung, Henry., and others. 2008. Information Fusion Based Smart Home Control System and Its Application, IEEE Transactions on Consumer Electronics 54:3, 1157-1165. Zhang, Y., Ilic, M.D., Tonguz, O.F., 2007. Application of support vector machine classification to enhanced protection relay logic in electric power grids. Proceedings of Large Engineering Systems Conference on Power Engineering, pp.30-37. October 10-12, Montreal. Canada. Zhang, H.G., Yan, F. and others., 2010. Research on theory and key technology of trusted computing platform security testing and evaluation. SCIENCE CHINA Information Sciences, 53 (3): 434-453. Zhao, W., 2010. WSN information in the physical system topology control and MAC Protocol. Thesis (PhD). Central South University. (in Chinese version) Zweigle, O., Andrea, R., Haussermann, K., 2009. RoboEarth: connecting robots worldwide. Proceedings of the 2nd International Conference on Interaction Sciences: Information Technology, Culture and Human, pp184-191, Nov.24-26, Seoul.

AUTHORS BIOGRAPHY WANG Zhong-Jie is a Professor and Ph.D. supervisor in the Department of Control Science and Engineering, College of Electronics and Information Engineering, Tongji University, Shanghai, China. Her research interests focus on the modeling, simulation and scheduling of complex environments and hybrid systems, including Dynamic Programming and Optimal Control; Cyber-Physical Systems; Cloud Manufacture; Smart Grids and Social Networks. XIE Lu-Lu is a Ph. D. candidate in the Department of Control Science and Engineering, Tongji University, Shanghai, China. Her research interests are in the field of intelligent modeling, control and optimization techniques of Distributed Cyber-Physical Systems.

136

TRANSPORT NETWORK OPTIMIZIATION: SELF-ORGANIZATION BY GENETIC PROGRAMMING J. Göbel(a), A.E. Krzesinski(b), B. Page(a) (a)

(b)

Department of Informatics, University of Hamburg, 22527 Hamburg, Germany Department of Mathematical Sciences, University of Stellenbosch, 7600 Stellenbosch, South Africa (a)

{goebel,page}@informatik.uni-hamburg.de, (b)[email protected]

improvements to the network topology itself (e.g. increasing nodes’ capacities, establishing additional links), the only degree of freedom for achieving such targets is the logic the nodes use to determine which entity (e.g. vehicle, item, IP packet) to process next. This logic for entity prioritization can be set up by a central authority, which is typically provided with global knowledge of the network state. For example, based on global (estimated) traffic density data, signals can be coordinated such that platoons of vehicles are able to traverse the network without stopping (“green waves”). However, such attempts to centrally optimize such networks typically imply exponential computational complexity (Holland 1995), yielding bad scaling behaviour. Further assuming the requirement to adaptively adjust to dynamic changes in traffic patterns, such approaches depend on the availability of the central server and the communication to this authority. This motivates applying decentralized optimization: Without being dependent on a central authority, each node (router, workstation, traffic light) independently decides which entity is processed next. This decision is based on information that is locally available (e.g. queue lengths, local flow estimations) only, enabling nodes to act autonomously if assuming means of obtaining theses data (e.g. induction loops and cameras and image processing capabilities at an urban intersection). Literature – e.g. in Bazzan 2005, Cools 2007, Gershenson 2005, Helbing 2008, Lämmer 2007 attempting decentralized urban traffic optimization – already provides local node control logic performing almost as well as or better than centrally controlled systems for some special network topologies, e.g. for Manhattan grids with one-way traffic (i.e. north to south and west to east traffic only, Gershenson 2005) or intersection inflows from different directions assumed to be mutually exclusive (Lämmer 2007). The remainder of this paper is organized as follows: Section 2 provides further details about decentralized transport network optimization. Section 3 proposes genetic programming (GP) as potential solution, evaluated in Section 4 by experiments in a simulation environment. The paper concludes with a summary and outlook about further work in Section 5.

ABSTRACT This goal of the paper is transport network optimization. Transport networks are defined as network topologies where entities are forwarded from node to node constrained by capacity restrictions both on nodes and links. Examples include urban traffic (vehicles/signalized intersections) and IP networks (packets/routers). Optimization of such networks particularly has to provide the logic the nodes use to determine which entity to process next. Such logic can be imposed by a central authority based on global knowledge of the network state. In contrast, a selforganizing network solely relies on local decision rules to prioritize entities. At the cost of a potential loss in performance, such a decentralized network control is scalable and robust. This paper proposes genetic programming to evolve local node rules. Results indicate that the performance is similar to centrally (near-optimally) controlled systems. Keywords: transport network optimization, genetic programming, discrete event simulation, simulation framework, Java 1. INTRODUCTION The target of the work is the optimization of a general class of networks, namely transport networks. These are defined as graph topologies formed by nodes and links; continuously, entities “appear” at originating nodes (all nodes or only a subset may qualify as originating nodes). Such entities are queued for being “processed” by their origin nodes. After processing, the entities traverse links, thus queuing for processing at the next nodes on their routes until eventually reaching their destination nodes. Examples of such networks include urban traffic (vehicles advancing from one intersection to the next), conveyor-based manufacturing systems (items processed successively by different workstations) and telecommunication networks (e.g. IP packets being forwarded from one router to the next). Optimizing such a transport network typically may involve minimizing waiting or travel times or maximizing throughput. Apart from discarding entities or adjusting their routes (which may or may not be feasible, depending on the network type) and long-term

137

2.

DECENTRALIZED TRANSPORT NETWORK OPTIMIZATION Transport network optimization can be conducted by a central authority to which all relevant information is made available; examples include Diakaki 2003, Pohlmann 2010 or commercial systems like SCOOT (see e.g. United Kingdom Department for Transport 1995). However, apart from the dependence on the communication of each node to this central authority, the run-time performance of such approaches scales badly with the network size, compare Section 1. Furthermore, as optimization is typically conducted in cycles of 15-60 minutes, reaction to patterns of traffic shifting or the failure of an adjacent node is delayed. This motivates applying decentralized optimization, analogously to the concept of selforganization in thermodynamic and other natural sciences: A self-organizing system autonomously acquires and maintains order despite external influence subject to perturbations, which is typically achieved by a set of microscopic (local) decision rules independently used by each component (Wolf 2005). However, the development of such rules if not known in advance is difficult; designing a self-organizing system can be interpreted as reverse-engineering such rules from the desired macroscopic behaviour of the system: This behaviour (e.g. efficient transportation) is an emergent property of such rules, for which research thus far does not offer an agreed-upon and universally applicable means of obtaining (Zambonelli 2004). Nonetheless, the development of such local rules facilitating decentralized optimization of a transport network can be approached from the input side, i.e. the information locally available at the nodes: From Bazzan 2005, Cools 2007, Gershenson 2005, Helbing 2008, Lämmer 2007 and other approaches to solve special cases of the network optimization problem, a set of criteria can be obtained, upon which the decision as to which entities to prioritize at a given instant is based. Using from now on urban traffic terminology for example, these criteria apply either to a specific lane (a queue for vehicles arriving at a node on a certain link, waiting for processing and departure on one or more other links), to an intersection as whole (e.g. maximum queue length, total estimated arrival rate), or even to the overall network (e.g. switching penalty, during which all lanes are “red” for safety reasons). See Table 1 for further examples for such criteria. Lämmer 2007 has also shown that any node logic can be expressed as function which he refers to as the priority index: Based on a subset of these criteria, one can determine the priority index of each lane; serve the vehicles from the lane with the highest priority, unless network stability would be violated if a lane did not receive any service for some maximum waiting period. Table 2 shows an extension of this mechanism using a set of lanes instead of single lanes, taking into account intersections serving more than one lane simultaneously (e.g. opposing traffic from north and south proceeding straight ahead).

&ůŽǁ;ůĂŶĞͿ YƵĞƵĞůĞŶŐƚŚ;&͕>Y>Ϳ >ŽŶŐĞƐƚǁĂŝƚŝŶŐƚŝŵĞ;&͕>tdͿ Ɛƚ͘ĂƌƌŝǀĂůƌĂƚĞŽǀĞƌĂůů;&͕>KͿ Ɛƚ͘ĂƌƌŝǀĂůƌĂƚĞĐƵƌƌĞŶƚƉĞƌŝŽĚ;&͕>WͿ ƵƌƌĞŶƚƉŚĂƐĞƐŝŶĐĞ;&͕>W^Ϳ EŽŐƌĞĞŶƐŝŶĐĞ;&͕>'^Ϳ hƚŝůŝnjĂƚŝŽŶŽĨŝŶĐŽŵŝŶŐůŝŶŬ;&͕>/hͿ DĂdžŝŵƵŵƵƚŝůŝnjĂƚŝŽŶŽĨĂůůŽƵƚŐŽŝŶŐůŝŶŬƐ;&͕>KhͿ Ɛƚ͘ĂƌƌŝǀĂůƐŶĞdžƚƉĞƌŝŽĚ;&͕>Ϳ Ɛƚ͘ĚƵƌĂƚŝŽŶƵŶƚŝůŶĞdžƚƉůĂƚŽŽŶĂƌƌŝǀĂů;&͕>WͿ Ɛƚ͘ĨĞĂƐŝďůĞĨůŽǁͬĂďƐŽůƵƚĞ;&͕>&Ϳ Ɛƚ͘ĨĞĂƐŝďůĞĨůŽǁͬƌĞůĂƚŝǀĞ;&͕>&ZͿ EŽĚĞ;ŝŶƚĞƌƐĞĐƚŝŽŶͿ DĂdž͘ƋƵĞƵĞůĞŶŐƚŚ;&͕EY>Ϳ >ŽŶŐĞƐƚǁĂŝƚŝŶŐƚŝŵĞ;&͕EtdͿ Ɛƚ͘ĂƌƌŝǀĂůƌĂƚĞͬŽǀĞƌĂůů;&͕EKͿ Ɛƚ͘ĂƌƌŝǀĂůƌĂƚĞͬĐƵƌƌĞŶƚƉĞƌŝŽĚ;&͕EWͿ ƵƌĂƚŝŽŶŽĨĐƵƌƌĞŶƚƉŚĂƐĞ;&͕EW>Ϳ /ĚůĞ;͕E/Ϳ ďůŽĐŬĞĚ;͕E>Ϳ 

'ůŽďĂů;ŶĞƚǁŽƌŬͿ ǀĞƌĂŐĞĂĐĐĞůĞƌĂƚŝŽŶ;&͕'Ϳ ^ǁŝƚĐŚŝŶŐƉĞŶĂůƚLJ;&͕'^WͿ WĂƐƐŝŶŐƉŽƐƐŝďůĞŝŶƋƵĞƵĞƐ;͕'WWͿ Table 1: Examples of criteria for entity (vehicle) prioritization in urban traffic; the suffix states the data types, either floating point numbers (F) or Boolean (B), and an abbreviation. Lane set selection can for example be conducted on greedy basis: The lane with the highest priority is set to green. Repeatedly, from all lanes not mutually exclusive to a lane already set to green, the lane with the highest priority is set to green until no further lane exists that is neither set to green nor mutually exclusive to a lane already set to green. 



ĞƚĞƌŵŝŶĞƉƌŝŽƌŝƚLJŝŶĚĞdžĨŽƌĞĂĐŚůĂŶĞ  hŶďůŽĐŬƐĞƚŽĨůĂŶĞƐǁŝƚŚŚŝŐŚĞƐƚĂĐĐƵŵƵůĂƚĞĚ  ƉƌŝŽƌŝƚLJ ZĞƉĞĂƚĞĚůLJƌĞͲĞǀĂůƵĂƚĞƉƌŝŽƌŝƚLJŝŶĚŝĐĞƐĂŶĚƐǁŝƚĐŚ  ƚŽĂĚŝĨĨĞƌĞŶƚƐĞƚŽĨůĂŶĞŽŶĐĞƚŚĞŝƌ ĂĐĐƵŵƵůĂƚĞĚƉƌŝŽƌŝƚLJĞdžĐĞĞĚƐƚŚĞƉƌŝŽƌŝƚLJŽĨ ƚŚĞĐƵƌƌĞŶƚĨůŽǁƐďLJĂƉŽƐŝƚŝǀĞȴ  ŶƐƵƌĞŵŝŶŝŵƵŵŐƌĞĞŶĂŶĚŵĂdžŝŵƵŵǁĂŝƚŝŶŐ ƉĞƌŝŽĚƐĂƌĞŶŽƚǀŝŽůĂƚĞĚ Table 2: Link choice based on priority indices, extended from Lämmer 2007 With the input (the above-mentioned criteria) and the result (a priority index function) being specified, the open problem is determining how to obtain the latter from the former. For special network topologies, this

138

blocks”) representing functions that can be flexibly combined, with the only restriction that arguments and result data types of adjacent node functions have to match. An example priority index function is shown in Figure 1, composed of functions like for instance “Multiply” (bottom right), representing a function accepting two floating point numbers (F) as arguments and returning the product as floating point number as result. The overall result of the GP program is the return type of the top-level function (“Node programming logic”), which itself requires arguments determined by functions on the second level and so forth. Note that the closure of genetic programs assumes the availability of functions not requiring any argument, so-called terminals, e.g. numerical or Boolean constants or values read from variables (notation V/… to distinguish from constant terminals), like the node’s maximum queue length (V/NQL), the longest waiting time of an entity at a lane (V/LWT), a lane’s queue length (V/LQL) or a lane’s estimated overall arrival rate (V/LAO), compare Table 1. A typical genetic programming cycle is summarized by Figure 2: Based on an initial population of programs generated at random, so-called genetic operators are applied to simulate biological evolution: The fitter (i.e. better network control performance) a program, the higher its probability of being selected for offspring composition by means of recombination; Figure 3 shows examples of recombination operators, particularly chromosome transfer X, aggregation Y, projection Z, and swapping [.

problem is already solved (compare literature quoted above), but not for the general case. Part of a more general solution might be setting local rules such that certain desirable patterns – like green waves – are facilitated, see Göbel 2009. However, for network configurations, in which no such desirable patterns are known or where patterns available do not suffice for fully specifying a priority index function, we propose a different approach based on genetic programming (GP) in Section 3. 3.

NODE LOGIC EVOLTION BASED ON GENETIC PROGRAMMING Our target is to obtain priority index functions for decentralized transport network control based on the input criteria from Table 1. In the absence of any restrictions of which of these criteria to use and how to algebraically and logically determine a priority index (PI) from their respective values, genetic programming, see e.g. Koza 1992 or Poli 2008, has been chosen because is a flexible and robust search technique not relying on any preconditions which would limit is applicability to special cases of transport networks: Mimicking nature, programs to determine priority indices are evolved by successively improving existing programs. Note that GP can be interpreted as a generalization of genetic algorithms (GA): While the search space of a typical GA is a static chromosome on which parameter values to optimize are stored, this structure the chromosome in GP itself is subject to evolution. ŝƚLJ ŝŽƌ dž Wƌ ŝŶĚĞ

Ğϭ > ŝŶ

ĞϮ >ŝŶ

Ğϯ >ŝŶ

Ğ< >ŝŶ

/ŶŝƚŝĂůŝƐĞ Ͳ ŐĞŶĞƌĂƚĞƉŽƉƵůĂƚŝŽŶ Ͳ ĞǀĂůƵĂƚĞƉŽƉƵůĂƚŝŽŶ ŝŽƌ Wƌ

ŝƚLJ

ĞŶ dŚ

ŶĚ Ž yх

LJĞƐ

Ğ ůƐ





ƚĞƌŵŝͲ ŶĂƚŝŽŶ͍

ůƵĞ sĂ

z ůƵĞ sĂ

y ůƵĞ sĂ

ŽŵƉŽƐĞŽĨĨƐƉƌŝŶŐƐ Ͳ ƐĞůĞĐƚŝŽŶ;ŵĂƚŝŶŐͿ Ͳ ƌĞĐŽŵďŝŶĂƚŝŽŶ

z

DƵƚĂƚŝŽŶ Ͳ ĐƵƌƌĞŶƚŐĞŶĞƌĂƚŝŽŶ Ͳ ŽĨĨƐƉƌŝŶŐƐ

ůƵĞ sĂ

y

ŶŽ

z ůƵĞ sĂ

hƉĚĂƚĞƉŽƉƵůĂƚŝŽŶ Ͳ ƐĞůĞĐƚŝŽŶ;ƐƵƌǀŝǀĂůͿ Ͳ ĞǀĂůƵĂƚĞƉŽƉƵůĂƚŝŽŶ

Figure 1: Example of a GP node priority index function

KƵƚƉƵƚ Ͳ ďĞƐƚƐŽůƵƚŝŽŶ;ƐͿ

To facilitate the evolution of genetic programs of arbitrary complexity, genetic programs typically have a tree-like structure consisting of nodes (“building

'ĞŶĞƌĂƚŝŽŶĐLJĐůĞ

Figure 2: GP evolution, based on Koza 1992

139

Figure 3: GP recombination operators

WƌŽŐƌĂŵ

WŽůLJŐŽŶϭ

͘Ő͘ĚƌĂǁ

WŽůLJŐŽŶϮ >ŝŶĞϭ >ŝŶĞϮ WŽůLJŐŽŶϯ

͘Ő͘ĐŽŵƉĂƌĞƚŽ

ǀĂůƵĂƚŝŽŶ

&ŝƚŶĞƐƐ

WƌŽŐƌĂŵ

/ŶƐƚĂŶĐŝĂƚĞ ŽŶĐĞƉĞƌŶŽĚĞ

ƌĞƉĞĂƚ ƌĞƉĞĂƚ ƌĞƉĞĂƚ ǁŚŝůĞ;ĐŽŶĚϭͿ

dƌĂĨĨŝĐ ^ŝŵƵůĂƚŝŽŶ

ƌĞƚƵƌŶW/с͙ ǁŚŝůĞ;ĐŽŶĚϮͿ ƌĞƚƵƌŶW/с͙ ŝĨ;ĐŽŶĚϯͿ ƌĞƚƵƌŶW/с͙

ǀĂůƵĂƚŝŽŶ

ĞůƐĞ ƌĞƚƵƌŶW/с͘͘͘

&ŝƚŶĞƐƐ сĐƵƌƌĞŶƚ;ƌĞͲͿĞŶƚƌLJƉŽŝŶƚ Figure 4: Alternative means of evaluating the fitness of a program 140

Both existing programs and new offsprings have a small chance of undergoing a random mutation, e.g. functions replaced by other functions with the same input and result parameter types, including terminals potentially being replaced by other terminals of the same type. For the resulting set of programs, a fixed number is selected to “survive” and form the next generation. The set of functions typically used to GPbased mathematical functions includes numerical (plus, minus, multiply, divide, power, root, exp, log, abs) and logical operators (and, or, xor, not, greater, smaller) as well as statements to define piecewise functions (if/else, switch). Inversion of control and state-dependent priority indices The standard GP approach of determining the fitness of a program is shown in Figure 4, upper half: The full program – e.g. serving the purpose of creating a Mona Lisa forgery by drawing lines and filled polygons on a canvas (courtesy of Meffert 2011) – is executed once, followed by evaluating the fitness by comparison to the “real” Mona Lisa. For two reasons, this paradigm is not appropriate for determining node priority index (PI) functions: First, the evaluation of a PI function is driven by a simulation experiment calling the PI function whenever desired.

Secondly, local PI functions as proposed by Gershenson 2005 or Lämmer 2007 are dependent on previous calls to the PI functions since they use partial conditionals (e.g. “if” without “else”, leaving PI calculation to the next line if the condition does not apply) and loops. Allowing such program components potentially yields programs evaluated only partially, compare Figure 4, bottom half: An equation determining the PI inside the “while”-function for instance may be evaluated repeatedly until the “while”condition is no longer fulfilled. Our implementation, based on JGAP (Java Genetic Algorithms Package, see Meffert 2011) reflects this control flow. Particularly, we propose a special function referred to as “Node Programming Logic” (see also Figure 1), which keeps a reference to the current (re-)entry point, i.e. the “line” to use at the moment to determine the priority of a lane, thus emulating program execution. Note that multiple nodes in a network require multiple program instances with a different (re-)entry point token each to reflect nodes potentially being in different states. The return type of “Node Programming Logic” is a PI floating point value, which allows for for hierarchically nesting partial conditionals and loops using multiple “Node Programming Logic” functions. After all lines have been used for determining a set of lanes to set to green, the program execution resumes at the first line. Wrapping up, the program from Figure 1 will assign the highest priority to the lane of the vehicle with longest waiting time, which yields a FIFO (First in, first out) service, see the “then” branch of the “if” clause, as long as congestion is moderate (less than 50 entities queued in total). Otherwise, lane priority is the product of queue length and arrival rate: As congestion increases, the function tends to prefer main roads and to serve multiple vehicles before switching to other links. Section 4 will investigate the performance of this GP approach of decentralized network optimization for three example networks (one intersection, two intersections, a small city area).





identical conditions (same speed limit of 50 km/h, single lanes). In S2, two intersections are located 100 meters apart. Symmetrical traffic flow is restricted to WN and ES, yielding mutual exclusiveness at both intersections (see Figure 5, assuming right-hand traffic). S3 is a network consisting of 11 intersections from southern Hanover/Germany subject to various flows from almost any entry to almost any exit, see Pohlmann 2010.

Figure 5: Scenario 2 (Two intersections)

4. EVALUATION We have built a discrete event simulation environment which is sufficiently parametrizable to represent different kinds of transport networks like urban traffic and IP packet routing; see Göbel 2009 for details about this environment. The mesoscopic logic of entity movement is derived from the traffic queuing model proposed by Nagel 2003. This simulation is based on DESMO-J, a framework for discrete event modelling and simulation in Java (see Page 2005 and the web page at http://www.desmoj.de), developed at the University of Hamburg. To evaluate the performance of the described GP approach of transport network optimization, we have investigated three network scenarios S1, S2, S3: • S1 consists of a single isolated intersection with incoming traffic from two directions with

Figure 6: Scenario 3 (Hanover) The optimization target is to minimize the vehicles’ average overall waiting times. For S1, a near-optimal strategy is alternatingly (“round robin”) serving each flow until the queue is empty: Switching earlier is not optimal as intersection capacity would be given away due to the switching penalty incurred in terms of a two second safety period in which all traffic lights have to be red; switching later most likely wastes capacity as no vehicle is served (unless the next arrival of a vehicle not yet queued is very close at hand). Since the space between the intersections is limited in S2, the optimization problem particularly involves synchronizing their traffic lights such that both intersections never waste capacity by being unable to server either of the flows. This undesired situation

141

states average waiting time and vehicle throughput during 5 hours (average of 10 runs). S1 has been evaluated with three different load levels (approx. 3200, 6400 and 9200 vehicles offered). The GP fitness function was the average waiting duration (the lower, the better), subject to a penalty proportional to the nodecount of the genetic program, thus implicitly bounding the complexity the programs evolved. Table 4 shows the “fittest” programs found in each of the runs S1/S2 and S3. Comparing the results of GP to the (near-)optimal solutions in the case of S1 and S2 or to the heuristic in S3 yields a GP performance similar or better (with the exception of the second S1 load level): Although not applying centralized control, e.g. explicit traffic light synchronization in S2, the GP solutions perform approximately equally well or in some cases even slightly better by exploiting the marginal remaining optimization potential, e.g. asymmetrical link clearance in S2 used to advance the traffic light switch at the relevant intersection which is advantageous in terms of overall waiting durations if the flow set to green is slower than its counterpart receiving green later.

occurs if the link towards the other intersection is fully congested (thus no further incoming vehicles from W at the western intersection and from E at the eastern intersection can be served) while at the same time no vehicles bound for N/S already served by the other intersection are waiting for service. A near-optimal centralized solution for S2 is “fill and clear”, exploiting the symmetry of the traffic flows offered: At both intersections the incoming flows from W at the western intersection and from E at the eastern intersection receive green until the link between the intersections is filled or until both queues incoming are empty. Synchronously, the vehicles bound for N/S now receive green at both intersections until the link between the intersections is cleared. Assuming enough “supply” of incoming vehicles, neither of the intersections ever wastes capacity apart from symmetry deficits caused by stochastic noise, e.g. one intersection clearing its vehicle queue on the link between the intersections faster than the other. For S3, a heuristic is used serving the longest queues (typically four protected flows on a four-way intersection, e.g. WS, E, N and NW in right-hand traffic) until the queues for different combinations of flows are least 25 vehicles longer, yet at least for 5 seconds.

5. SUMMARY AND OUTLOOK This paper has presents a GP-based approach to decentralized, transport network optimization, providing local rules in terms of priority index functions. To the standard paradigm of GP evolution (Koza 1992), adjustments were necessary to cover inversion of control in fitness evaluation (simulation calling the program to be evaluated, not vice versa) and statedependently only partially executing the program to be evaluated; these adjustment were implemented extending JGAP (Java Genetic Algorithms Package, see Meffert 2011). Experiment results indicate that the performance is similar to centrally (near-optimally) controlled systems while at the same node control is scalable and not dependent on a central authority. “Performance” of course is not restricted to minimizing waiting times as conducted in the experiments in Section 4; the GP-based transport network is sufficiently flexible to use any fitness function, e.g. a weighted combination of waiting times and fuel consumption/emission production. Further work will address the run-time performance of the GP evolution of node PI functions: As the fitness evaluation of a single program is relatively expensive due to the discrete event simulation runs to be executed (approx. 1 day for scenarios S1/S2, approx. 4 days for scenario S3 on a single machine), recognizing and not evaluating inferior programs may provide large improvements in run-time performance. Examples of such inferior programs are all programs containing branches that are never executed (e.g. all lines after a ǁŚŝůĞ;ƚƌƵĞͿ΂͙΃ statement) or programs not containing a single lane-specific criterion (compare Table 1) as they yield the same priority for all lanes at an intersection. The convergence of the GP can also be improved by providing “higher level” criteria, e.g. including the

Avg. ThroughWait put Round robin 5.7 3167 S1 (low load) GP 5.4 3161 Round robin 17.4 6243 S1 (high load) GP 19.0 6085 Round robin 141.6 6376 S1 (overload) GP 135.5 6459 Fill and clear 13.6 5170 S2 GP 13.0 5201 Longest queue 77.3 18663 S3 GP 52.9 23377 Table 3: Experiment performance results Network

Network

S1/S2

S3

Control

Program ƌĞƉĞĂƚ  ŝĨ;sͬE/Ϳ   ƌĞƚƵƌŶW/с;sͬ>/hͿ  ĞůƐĞ   ƌĞƚƵƌŶW/сĞdžƉ;sͬ>&Ϳ ƌĞƉĞĂƚ  ŝĨ;sͬE/Ϳ   ƌĞƚƵƌŶW/сsͬ>K  ĞůƐĞ   ƌĞƚƵƌŶW/сĞdžƉ;sͬ>W&ͿΎsͬ>td  ǁŚŝůĞ;EͬEY>хϲϴ͘ϰϰϰϱͿ   ƌĞƚƵƌŶW/сsͬ'ΎĞdžƉ;ĞdžƉ;sͬ>Y>ͿͿ

Table 4: GP evaluation results Table 3 compares these means of intersection control to the best GP solution found in 100 generations of size 100 for S1/S2 (combined) and for S3; the table

142

Koza, J. R., 1992. On the programming of computers by means of natural selection. Cambridge (Massachusetts, USA): MIT Press. Lämmer, S., 2007. Reglerentwurf zur dezentralen Online-Steuerung von Lichtsignalanlagen in Straßennetzwerken. PhD Thesis, Technical University of Dresden (Germany). Meffert, K. et al., 2011: JGAP – Java Genetic Algorithms and Genetic Programming Package. URL: http://jgap.sf.net Nagel, K., 2003. Traffic networks. In S. Bornholdt, H. G. Schuster (eds): Handbook on networks. New York (NY, USA): Wiley. Page, B. and Kreutzer, W., 2005. The Java Simulation Handbook – Simulating Discrete Event Systems with UML and Java. Aachen (Germany): Shaker. Pohlmann, T., 2010. New Approaches for Online Control of Urban Traffic Signal Systems. PhD Thesis, Technical University of Braunschweig (Germany). Poli, R., Langdon, W. B. and McPhee, N.F., 2008. A Field Guide to Genetic Programming, Raleigh (North Carolina, USA): Lulu.com United Kingdom Department for Transport, 1995. “SCOOT” Urban Traffic Control System. United Kingdom Department for Transport Traffic Advisory Leaflet 04/1995. Wolf, T. de and Holvoet, T., 2005. Emergence Versus Self-Organisation. In S. A. Brueckner, et al. (eds): Engineering Self-Organising Systems, pp. 1-15. Berlin (Germany): Springer. Zambonelli, F., Gleizes, M.-P., Mamei, M., and Tolksdorf, R., 2004. Spray Computers: Frontiers of Self-Organization. Proceedings of the First IEEE International Conference on Autonomic Computing (ICAC'04), pp. 268-269, May, Miami (Florida, USA).

optimal priority for an isolated intersection subject to uniform flows (no stochastic noise) as determined by Lämmer 2007, thus relieving the GP evolution from producing such terms. Another means of facilitating GP convergence is removing the need for co-evolution by allowing a subtree referenced more than once: If the results determined by a certain sub-tree, the current GP approach would be required to create this repeating pattern more than once (Figure 7, left). Allowing multiple references (Figure 7, right) has to ensure infinite recursion is avoided, yet provides smaller programs without need to multiple branches undergoing the same evolution.

Figure 7: Multiple references to a sub-tree REFERENCES Bazzan, A., Oliveira, D. de, and Lesser, V., 2005. Using Cooperative Mediation to Coordinate Traffic Lights: A Case Study. Proceedings of Fourth International Joint Conference on Autonomous Agents and Multiagent Systems, pp. 463-469, July 25-29, Utrecht (The Netherlands). Cools, S.-B., Gershenson, C., and D'Hooghe, B., 2007. Self-organizing traffic lights: A realistic simulation. In M. Prokopenko (ed): SelfOrganization: Applied Multi-Agent Systems, pp. 41-49. London (UK): Springer. Diakaki, C., Dinopoulou, V, Aboudolas, K., Papageorgiou, M., Ben-Shabat, E., Seider, E., and Leibov, A., 2003. Extensions and new applications of the traffic signal control strategy TUC. Transportation Research Board, 1856:202-211. Gershenson, C., 2005. Self-Organizing Traffic Lights. Complex Systems 16(1):29-53. Göbel, J., 2009. On Self-Organizing Transport Networks – an Outline. Proceedings of the 6th Vienna International Conference on Mathematical Modelling (MATHMOD) 2009, p. 82. Feb 11-13, Vienna (Austria). Helbing, D., and Lämmer, S., 2008. Self-Control of Traffic Lights and Vehicle Flows in Urban Road Networks. Journal of Statistical Mechanics: Theory and Experiment 4(P04019):1-33 Holland, J. H., 1995. Hidden Order – How Adaption builds complexity. New York (New York, USA): Basic Books.

AUTHORS BIOGRAPHY J. Göbel holds a diploma in Information Systems from the University of Hamburg, Germany. He is scientific assistant and PhD candidate at the Center of Architecture and Design of IT-Systems at the University of Hamburg; his research interests focus on discrete event simulation and network optimization. A. E. Krzesinski obtained the MSc from the University of Cape Town and the PhD from Cambridge University, England. He is a Professor of Computer Science at the University of Stellenbosch, South Africa. His research interests centre on the performance evaluation of communication networks. B. Page holds degrees in Applied Computer Science from the Technical University of Berlin, Germany, and from Stanford University, USA. As professor for Applied Computer Science at the University of Hamburg he researches and teaches in the field of Modelling and Simulation as well as in Environmental Informatics.

143

3D Physics Based Modeling and Simulation of Intrinsic Stress in SiGe for Nano PMOSFETs Dr. A. El Boukili Al Akhawayn University, Ifrane 53000, Morocco Email: [email protected]

Abstract

the intrinsic stress is a user defined input. And, it is not calculated internally by the simulator. In this paper, we are attempting to extend the 2D models found in the literature to 3D in the context of process simulation. We are following the 2D model of Van de Walle (Van de Walle and Martin 1986). Most of nano semiconductor device manufacturers as Intel, IBM and TSMC are intentionally using this intrinsic stress to produce uniaxial extrinsic stress in the Silicon channel. And, it is now admitted that the channel stress enhances carrier mobilities for both nano PMOS and NMOS transistors by up to 30% (Krivokapic 2003). The need of the hour is the development of accurate physics based models and the use of TCAD simulation tools to understand the physics of intrinsic and extrinsic stress and how to attain the desired stress in the channel. This paper is organized as follows. Section 2 outlines the different sources of intrinsic stress in Si1−x Gex pockets generated during deposition. Section 3 describes the proposed 3D model to calculate accurately the intrinsic stress due to lattice mismatch between Si1−x Gex and Silicon. After deposition process, intrinsic stress produces an extrinsic stress distribution in the whole device. This section, also outlines the 3D elastic model for the extrinsic stress and how it is related to the intrinsic stress. Section 4 presents 3D simulation results and analysis of extrinsic stress distribution in 45nm Intel strained PMOSFETs (Ghani 2003) using the proposed intrinsic model. This section will also present 3D numerical results showing the effects of Germanium (Ge) mole fractions and crystal orientations on the intrinsic stress. For qualitative and quantitative validations, the channel extrinsic stress profiles will be calculated using the proposed 3D intrinsic stress model and will

We are proposing a new analytical model, in three dimensions, to calculate intrinsic stress that builds during deposition of Silicon Germanium pockets in source and drain of strained nano PMOSFETs. This model has the advantage of accurately incorporating the effects of the Germanium mole fraction and the crystal orientation. This intrinsic stress is used to calculate the extrinsic stress distribution in the channel after deposition. Simulation results of channel stress based on this model will be presented and discussed for Intel technology based nano PMOS transistors. Keywords: 3D Modeling, Intrinsic Stress, Silicon Germanium, Nano PMOSFETs

1

Introduction

The originality of this paper is the development of new analytical model, in three dimensions (3D), to calculate accurately the intrinsic stress in Silicon Germanium (Si1−x Gex ) due to lattice mismatch between Si1−x Gex and Silicon where x represents the Germanium mole fraction. This intrinsic stress is generated during deposition of Si1−x Gex pockets in source and drain of Intel nano PMOSFETs (Ghani 2003). In the literature, there are only few papers and only in two dimensions (2D) about the modeling of intrinsic stress in SiGe (Rieger and Vogl 1993; Van de Walle and Martin 1986; Fischetti and Laux 1996; Brash, Dewey, Doczy, and Doyle 2004). These papers were developed in the context of device simulations for mobility modeling under the effects of stress. On the other hand, in most advanced commercial or noncommercial process simulators as FLOOPS, Sentaurus, or Athena, 144

1

3

PROPOSED 3D ANALYTICAL MODEL

be compared with channel stress profiles found in the literature. At this point, we could not find any experimental values of intrinsic stress in 3D. Therefore, we could not provide any comparisons with experiments.

2

substrate according to their thermal expansion coefficients. This creates an intrinsic strain and stress in the film and also in the substrate. The thermal expansion coefficient is defined as the rate of change of strain with temperature.

2.3

2

Sources of intrinsic stress in SiGe

The deposition process plays a key role in determining the intrinsic stress in Si1−x Gex films. At first, we should note that the deposition takes place at elevated temperatures. When the temperature is decreased, the volumes of the grains of Si1−x Gex film shrink and the stresses in the material increase. The stress gradient and the average stress in the Si1−x Gex film depend mainly on the Silicon-Germanium ratio, the substrate temperature and orientation, and the deposition technique which is usually LPCVD (low pressure chemical vapor deposition) or PECVD (plasma enhanced chemical vapor deposition). It was observed that the average stress becomes more compressive, if the Ge concentration decreases (Hollauer 2007). Thus, it is expected that a film with higher Ge concentration has a higher degree of crystallinity and larger grains, which leads to higher film density and to higher intrinsic stress. The intrinsic stress observed in thin films has generally the following main sources.

2.1

Intrinsic stress due to lattice mismatch

During deposition, thin films are either stretched or compressed to fit the substrate on which they are deposited. After deposition, the film wants to be smaller if it was stretched earlier, thus creating tensile intrinsic stress. And similarly, it creates a compressive intrinsic stress if it was compressed during deposition. In this paper, we are focusing on developing an analytical model in 3D for this type of intrinsic stress.

2.2

Intrinsic stress due to thermal mismatch

Thermal mismatch stress occurs when two materials with different coefficients of thermal expansion are heated and expand/contract at different rates. During thermal processing, thin film materials like Si1−x Gex , Polysilicon, Silicon Dioxide, or Silicon Nitride expand and contract at different rates compared to the Silicon

Intrinsic stress due to dopant

Boron doping in p-channel source/drain regions introduces a local tensile strain in the substrate due to its size mismatch with Silicon. Boron (B) atom is smaller in size than Silicon atom and when it occupies a substitutional lattice site, a local lattice contraction occurs because the bond length for Si-B is shorter than for Si-Si (Randell 2005, Horn 1955). We will deal with the 3D modeling of intrinsic stress due thermal mismatch and doping in future work.

3

Proposed 3D analytical model

In this section, we are going to describe the proposed analytical model in 3D to calculate the three normal components (σ0xx , σ0yy , σ0zz ) of the intrinsic stress in Si1−x Gex due to lattice mismatch at the interfaces between Silicon and Silicon Germanium. We did follow the same strategy used in the 2D model of Van de Walle (Van de Walle and Martin 1986). We first calculate the strained lattice constants parallel and perpendicular to the interfaces in x, y and z directions. Then, from these lattice constants, we calculate the strain parallel and perpendicular to the interfaces in x, y and z directions . And, finally, we get the 3 normal stress components in 3D from the calculated strain using a modified Hookes’s law. The restriction of the proposed 3D model to 2D gives exactly the 2D model of Van de Walle. And, this is a great advantage for validations and even comparison issues. In 3D PMOSFET with SiGe source and drain as shown in the Figure 2, there are two interfaces between Si and SiGe: a vertical interface and a horizontal interface. In the Figure 2, the vertical interface is defined in the yz-plane and the horizontal interface is defined in the xz-plane. Let’s assume that Si1−x Gex pocket grown in source or drain area has thickness h v DSiGe and DSiGe at horizontal and vertical h v and DSi be the interface respectively. Let DSi thickness of the Silicon substrate at horizontal and vertical interface respectively. Strains will be generated due to lattice mismatch of 145

3

PROPOSED 3D ANALYTICAL MODEL

the lattice constants. Let ASi and ASiGe be the lattice constants of unstrained Silicon and Si1−x Gex . Let Ai,h,x , Ai,h,z and Ai,v,y be the strained lattice constants parallel to the horizontal interface in x and z directions and parallel to vertical interface in y direction. The index i represents the Silicon or Silicon Germanium materials. Let Ai⊥,h,x , Ai⊥,h,z , and Ai⊥,v,y , be the strained lattice constants perpendicular to the horizontal interfaces in x and z directions and perpendicular to the vertical interface in y direction. Please see Figure 1 to have an idea about the lattice constants parallel and perpendicular to a given interface between Silicon and Silicon Germanium. In 2D, Van de Walle assumed that SiGe Si SiGe In ASi ,h,x = A,h,x and A,v,y = A,v,y . Si 3D, we are also assuming that A,h,z = ASiGe ,h,z . For simplicity, let’s assume that A,x = Ai,h,x , A,z = Ai,h,z , and A,y = Ai,v,y . The 2D model of Van de Walle gave expressions to calculate the strained lattice constants parallel and perpendicular to the interfaces in x, y directions. For the interface in z direction, we use the same expression given by: h = (ASi GzSi DSi + A,z z h z h z h ). ASiGe GSiGe DSiGe )/(GSi DSi + GSiGe DSiGe

Ai⊥,h,z = Ai [1 − Diz (

A,z Ai

− 1)]

The shear modulus Gzi for Silicon and SiGe depend on the elastic constants of the material i and depend on the the orientation of interface in z direction. It is given by: i i + 2C12 )(1 − Gzi = 2(C11

3

Walle and Martin 1986 ): z Di,(001) z Di,(110) z Di,(111)

Ci

=

2 C12 i

=

i i i C11 +3C12 −2C44 i +C i +2C i C11 12 44

=

i i i C11 +3C12 −2C44 i +C i +C i C11 44 12

11

(1)

The ratio of strained lattice constants A,x , A,y , A,z and Ai⊥,h,x , Ai⊥,v,y , and Ai⊥,h,z to unstrained lattice constants Ai determines the intrinsic strain parallel and perpendicular to the interfaces in x, y, and z directions: (,h,x , ⊥,h,x , ,v,y , ⊥,v,y , ,h,z , ⊥,h,z ). At horizontal interfaces in x and z directions, and at vertical interface in y direction, we have: ASiGe

A

,x − 1), ,h,x = ( ASiGe

⊥,h,x = ( A⊥,h,x − 1) SiGe ASiGe

A

⊥,h,z = ( A⊥,h,z − 1) SiGe

A

⊥,v,y = ( A⊥,v,y − 1) SiGe

,z ,h,z = ( ASiGe − 1), ,y ,v,y = ( ASiGe − 1),

ASiGe

Let σ0xx,h,x, σ0zz,h,z , σ0xx,v,y and σ0yy,v,y be the intrinsic stress at the horizontal interface in x and z directions and at the vertical interface in y direction respectively. We use a modified Hookes’s law to get these intrinsic stress components from the intrinsic strains as follows: σ0xx,h,x = (C11 + C12 ),h,x + C12 (⊥,h,x + ,h,z ) σ0zz,h,z = E⊥,h,z σ0yy,v,y = (C11 + C12 ),v,y + C12 (⊥,v,y + ,h,z )

Diz 2 )

In this paper, the constant Diz depend on i i i the elastic constants C11 , C12 , C44 of each material i. And, they also depend on the interfaces’s orientations that are (001), (110), or (111). In this work, the elastic constants C11 , C12 , C44 for Si1−x Gex depend on the Germanium mole fraction x and on the elastic constants C11 , C12 , C44 of Silicon and Germanium that we get from Van Der Walles Table I. We are going to use a nonlinear extrapolation method of ( Rieger and Vogl 1993) to calculate C11 , C12 , and C44 for Si1−x Gex . We calculate the constant Diz that depends on the orientation of the interface in z direction by following the 2D model of (Fischetti and Laux 1996; Van de 146

σ0xx,v,y = E⊥,v,y The elastic constants C11 , C12 and the Young’s modulus E are those of Si1−x Gex . Finally, the normal intrinsic stress components σ0xx , σ0yy , and σ0zz in Si1−x Gex in the proposed 3D model are calculated as follows: σ0xx = σ0xx,h,x + σ0xx,v,y σ0yy = σ0yy,v,y

(2)

σ0zz = σ0zz,h,z We should note that the proposed 3D model given by the equations (2) reduces to 2D model of Van de Walle if we take σ0zz,h,z = 0 and ,h,z = 0. The 3 shear intrinsic stress

REFERENCES

4

components σ0xy , σ0yz , and σ0zx are taken to be zero. Then, the 6 components of the intrinsic stress tensor σ0 in Si1−x Gex are given by: σ0 = (σ0xx , σ0yy , σ0zz , 0, 0, 0). The intrinsic stress tensor σ0 is used as a source term to calculate, in the whole 3D nano MOSFET structure, the extrinsic stress tensor σ = (σ xx , σ yy , σ zz , σ xy , σ yz , σ zx ). We note that σ xx , σ yy , and σ zz represent the extrinsic stress along the channel, vertical to the channel, and across the channel. We assume that Silicon and Silicon Germanium are elastic materials. And, to calculate the stress tensor σ, we use the elastic stress model based on Newton’s second law of motion, and the following Hookes law relating stress to strain: σ = D( − 0 ) + σ0 . Here D is the tensor of elastic constants C11 , C12 , C44 , 0 = 0 is the intrinsic strain and σ0 is the intrinsic stress given by the equations (2). A detailed description of this elastic model is given in (El Boukili 2011). Table 1: Ge% 17 20 30 40 50

Ge Mole Fraction Effects on Stress σ0xx σ0yy σ0zz 10 9 −1.432e 3.269e 1.752e9 10 9 −1.674e 3.821e 2.047e9 10 9 −2.454e 5.607e 3.000e9 10 9 −3.197e 7.312e 3.914e9 10 9 −3.900e 9.321e 4.780e9

in Si1−x Gex . From Table 1, we observe that the intrinsic stress along channel becomes more compressive, if the Ge concentration decreases. This is in great agreement with what was reported in (Hollauer 2007). Table 2 show the effects of substrate orientations on the intrinsic stress. The results in Figures 3 and 4 show that the stress components σxx and σzz along the channel, and across the channel respectively are all significant. A similar stress distribution has been reported in (Victor 2004). The values of the calculated 3D extrinsic stress are also qualitatively and quantitatively in good agreement with those calculated in (Victor 2004). Figure 5 shows that the distribution of x stress component is compressive along channel as expected. Figure 6 shows that the distribution of the z stress component is really nonuniform in the channel. A similar result was reported in ( Victor 2004 ). These numerical results confirm that our implementation of intrinsic and extrinsic stress models in 3D provide valid and correct results. We also believe that these results are of great interest to the semiconductor community including industrials and academia.

Table 2: Substrate Orientation Effects on Stress Orientation σ0xx σ0yy σ0zz 9 10 (100) −8.859e 1.186e 6.361e9 10 9 (110) −1.432e 3.269e 1.752e9 10 9 (111) −1.556e 1.327e 7.116e8

4

3D Numerical Results and Analysis

The proposed 3D model for intrinsic stress given by the equations (2) is used to simulate numerically the 3D extrinsic stress in the channel of an Intel 45nm gate length PMOSFET shown in Figure 2. For the following numerical results we used (001) for the substrate orientation and 17% as the Germanium mole fraction. In the future, we will do more investigations using different gate lengths (32nm, 22nm and below), different substrate orientations and Germanium mole fractions. The Table 1 shows the effects of Germanium mole fraction on the intrinsic stress 147

Figure 1: Parallel and perpendicular strained lattice constants.

References [1] Brash, B., Dewey, G., Doczy, M., Doyle, B., 2004. Mobility enhancement in compressively strained SiGe surface channel PMOS transistors with HFO2/TIN gate stack. Elec-

REFERENCES

5

Figure 2: Materials and mesh of the simulated structure.

Figure 4: 3D distribution of z stress component across channel.

Figure 3: 3D distribution of x stress component along channel.

Figure 5: Cut along channel of x stress component.

trochemical society proceedings,12-30, 2004, Volume 07, San Antonio, California, USA.

[4] Ghani T. et al., 2003. A 90nm High Volume Manufacturing Logic Technology Featuring Novel 45nm Gate Length Strained Silicon CMOS Transistors. Proceedings of IEDM Technical Digest, 978-980, December 2003. Washington, DC, USA.

[2] EL Boukili, A., 2011, 3D Stress Simulations of Nano Transistors. To appear in Proceedings of the 16th European Conference on Mathematics for Industry. July 26-30, 2010. Wuppertal, Germany. [3] Fischetti, M., Laux, S., 1996. Band Structures, Deformation Potentials, and Carrier Mobility in Strained Si, Ge, and SiGe Alloys. J. Appl. Phys. Vol. 80: 2234-2240.148

[5] Hollauer, C., 2007. Modeling of thermal oxidation and stress effects. Thesis (PhD). Technical University of Wien. [6] Horn, F., 1955. Densitometric and Electrical Investigation of Boron in Silicon. Physi148

REFERENCES

Figure 6: Cut in z-direction of z stress component. cal Review Vol. 97: 1521-1525. [7] Krivokapic, Z. et al., 2003. Locally strained ultra-thin channel 25nm Narrow FDSOI Devices with Metal Gate and Mesa Isolation. Proceedings of IEDM, IEEE International, 445-448, 2003. Washington, DC, USA. [8] Randell, H., 2005. Applications Of Stress From Boron Doping And Other Challenges In Silicon Technology. Thesis (Master). University of Florida.

6

and the MSc degree in Numerical Analysis, Scientific Computing and Nonlinear Analysis in 1991 at Pierre et Marie Curie University in Paris-France. He received the BSc degree in Applied Mathematics and Computer Science at Picardie University in Amiens-France. In 1996 he had an industrial Post-Doctoral position at Thomson-LCR company in OrsayFrance where he worked as software engineer on Drift-Diffusion model to simulate heterojunction bipolar transistors for radar applications. In 1997, he had European Post-Doctoral position at University of Pavia-Italy where he worked as research engineer on software development for simulation and modeling of quantum effects in heterojunction bipolar transistors for mobile phones and high frequency applications. In 2000, he was Assistant Professor and Research Engineer at the University of OttawaCanada. Through 2001-2002 he was working at Silvaco Software Inc. in Santa Clara, CaliforniaUSA as Senior Software Developer on mathematical modeling and simulations of vertical cavity surface emitting lasers. Between 20022008, he was working at Crosslight Software Inc. in Vancouver-Canada as Senior Software Developer on 3D Process simulation and Modeling. Since Fall 2008, he is working as Assistant Professor of Applied Mathematics at Al Akhawayn University in Ifrane-Morocco. His main research interests are in industrial TCAD software development for simulations and modeling of opto-electronic devices and processes. http://www.aui.ma/perosnal/ A.Elboukili.

[9] Rieger, M., Vogl P., 1993. Electronic-band parameters in strained Si(1-x)Ge(x) alloys on Si(1-y)Ge(y) substrate. Phy. Rev. B., Vol. 48 No 19: 14276-14287. [10] Rim, K., et al., 2000. Fabrication and Analysis of Deep Submicron Strained-Si NMOSFETs. IEEE Transactions on Electron Devices, Vol. 47, No.7: 1406-1415. [11] Takagi, S. et al., 2003. Channel Structure Design, Fabrication and Carrier Transport Properties of Strained-Si/SiGe-On-Insulator (Strained-SOI) MOSFETs. IEDM Technical Digest,57-60, 10 December , Washington, DC, USA. [12] Van de Walle, C., Martin, R., 1986. Lattice constants of unstrained bulk Si(1-x)Ge(x). Phy. Rev. B., Vol. 34: 5621-5630. AUTHOR’s BIOGRAPHY Abderrazzak El Boukili received both the PhD degree in Applied Mathematics in 1995, 149

Simulation model for the calcination process of cement.

(a)

Idalia Flores(a)Guillermo Perea(b) (a) Facultad de Ingeniería, UNAM (b) Facultad de Ingeniería, UNAM

[email protected], (b)[email protected]

Abstract Simulation is an important tool when a phenomenon or input-output relationships of a system makes its operation or testing impossible, expensive, dangerous or impractical. This paper develops a simulation model for the burning process of Portland cement. The methodology used is the one used in simulation, which establishes the definition of the problem, analysis of the variables to be modeled, executes a basic model, a detailed model development, validation, reports and conclusions.

Figure 1 Cement components 2. Manufacturing process. 2.1 Obtaining raw materials. The cement manufacturing process begins with the extraction of raw materials that are found in deposits, usually in open quarries. The quarries are operated by controlled blasting in the case of hard materials such as limestone and slates, while excavators are used to dig out the soft materials (clays). Once the material is extracted and classified, it is then crushed to a particle size suitable for the mill product and is transported by conveyer belt or truck to the factory for storage in the prehomogenization pile.

Keywords: Simulation, calcinations, clinker, Cruz Azul cement, Arena, Simio. 1. Introduction

Cement is one of the main inputs in the construction industry in Mexico; domestic production was 42 million tons in 2010. The calcinations of cement unit is a system consisting of a Preheater, Kiln and Cooler (PHE), which raises the temperature of the limestone powder to 1.450° C, causing physicochemical changes and the formation of silicates in a granular mixture called clinker. Simulating the system PHE will allow us to analyze the formation of clinker through a mass-energy balance.

2.2 Homogenization and grinding of raw. In the prehomogenization pile, the crushed material is stored in top layers to be selected later in a controlled manner. The blending bed can prepare the proper dosage of components by reducing variability. Subsequently, these materials are ground in ball or vertical mills to make them smaller and thus make it easier to fire

To optimize this process we require a model that allows us to manipulate the different variables of the system. The aim of this paper is to build a simulation model of the calcination process in cement production, assessing the behavior of the input, distribution of the process, and output variables. Figure 1 shows some components of the cement.

150

them in the kiln. In the vertical mill, the material is crushed by the pressure of its roller on a turntable. From there, the raw material (powder or rawmix) is stored in a silo to increase the uniformity of the mixture.

the balls collide, crushing the clinker and additives to a fine homogeneous rawmix: cement. 2.5 Distribution. Finally, the cement is stored in silos, separated according to its various classes before being bagged or loaded onto a truck for transport by road or rail.

2.3 Preheater, kiln and cooler (PHE). The kiln is powered by means of the cyclone preheater that heats the feedstock to facilitate firing. Ground material or rawmix is inserted through the top of the tower and drops through it. Meanwhile, the gases from the kiln, which are at a high temperature, rise against the current, thus the rawmix is preheated before entering the kiln. As the rawmix progresses in the kiln while it rotates, the temperature increases to reach 1.500 ° C. At this temperature complex chemical reactions occur that result in the clinker. To achieve the temperatures required for firing the raw materials and the production of clinker, the kiln has a main flame that burns at 2, 000 º C. In some cases there is also a secondary flame located in the combustion chamber in the preheater tower. Once the clinker leaves the kiln, a cooler is introduced in inject cold air to lower the temperature from 1.400 ° C to 100 º C. The hot air generated in this device is returned to the kiln to support combustion, thereby improving the energy efficiency of the process.

3. The simulation model for the cement. 3.1. Calcination process analysis. The reactions that occur in the calcination process are: • Evaporation of water from the mixture. • Elimination of combined water in the clay. • Dissociation of magnesium carbonate. • Dissociation of calcium carbonate. • Reaction in the kiln, mixing the lime and clay. The kiln (heat exchanger-cooler) is the equipment that determines the production, being the most important part of the process. The clinker is produced by heating the properly dosed rawmix at high temperatures in an oxidizing atmosphere generally. The reactions of clinker produced essentially four main elements: CaO, SiO 2 , Al 2 O 3 , Fe 2 O 3 to form silicates with hydraulic properties. In overall, the clinker formation process can be divided into four parts:

2.4 Grinding of the clinker. Once the clinker is obtained, it is mixed in a cement mill with gypsum and additives, in the right proportions. Inside, the materials are ground, mixed and homogenized. The mills can consist of (horizontal and vertical) rollers or balls. The later consists of a large rotating tube with steel balls inside. Thanks to the rotation of the mill, 151

Temperature (ºC)

Reactions

ferroaluminato

20-100

CaCO 3 .MgCO 3 .Al 2 O 3 .SiO 2 .Fe 2 O 3 .H 2 O » CaCO 3 .MgCO 3 .Al 2 O 3 .SiO 2 .Fe 2 O 3 + H 2 O (V)

1260-1450

Formation of tricalcium silicate (C3S) from C2S and free lime CaO + 2CaO.SiO 2 » 3CaO.SiO 2

Dehydration of the mixture(evaporation of free water) 100-400

400-900

Temperature (ºC)

Reactions

It expels water of crystallization (water removal combined with clay)

1450

Belita Formation of and Alita

1300-1240

Crystallization of aluminates and ferrites

Chemical water is released.

Table 4 Cooled

Reactions

CaCO 3 » CaO + CO 2 Decarbonation MgCO 3 » Mg + CO 2 Dissociation of magnesium carbonate CO 2 is expelled Debajo de CaO + Al 2 O 3 » CaO.Al 2 O 3 800 Formation of calcium aluminate CaO + Fe 2 O 3 » CaO.Fe 2 O 3 Formation of ferrous oxide 800-900 CaO + SiO 2 » CaO.SiO 2 Formation of calcium silicate 900-950 5CaO + 3Al 2 O 3 » 5CaO.3Al 2 O 3 Formation of calcium trialuminato 950-1200 2CaO + SiO 2 » 2CaO.SiO 2 Formation of dicalcium silicate (C2S) Table 2 Calcination Temperature (ºC)

Reactions

1200-1300

3CaO + Al 2 O 3 » 3CaO.Al 2 O 3

3.2 System mass balance-preheater kiln-cooler (PHE). PHE Process is developed in the following steps: 3.2.1. Precalcination of the raw mixture. The preheater has a preheater, which heats the raw mixture of 60-70 ° C to 800-850 ° C, usually fueled by natural gas as fuel and use the waste gases from the kiln. 3.2.2. Formation of clinker. The formation of the clinker takes place in the rotary kiln, which is fed with the raw mixture from the preheater and, in turn, introduces hot air (secondary) cooler. 3.2.3. Cooling of clinker. The cooler consists of fans with variable flow through variable speed drives. 3.2.4 Cooling gases. The waste gases are cooled in a cooling tower, which is constituted by a system of nozzles and decanting to separate the oil carried by the gases. However, decanting is not enough, so an electrostatic filter is also used.

Formation of tricalcium aluminate (C3A) 1260

Table 3 Sintering and clinker

Al 2 O 3 .2SiO 2 .H 2 O » Al 2 O 3 + SiO 2 + 2H 2 O (v)

Table 1 Drying

Temperat ure (ºC) 500-900

3CaO + SiO 2 » 3CaO.SiO 2

4CaO + Al 2 O 3 + Fe 2 O 3 » 4CaO.Al 2 O 3 .Fe 2 O 3

3.2.5. Separation of dust from waste gases.

Formation of Tetracalcium

152

An electrostatic precipitator, consisting of plate-rapping systems and electric fields, is used to separate or precipitate dust from raw waste gases.

Due to the complexity and automation of most processes in the Cooperativa La Cruz Azul SCL, it is somewhat difficult to find areas of opportunity for improvement. At the present time, Cruz Azul has project engineering areas and an optimization department whose job is to constantly search for innovative technologies or technical information that would maximize existing resources, evaluate the replacement of equipment, performance and energy fuels and, if necessary, to supervise the construction of an entirely new factory. Usually these areas (or project optimization) work separately and there is a significant difference between the scope of each of them.

3.2.6. Separation of dust from the cooler. It uses an electrostatic filter that separates the particles of dust from clinker cooler air Mass balance: The aforementioned integrated process is summarized in the following block diagram:

However, in both the Project area and the area of optimization, the firing is the key process in cement manufacturing. From the roasting process is designed the size and capacity of the kiln, which in turn determines the capacity of the preheater building itself and, in consequence, the various skills of the teams that will take part in the design of a production line or a complete plant.

Figure 2 Mass balance

To calculate the mass flow rates, the incidence matrix for mass was developed, according to the mass flows that get in and out of the equipment:

The following questions arise: • Why not analyze the input-output flows under a controlled environment of massenergy and time? • Why not gradually change the way people work in the process engineering department, using a scientific methodology provided by simulation?

Table 5 Incidence matrix for mass

3.3 Simulation model for calcination unit No. 9 of Cruz Azul cement plant, Hidalgo, Mexico. Cooperativa La Cruz Azul S.C., a homegrown company from the Mexican state of Hidalgo, currently ranks third in national cement production after Cemex and Apasco.

The system to be analyzed consists of a combustion process as show below.

153

3.3.1 Collecting the information. Solving the system of mass balance equations with real data of the calcination unit No.9, we have:

Figure 4 Basic model with Arena

3.3.3. Complete Model. From the calculated mass balance, we developed a new model from the inputs and outputs, by considering mass and the stoichiometric analysis of the raw materials and the heat capacity of the fuel. APE software was used.

Table 6 Mass flow of calcination unit 9

3.3.4 Flowcharts According to the block diagram in which inputs and outputs represent the mass, we proceeded to develop flow charts for the calcination process.

3.3.2 Basic model. The basic model was developed in Arena, which is an initial flowchart where the flow of rawmix fed into the PreheaterKiln-Cooler system (PHE), and the chaotic movement is undergoes in the cyclone preheater (allocation probability) determines, in a linear fashion, both the consumption of fuel and electrical power.

Figure 5 Flowchart for the calcination process

3.3.5 Display We subsequently assigned the variables representing the masses (m 1 , m 2 , m 3 ...) to form the input-output system mass. The arrival of a continuous entity called rawmix is determined to simulate a power of 265 Ton / hr within the system.

Figure 3 Flow chart for the PHE

154

Figure 7 Simulation experiments

Figure 6 Simulations with SIMIO

As a result, we obtained the following data:

We observe in the Arena model that there is a chaotic movement of rawmix particles in the cyclone, so that we determine that they are probability fluxes (approximately 65% -35%). The temperature is monitored at the inlet to the Preheater, and at the inlet , center and outlet from the kiln, at the inlet and outlet of the clinker cooler. Fuel consumption is based on the flow of rawmix, which has already been decarbonated in order to achieve more efficient calcination. It simulates the consumption of coke (petroleum), whose consumption is approximately 7.500 kg / hr in the preheater (preheating) and 8.800 kg / hr in the main burner (kiln). The flow in the supply of rawmix is between 64.7 - 65 Ton / hr, so we used the linear function L (64.7-65) to simulate the behavior. 3.3.6 Experiments

In order to have more results, 10 experiments were performed, as shown below:

Table 7 Simulation experiments 3.3.7 Model Validation

In order to validate the model, we consider the nominal production of the calcination unit No.9 Cruz Azul, according to the information provided by the area of new projects, nominal output 155

is 157.65 tons / hr of clinker. Considering this fact, along with the 10 experiments and the linear nature of a controlled process, we can validate the statistical behavior of the production of clinker simulated by using the Student t test as shown in the following table:

of this operation on the final quality of the cement. It has been demonstrated to simulate a controlled continuous production process (24 hours a day 360 days a year) finally yields results that are very close to reality, regardless of the number of variables involved. The simulation that we developed was a process "in constant motion". There are several international companies that develop the engineering and construction for cement plants, using very complex mass and energy balances to determine the specific capacity of each piece of equipment to be installed, though, of course, the heart of the system is the installed kiln capacity and overall clinker production rate of the calcination unit. Simulation can give different scenarios for the future and allows the company to change or modify important parameters in the production of the cement. References Aguilar Barona Byrthzee Rubén. Tesis de Maestría: “La Simulación de Sistemas en la mejora de procesos de manufactura y servicios”. Facultad de Ingeniería UNAM. Año 2004. Barceló Jaime. “Simulación de Sistemas Discretos”. Primera Edición. Año 1996. Editorial Isdefe. Madrid España. García Dunna Eduardo, García Reyes Heriberto y Cárdenas Barrón Leopoldo. “Simulación y análisis de sistemas con ProModel”. Primera Edición. Editorial Prentice Hall, México 2006. Kelton David, Sadowski Randall y Sturrock David. “Simulación con Software Arena”. Cuarta edición. Editorial Mc Graw Hill. Año 2008. Rivett Patrick. “Construcción de modelos para análisis de decisiones”. Primera

Using the t-test, the mean and the standard deviation is calculated with a sample of the 10 experiments:

4. Conclusions

The linearity of an industrial process such as cement calcination represents very slight changes in control variables, because of the importance and criticality 156

(UPIICSA), currently studying a Masters in Systems Engineering with specialization in Operations Research at the National Autonomous University of Mexico (UNAM). He has worked in the cement industry in the planning and control of projects specifically in Cementos Cruz Azul. Within the concrete industry in the same group has developed databases and applied statistical tools that have increased the efficiency of the information in the technical area He is currently coordinator of new projects.

Edición. Año 1983. Editorial Limusa. México D.F. Deolalkar S P. “Handbook for Designing Cement Plants”. Añ0 2009. BS Publications. Ackoff Russell L., Sasieni Maurice W. “Fundamentos de Investigación de Operaciones” Sexta Edición. Año 1984. Editorial Limusa. México D.F. Web sites http://canacem.org.mx/canacem.htm http://www.cruzazul.com.mx http://www.holcim.com.mx http://www.ieca.es http://minerals.usgs.gov/minerals/pubs/mc s/2011/mcs2011.pdf http://www.investigacionoperaciones.com/Historia.htm http://www.vaticgroup.com/unlimitpages.a sp?id=81&pid=-1 http://www.flsmidth.com

AUTHORS BIOGRAPHY Dra. Idalia Flores de la Mota… Dr. Idalia Flores de la Mota is mathematics at the Faculty of Sciences of the UNAM and studied the Masters and Ph.D. in Operations Research at the Faculty of Engineering of the UNAM. She has published notes, chapters in books, booklets and articles disclosed in international journals. He has been referee of the journals, Simulation, the Journal of Accounting and Administration, Computing Reviews online and Revista Iberoamericana de Automática e Informática Industrial. It belongs to the Institute for Operations Research and the Management Science and is Director of the Center for Simulation McLeod in Mexico. Ing. Guillermo Perea Rivera. Industrial engineer, graduated from the National Polytechnic Institute

157

JOB SATISFACTION MODELLING IN AGENT-BASED SIMULATIONS Alexander Tarvid University of Latvia [email protected]

Granovetter (2005) mentions two reasons why social networks are so much used in the hiring process. Firstly, they help mitigate the problem of bilateral asymmetric information, when both prospective employers and employees do not know the other side's quality. In these settings, they search for more information about one another from personal sources they can trust. Secondly, the cost of searching for a new employee in existing social networks, which are maintained mainly for non-economic reasons, is far lower than using the formal channels. One could argue that existing employees may inflate the real qualifications of the friend they recommend, but this would contradict their long-term interest in the company. Therefore, using referral hiring is a theoretically clean way of reducing costs. At this point, we would like to stress that we do not touch upon normative theories of human resource management concerning whether appointing one’s friend to a position inside the company is a right way of doing management. Rather, we adhere to the literature on labour economics and observe what actually happens in real-world labour markets. Realizing the importance of social networking for labour markets, researchers started investigating the interplay between social networks and the economic situation of workers and firms. Several theoretical results (see, e.g., Bramoullé and Saint-Paul 2010, Calvó-Armengol and Jackson 2007, Krauth 2004) have been obtained for the steady-state of Markov processes describing employment and social network dynamics. To analyse dynamic non-equilibrium short- and medium-term effects, agent-based models were built. However, there still are restrictive assumptions under many of them. In some models, the unemployed were static—they were simply taking any vacancy the labour market proposed them. Abdou and Gilbert (2009) focus on the level of homophily driving the probability of both changing the social network and changing the employment status in a particular firm. They assume, however, that only social networking and homophily are the main determinants of labour status. Gemkow and Neugart (2011) use the experience-weighted attraction algorithm to guide agents in their network formation decisions. Nevertheless, by assuming that workers apply to all available vacancies, they do not

ABSTRACT Theoretical labour market models that incorporate social networks have largely focused on the steady-state of the system, ignoring their short- and medium-term dynamic effects. In many agent-based models of job search, the unemployed were either static, taking any vacancy proposed to them, or chose among vacancies based on either proposed wage or whether there were any of their friends employed in the firm. Thus, job satisfaction, an important multi-faceted concept in the labour market literature, has been overlooked. We propose a way to measure job satisfaction and illustrate how it can be incorporated in an agent-based model of the labour market. We use a simulation to study the dynamics of this model. Keywords: job satisfaction, agent based modelling, labour market, social network 1. INTRODUCTION Empirical studies show that social networks are important in the labour market. Bewley (1999) reports that 96 out of 161 (or 60 per cent) US businesses interviewed use personal contact networks to find job candidates, where in most cases, this meant employee referrals. Based on a survey of 6066 employers in Latvia, Hazans (2011) found that networking is the most popular recruitment method used by enterprises (depending on language used in enterprises, 30% to 50% of them hire by referral), but the intensity of systematic use of social networks decreases with firm size. Latvia is not an exception—indeed, Kuddo (2009) notes that in all Eastern European countries, a usual way of finding and hiring for vacancies is through informal channels (relatives, friends, acquaintances), especially in the small and medium enterprise sector. Employees also use their social networks in the process of job search. Montgomery (1991) cites several studies reporting that around 50 per cent of employees in the US found their jobs through friends and relatives. In Estonia, using Estonian Labour Force Survey data, we find that every year during 2001-2009, 30 per cent of respondents reported asking relatives and friends as their most important step taken to find a job (30% mentioned watching job ads, 15%—directly contacting employers, and 15% found it most helpful to seek through the state employment office).

158

that implements the labour market model and the analysis of its dynamics. The last section concludes.

model choice between them. The probability of being employed in their model depends only on whether an applicant has friends in the firm hosting the vacancy. Tassier and Menczer (2008) assume that agents learn about open vacancies after a formal search and based on information from their friends. Nevertheless, all vacancies are identical. Other models introduced heterogeneous vacancies. For instance, Tassier and Menczer (2001) present an evolutionary model where vacancies differ by the associated wage rate, and the person chooses the vacancy with the highest proposed wage. However, the social network plays only a role of informing its member on the vacancies available. In reality, individuals’ decisions on which vacancy to choose or whether to leave the current job depend on a combination of monetary and social rewards, rather than on each of these in isolation. In particular, it is quite well-known that job satisfaction (JS) is an important predictor of the decision to quit (Acker 2004, Manger and Eikeland 1990, Parry 2008). Carless and Arnup (2011) found that JS increases statistically significantly after a job change, which means that workers take into account expected job satisfaction when choosing among several job proposals. Kalleberg (1977, p. 126) defines JS as “an overall affective orientation on the part of individuals toward work roles which they are presently occupying” and views it as the result of an interplay between the values workers attach to job characteristics and the extent to which these values are satisfied. He proposes six dimensions of values: intrinsic (associated with the task itself), convenience, financial, relationships with coworkers (satisfaction of social needs), career, and resource adequacy. While he did not find that relationships with co-workers significantly affect JS, this does not mean that co-workers are irrelevant to it. Indeed, in his definition of resources he extensively mentions help, authority, information, supervision, and competency of co-workers, and these resources are found to significantly influence JS. Harris, Winskowski, and Engdahl (2007) arrive at a similar conclusion with a more recent dataset. Empirically, the level of social support from co-workers was found to be significant in many occupations (Alexander, Lichtenstein, Oh, and Ullman 1998; Brough and Frame 2004; Cortese, Colombo, and Ghislieri 2010; Ducharme and Martin 2000; Roxburgh 1999). This importance of the social support resource may come from it being a buffer against high job demands to prevent job strain and because it affects motivation and productivity, according to the Job Demands-Resources model (Bakker and Demerouti 2007). This paper proposes a way to incorporate JS in an agent-based model of labour market, explicitly making several factors affect individuals’ decision-making. The paper is structured as follows. Section 2 introduces our method of modelling JS. In Section 3, we put our JS model in the context of an artificial labour market. Section 4 provides the description of simulation setup

2. MODELLING JOB SATISFACTION To formally model job satisfaction (JS), we propose to separate it into two components: expected JS and current JS. The difference between the two is that the former can be measured for any job, as it depends only on the current situation in the firm hosting the job. The latter depends on the former but, in addition, incorporates a stochastic component tracking the experience of the person on the job. In this paper, we assume that expected JS, ‫ݏ‬௝௘ , is a function of the ratio of the wage of agent ݅ to his reservation wage, ‫ݓ‬௜௝ /‫ݓ‬௜௥ , and the ratio of the number of friends he has in the firm hosting the job (also referred to as “the number of local friends”) to the ௙ maximum number of friends he can have, ݊௜ /݊௜ . In other words, a person expects a certain level of monetary compensation and social support on the job. (Note that with the number of local friends we approximate a broad notion of social support rather than dividing it into support from management and from colleagues. More sophisticated frameworks should account for these two facets separately.) This is quite realistic, as normally, this is the only information individuals have before actually starting working in the firm. We also assume the following properties of expected JS: x Its partial derivatives with respect to both parameters are decreasing functions of the absolute values of the respective parameters. Thus, any next friend working in the same firm would add less to job satisfaction. The same would go for any next dollar of wage change; x Its range is bounded: ‫ݏ‬௝௘ ‫ א‬ൣ‫ݏ‬, ‫ݏ‬൧. Firstly, the level of satisfaction cannot be arbitrarily large or arbitrarily low. Secondly, this requirement is consistent with empirical data from surveys, where satisfaction is normally measured on a Likert scale, which would help in validating the model; x The same level of job satisfaction can be gained by different combinations of the relative wage and the number of friends. On the contrary, when the person starts working, many other parameters start influencing his current JS—working conditions, job demands, role clarity, and other facets. As already noted above, we assume that these factors are pure noise captured by a random disturbance ߝ~ܰ(0, ߪఌଶ ). The change in current JS in period ‫ݐ‬, therefore, is the sum of this random disturbance and the change in expected JS (due to changes in wage and the number of local friends): ௖ ௖ ௘ ௘ ൯ + ߝ௧ ‫ݏ‬௜௝௧ െ ‫ݏ‬௜௝,௧ିଵ = ൫‫ݏ‬௜௝௧ െ ‫ݏ‬௜௝,௧ିଵ (1) ௖ ௘ and ‫ݏ‬௜௝௧ should remain in ൣ‫ݏ‬, ‫ݏ‬൧. Note that both ‫ݏ‬௜௝௧

159

candidates receive acknowledgements. If a person receives acknowledgements for several applications, he chooses the one with the highest expected job satisfaction. He then sends an acknowledgement in reply to the chosen vacancy and starts working immediately. If the vacancy failed to attract a new employee, the hosting firm re-posts this vacancy in the next period, raising the proposed wage rate ‫ ݓ‬by the factor of ݄, which is the same for all firms. Required experience does not change, because the firm needs qualified personnel for its vacancies. Instead, the firm realises that the reason of the failure of the vacancy is a lack of motivation, that is, expected job satisfaction, which the firm can improve only by raising the proposed wage rate. If the re-posted vacancy also fails, it is completely removed from the vacancy list. For a working person, reservation wage is his current wage. For a person with no working experience, it is given by the minimum wage. For an unemployed, it is a decreasing function of his last wage and the length (in months) of the current unemployment period ‫ݐ‬௜௨ , ೠ (3) ‫ݓ‬௜௥ = ߮௧೔ ିଵ ‫ݓ‬௜ , ‫ݓ‬௜௥ ൒ ‫ݓ‬௠ The parameter ߮, 0 < ߮ < 1 is the same for all persons. Thus, the longer a person is unemployed, the lower wage rate he is ready to accept. Reservation wage, however, cannot fall below the minimum wage. We also model on-the-job search. If current JS falls below the minimal level, which is the same for everyone, the person starts seeking job as if he was unemployed. However, in this case, he only considers the appropriate vacancies with expected JS not lower than his current JS. If he is selected to fill a vacancy, he quits his current job and then starts working on the new position (it could be hosted by the same firm where he worked before).

3. SIMULATION CONTEXT To illustrate how the proposed job satisfaction model could be used in a real simulation, we incorporate it in the following artificial labour market. 3.1. General Characteristics The timing is discrete, one period representing one month, and 12 months constituting a year. Throughout the paper, we will use subscript ‫ ݐ‬to refer to the monthly periods and ߬ to the yearly ones. Most actions in the labour market, such as changes in job satisfaction and in the workforce of a firm, are made on a monthly basis. Changes in the population and in wages are made once a year. There are two types of agents in the economy: persons and firms. Initially, the economy is populated with ܰ଴ persons. Each year ߬ ൒ 1, ܰఛ new persons are added to the population, with the number of new entrants growing at a fixed rate of ݃, ܰఛ = ݃ܰఛିଵ. This can be regarded as an inflow of new secondary school graduates in the labour market. Persons are born with zero age and zero experience (including those in the initial population) and start seeking a job. They retire at the age of ܽ years, at which moment they are removed from the simulation. Firms, on the contrary, are assumed to live forever, the number of firms being fixed at ‫ܯ‬. 3.2. Job Search There is a unique vacancy list in the economy that is available to everyone for free. To find new labour, firms post vacancies on the vacancy list. Persons use the list to find new jobs. A vacancy is a three-tuple (݂, ‫ݔ‬, ‫)ݓ‬, where x ݂ is the firm hosting the vacancy; x ‫ א ݔ‬Ժ, 0 ൑ ‫ ݔ‬൑ ‫ ݔ‬is the required working experience measured in years, ‫ ݔ‬being the sufficient experience, which is common for all vacancies; x ‫ א ݓ‬Ժ, ‫ ݓ‬൒ ‫ݓ‬௠ is the proposed wage rate at the required experience ‫ݔ‬, ‫ݓ‬௠ being the minimum wage, which is the same for the whole economy. The proposed wage rate at experience ‫ݔ‬௜ ൑ ‫ ݔ‬is (2) ‫ݓ‬௜ = ‫ ݓ‬+ ‫ݔ(ݍ‬௜ െ ‫) ݔ‬, where ‫ ݍ‬is a constant equal for all vacancies. The proposed wage rate at experience ‫ݔ‬௜ > ‫ ݔ‬is given by the same equation with ‫ݔ‬௜ taken equal to ‫ݔ‬. A person in search for a job browses through the vacancy list and creates a sub-list of vacancies that require a working experience not higher than his experience and, for his experience (‫ ݍ‬is known by everyone), propose a wage rate not lower than his reservation wage. He then sends applications to ݇ vacancies from the sub-list with the highest expected job satisfaction, where ݇ is constant for all persons. A firm screens the applicant list for each of its vacancies and, if it finds its employees’ friends, it randomly chooses among them; in the other case, it chooses randomly from all applicants. Successful

3.3. Firm Inside Dynamics All firms start with no workforce. In the first month, each firm publishes ܰ଴ /‫ ܯ‬vacancies. Thus, at the beginning, all firms try to be of equal size. Each month, firms randomly change the size of their workforce by ߜ௙௧ persons, which is distributed uniformly in ൣെߜ, ߜ൧. If ߜ௙௧ < 0, the firm contracts its workforce by randomly firing หߜ௙௧ ห employees. If the change is positive, it publishes ߜ௙௧ vacancies. Each vacancy for a new position is created with the required experience ‫ ݔ‬uniformly chosen from [0, ‫] ݔ‬. Given that value of ‫ݔ‬, the corresponding proposed wage ‫ ݓ‬is set to the average wage currently received by the firm's employees having experience ‫ݔ‬. If no such persons are currently employed, the firm considers wages earned by the relevant employees who were working in the firm in the nearest month during the last year. If no such persons worked in the firm during the last year, it makes an interpolation from the point (0, ‫ݓ‬௠ ) using Eq. (2). A firm can also publish a vacancy that would substitute employee ݅ who just quit the firm, either

160

because of reaching the retirement age or due to a low job satisfaction (fired workers are not substituted). In this case, the vacancy is published with the required experience being two years smaller than that employee's experience, checking that the resulting experience is inside [0, ‫] ݔ‬. The proposed wage corresponding to the required experience is then set so that an applicant with experience equal to that employee's experience had the same proposed wage as that employee, correcting it if it falls below the minimum wage. This is summarized by the following equations: (4) ‫ݔ = ݔ‬௜ െ 2, 0 ൑ ‫ ݔ‬൑ ‫ݔ‬ (5) ‫ݓ = ݓ‬௜ െ ‫ݔ(ݍ‬௜ െ ‫) ݔ‬, ‫ ݓ‬൒ ‫ݓ‬௠ At the start of each year ߬ ൒ 1, each firm posts ߠ௙௧ ܰఛ vacancies characterised by the tuple (݂, 0, ‫ݓ‬௠ ), where ߠ௙௧ is the firm's current labour market share. Thus, firms try to hire a share of fresh graduates that is consistent with their current labour market share, providing these graduates with vacancies with the lowest experience requirements, but also proposing them the minimum wage. While we do not explicitly model production and selling, by placing a cap on the share of graduates that can be hired, we are preventing the situation when a small firm hires an arbitrarily large number of new workers, for which it may not have enough resources. For a firm's employee, wage can change only once a year—thus, we model stickiness of wages. Wages change as in the trinomial option pricing model, i.e., by a factor taken from {‫ݓ‬௨ , 1, ‫ݓ‬ௗ }, where ‫ݓ‬௨ > 1 and the corresponding probabilities ‫ݓ‬ௗ < 1 with ௨ ௡ ௗ} ௨ ௡ ௗ {‫݌‬௪ , ‫݌‬௪ + ‫݌‬௪ , ‫݌‬௪ , ‫݌‬௪ + ‫݌‬௪ = 1; all these parameters are fixed for all firms. In the beginning of the year, firms choose one of these factors and throughout the year, they change wages of all workers with expiring yearly contracts by this factor.

refuse the proposed friendship. If, due to additional ties created in the workplace, the number of friends exceeds the allowable ceiling, the person removes these extra friends. He first removes currently unemployed friends, starting with those with the longest period of unemployment. If this is not sufficient, he randomly removes friends who do not work with him in one firm. Finally, after there are no more such friends, he randomly removes his colleagues from his friendship circle. 4.

SIMULATION RESULTS

4.1. Simulation Setup We implemented a simulation of the model described in the two previous sections in Repast Simphony. Expected JS was represented as a sum of two logistic functions, which correspond to the three desirable properties stated in Section 2: ೑



଺௡







௘ ‫ݏ‬௜௝௧ = (1 െ ߣ௜ )ܲ ൬6 ൤௪ ೝ೔ െ 1൨൰ + 2ߣ௜ ቈܲ ቆ ௡ ೔ ቇ െ ଶ቉

(6)

In this formula, ܲ(ή) is the logistic function. In the definition of both summands, we took into account that ܲ(6) ൎ 1 and ܲ(െ6) ൎ 0. Thus, when ‫ݓ‬௜ ‫ݓ ا‬௜௥ , the first summand approaches zero, while it approaches one when ‫ݓ‬௜ = 2‫ݓ‬௜௥ . The second summand is zero when the person has no local friends and approaches one when he has all his possible friends working with him. We also take into account the importance of friends, ߣ௜ , so that the more important friends are for a person, the higher is the weight of the second summand. Since ߣ௜ ‫[ א‬0,1], ௘ ‫[ א‬0,1]. it follows that ‫ݏ‬௜௝௧ The maximum number of friends is determined according to the following linear relationship: (7) ݊௜ = ‫ڿ‬100ߣ௜ ‫ۀ‬ As 0 < ߣ௜ ൑ 1, Eq. (7) guarantees that maximum number of friends is distributed log-normally in the interval [1,100]. Table 1 reports the values of the simulation’s parameters. Annual population growth rate was taken approximately equal to the one characterising the situation in Europe in the last decade, as reported by Eurostat. Critical job satisfaction level is the level at which the person starts on-the-job search. Parameters of the friend importance distribution, ߤఒ and ߪఒ , were chosen so that the median importance is 0.07, leading to the median maximum number of friends equal to seven. The standard deviation of job satisfaction noise, ߪఌ , is not reported in Table 1, since we will compare model dynamics depending on the values of this parameter.

3.4. Social Network Dynamics According to Granovetter (2005, p. 34), “people have cognitive, emotional, spatial and temporal limits on how many social ties they can sustain.” Thus, maintaining a particular number of friends has an inherent cost for a person. We do not model such costs explicitly. Rather, we assume that each person has a maximal number of friends, ݊௜ , which depends on the importance of friends in his life, ߣ௜ ‫[ א‬0,1], which, in turn, is generated by ݈‫ߤ(ܰ݃݋‬ఒ , ߪఒଶ ). Lognormal distribution was chosen because it approximates the degree distribution in networks of friends quite well (see, e.g., Toivonen et al 2009). The functional form of ݊௜ (ߣ௜ ) can then be chosen so that maximal number of friends is distributed lognormally, too. The person starts his life with a random number of friends from his generation that does not exceed the maximal number of friends he is ready to make. These can be regarded as his school-friends. Coming to a new workplace, he tries to make new ȟ݊௜ = ‫݊ڿ‬௜ Τ10‫ ۀ‬friends working in the firm hosting this workplace. He succeeds in creating a friendship tie with a random firm's employee with probability 1Τ2, as the employee can

4.2. Analysis As the retirement age was set to 20 years and initially, everyone is of age zero, the size of the population grows rapidly until year ߬ = 20, at which time the persons born in the first periods start retiring and the labour force size grows much less rapidly. Thus, we analyse only the last 20 years of the simulation, when the artificial labour market should have already stabilised.

161

Table 1: Simulation Parameter Values Parameter — Simulation length (in months) ‫ ܯ‬Number of firms ܰ଴ Initial population size ݃ Annual population growth rate ܽ Retirement age (in years) ‫ݓ‬௠ Minimum wage ݄ Wage change factor for failed vacancy ‫ݓ‬௨ Wage increase factor ௨ Probability of wage increase ‫݌‬௪ ‫ݓ‬ௗ Wage decrease factor ௗ Probability of wage decrease ‫݌‬௪ ߤఒ Mean of friend importance ߪఒ Std. dev. of friend importance ߜ, ߜ Workforce change boundary ߮ Reservation wage modifying factor ‫ ݔ‬Sufficient experience (in years) ݇ Number of simultaneous applications ‫ ݍ‬Wage-experience multiplier — Critical job satisfaction level (%)

component of expected JS being less than that of persons with lower friend importance—other things equal, the former are less happy than the latter because they do not meet their needs for social interaction. Consequently, we expect persons with higher friend importance to change jobs more frequently. Logistic regressions of leaving the job because of a low current JS (see Table 2) confirm this: in all four models, the largest effects in absolute terms are from friend importance and squared experience, and friend importance effects are positive. Note that the effect of friend importance on the probability of quitting the job increases if JS is based solely on the wage component.

Value 480 20 200 1.005 20 100 1.1 1.05 0.6 0.95 0.1 -2 0.8 5 0.9 10 5 1.1 20

Table 2: Marginal Effects after Logit Regression of Leaving the Job Med noise Med noise, Low noise Low noise, Model .010*** Friend imp. # local friends .000*** .000*** Wage .001*** Age 2 .004*** Age /100 *** -.003 Experience 2 Exper. /100 -.009*** .0848 Pseudo-R2

We compare four setups of the model, which differ, firstly, on the values of current JS noise: medium (ߪఌ = 0.1) vs. low (ߪఌ = 0.05), and secondly, on whether JS incorporates the local friend component (the second summand in Eq. (6)) or it consists of the wage component only. In models where friends are irrelevant to JS, persons still make friends and firms continue to hire by referral. Persons simply do not take the number of local friends into account when choosing among vacancies or considering starting on-the-job search. Figure 1 compares the distribution of friend importance (which is identical to the distribution of the maximum number of friends) with the actual number of friends in the last period of the model. We can observe that the friend distribution is more peaked and has a thinner tail than the friend importance distribution.

***

*

.007*** .000** .000*** .001*** .003*** -.003*** -.008*** .0806

wage only .012*** -.001*** .000*** .000* .002*** -.004*** -.002** .1210

‫ < ݌‬0.01 ‫ < ݌‬0.05 ‫ < ݌‬0.1 The table shows median values for marginal effects and median pseudo-R2s for runs of each model. Standard errors allow for intragroup correlations, where a group is defined as all observations belonging to one person.

To see whether friend importance systematically differs by the labour status of a person, we ran Kolmogorov-Smirnov two-sample tests on the runs of each of the four setups. The tests show that in the models where JS contained information on friends, the unemployed generally have a lower friend importance than the employed. For the two models where JS is based solely on wages, however, the situation is reversed.

0

0

.2 .19 .18 .16

.17

5

5

Percent

Percent

10

10

Unemployment Rate

.21

.22

15

15

**

wage only .013*** -.001*** .000*** .000*** .002*** -.004*** -.004*** .1232

0

.2

.4 .6 .8 Friend Importance

1

0

20 40 60 80 Actual Number of Friends

20

100

25 Medium noise Low noise

Figure 1: Distributions of friend importance and actual number of friends for the last period of the simulation.

30 Year

35

40

Medium noise, wage only Low noise, wage only

Figure 2: Median annual unemployment rates for the runs of the four setups.

This means that persons with a high maximum number of friends generally fail to make that many friendship ties, which should lead to their second

Annual unemployment rates, however, do not differ much between the groups with differing JS

162

Table 5: Correlation coefficient between person’s wage and his friends’ average wage Med noise Med noise, Low noise Low noise, Model

functions—in all four cases shown in Figure 2, median unemployment rates are in the narrow interval (0.18,0.20). To check whether the difference in friend importance between the employed and the unemployed is significant in real terms, we compare the average of this characteristic for the two groups in each of the four model setups (see Table 3). The table shows that the differences between the groups are minor.

***

Our final check for the link between friend importance and unemployment concerns the length of the longest period of unemployment experienced by the person. Results (see Table 4) show that, firstly, higher friend importance tends to reduce time to find the job, and secondly, that this reduction is much larger when JS is based on wages only. Note also that for wage-only JS models, regression fit is considerably lower than for the other two models. Another result is that the lower is JS noise, the more pronounced is the friend importance effect. Table 4: Regression of the length of the longest period of unemployment experienced by the person Med noise Med noise, Low noise Low noise, Model

‫ < ݌‬0.01

wage only -1.448** -4.082*** 4.338*** 3.061*** -1.792*** -5.474*** -5.914*** -3.296*** 8.166*** 8.759*** 21.570*** 18.004*** .2736 .196

Model

‫ < ݌‬0.01 ‫ < ݌‬0.05 ‫ < ݌‬0.1 The table shows median values for regression coefficients and median R2s for runs of each model.

Medium noise

Next, we check how the situation with person’s friends relates to his own situation in the labour market. We find that the correlation coefficient between a person’s wage and his friends’ average wage is positive and quite high in all four model specifications (see Table 5), meaning that there is a certain degree of income homophily among persons. The coefficient is generally higher when friends are taken into account in JS. The coefficient value increases slightly with JS noise variance; moreover, for a lower noise, the difference in the correlation coefficient between JS specifications becomes larger.

Medium noise, wage only Low noise Low noise, wage only a

Firm Size

Avg. No. of Local Friends

Table 6: Firm-Level Statistics Avg. Number of Friends

*

wage only 0.780***

Average Friend Importance

**

0.809***

Number of Firms

***

wage only -3.691*** 2.821*** -5.098*** -3.173*** 8.374*** 17.846*** .1870

wage only 0.816***

Finally, we analyse situation within firms. We divide them into three groups by the average number of employees in the last month of each of the last 20 years of the simulation (see Table 6). While all firms were of the same size initially, several large (>500 employees) and medium-sized (50-500 employees) firms have evolved in the artificial economy. Note that with a lower JS noise, more large and medium firms evolve in the standard JS function setup. That can be explained by current JS changing less and, thus, decisions to quit taken less often—as a result, persons work longer at the same jobs, and firms can better accumulate workforce. Large firms have a higher average friend importance than medium and small firms in the models where JS incorporates friends, while the situation is reversed when JS depends on wage only, and the magnitude of JS noise does not affect the results. The same holds for average total number of friends of firms’ employees. Average number of local friends, on the contrary, is always greater, the larger the company, which was also expected. Note, however, that once friends become a component of current JS, the number of local friends in large and medium-sized companies increases 1.5 to 2 times, at the expense of a minor decrease of this characteristic for small companies. In other words, friendship networks are more clustered within companies than in models where JS depends only on wages.

Table 3: Mean Friend Importance by Labour Status Friend Importance Model Employed Unemployed Medium noise 0.18 0.17 Medium noise, wage only 0.18 0.19 Low noise 0.18 0.17 Low noise, wage only 0.18 0.19

Friend imp. -1.356* 4.025*** Age 2 .826*** Age /100 Experience -5.214*** Exper.2/100 5.527*** 21.117*** Constant 2 .2887 R

0.819***

Corr. coeff.

Large a Medium b Small c Large a Medium b Small c Large a Medium b Small c Large a Medium b Small c

2 2 16 2 4 14 3 3 14 2 4 14

.18 .15 .15 .17 .22 .22 .19 .15 .15 .17 .21 .21

15.47 12.93 12.03 14.38 18.93 18.50 15.84 12.83 11.93 14.28 17.99 18.26

10.27 3.76 1.68 6.10 2.62 1.82 10.75 4.42 1.67 5.67 2.60 1.80

Number of employees > 500. Number of employees in [50,500]. c Number of employees < 50. The table shows average results for runs of each model. b

163

REFERENCES Abdou, M., Gilbert, N.. 2009. Modelling the Emergence and Dynamic of Social and Workplace Segregation. Mind & Society, 8(2), 173-191. Acker, G.M., 2004. The Effect of Organizational Conditions (Role Conflict, Role Ambiguity, Opportunities for Professional Development, and Social Support) on Job Satisfaction and Intention to Leave Among Social Workers in Mental Health Care. Community Mental Health Journal, 40(1), 65-73. Alexander, J.A., Lichtenstein, R., Oh, H.J., Ullman, E., 1998. A Causal Model of Voluntary Turnover Among Nursing Personnel in Long-Term Psychiatric Settings. Research in Nursing & Health, 21(5), 415-427. Bakker, A.B., Demerouti, E., 2007. The Job DemandsResources Model: State of the Art. Journal of Managerial Psychology, 22(3), 209-238. Bewley, T.F., 1999. Why Wages Don't Fall During a Recession. Cambridge, MA: Harvard University Press. Bramoullé, Y., Saint-Paul, G., 2010. Social networks and labor market transitions. Labour Economics, 17(1), 188-195. Brough, P., Frame, R., 2004. Predicting Police Job Satisfaction and Turnover Intentions: The role of social support and police organisational variables. New Zealand Journal of Psychology, 33(1), 8-16. Calvó-Armengol, A., Jackson, M.O., 2007. Networks in labor markets: Wage and employment dynamics and inequality. Journal of Economic Theory, 132(1), 27-46. Carless, S.A., Arnup, J.L., 2011. A longitudinal study of the determinants and outcomes of career change. Journal of Vocational Behavior, 78(1), 80-91. Cortese, C.G., Colombo, L., Ghislieri, C., 2010. Determinants of nurses’ job satisfaction: the role of work–family conflict, job demand, emotional charge and social support. Journal of Nursing Management, 18(1), 35-43. Ducharme, L.J., Martin, J.K., 2000. Unrewarding Work, Coworker Support, and Job Satisfaction: A Test of the Buffering Hypothesis. Work and Occupations, 27(2), 223-243. Gemkow, S., Neugart, M., 2011. Referral hiring, endogenous social networks, and inequality: an agent-based analysis. Journal of Evolutionary Economics, 1-17. Granovetter, M., 2005. The impact of social structure on economic outcomes. Journal of Economic Perspectives, 19(1), 33-50. Harris, J.I., Winskowski, A., Engdahl, B.E., 2007. Types of workplace social support in the prediction of job satisfaction. The Career Development Quarterly, 56(2), 150-156. Hazans, M., 2011. Labor market integration of ethnic minorities in Latvia. In M. Kahanec, K.F. Zimmerman, ed. Ethnic diversity in European

5. CONCLUSION In the present paper, we proposed a way to model job satisfaction and to incorporate it into an agent-based model of labour market. In our model, job satisfaction depends on two components: monetary benefits and social support, which were found to be empirically important factors; the influence of other factors is gauged by a random disturbance term. We created an artificial labour market simulation with heavy usage of social networking—during referral hiring, in choosing among vacancies, and in considering whether to start on-the-job search. The two latter choices are actually made based on job satisfaction. We found that friend importance is an important determinant of the probability to quit the job and of the length of the longest unemployment period; both effects increase in absolute terms when job satisfaction does not depend on social support. We also found evidence of social clustering. Firstly, labour market dynamics resulted in the emergence of a small number of large and mediumsized firms. Secondly, there is a substantial positive correlation between friends’ wages, indicating income homophily of social groups. While, naturally, the average number of a person’s friends working in the same firm increases with firm size, this number is nearly two times higher when job satisfaction depends both on wages and on social support than when it depends on wages only. The model presented here has several limitations. Firstly, it does not distinguish among several components of social support, modelling it as a single factor. Moreover, it assumes that job satisfaction determinants other than monetary compensation and social support are pure noise, while they might be firm/job-specific and show serial correlation. Secondly, it portrays the behaviour of small and large firms analogically, while in a large firm, overall social support may be less important than social support in the department where the person works—thus, it may be useful to model the effects of social ties between workers of the same department and of different departments differently. In addition, empirical findings show that large firms rely on referral hiring to a lower extent than the small ones. Thirdly, it needs to be verified whether the same results hold when the production-consumption decisions are incorporated in the model. Further research should aim to overcome these limitations. ACKNOWLEDGMENTS The research was supported by the European Social Fund Project No. 2009/0138/1DP/1.1.2.1.2/09/IPIA/ VIAA/004 (“Support for Doctoral Studies at University of Latvia”). The author also thanks two anonymous reviewers for their helpful comments.

164

labor markets: Challenges and solutions. Cheltenham, UK: Edward Elgar. Kalleberg, A.L., 1977. Work Values and Job Rewards: A Theory of Job Satisfaction. American Sociological Review, 42(1), 124-143. Krauth, B.V., 2004. A dynamic model of job networking and social influences on employment. Journal of Economic Dynamics and Control, 28(6), 1185-1204. Kuddo, A., 2009. Employment Services and Active Labor Market Programs in Eastern European and Central Asian Countries. Washington DC: World Bank. Manger, T., Eikeland, O.-J., 1990. Factors predicting staff's intentions to leave the university. Higher Education, 19(3), 281-291. Montgomery, J.D., 1991. Social networks and labormarket outcomes: Toward an economic analysis. American Economic Review, 81(5), 1408-1418. Parry, J., 2008. Intention to leave the profession: antecedents and role in nurse turnover. Journal of Advanced Nursing, 64(2), 157-167. Roxburgh, S., 1999. Exploring the Work and Family Relationship: Gender Differences in the Influence of Parenthood and Social Support on Job Satisfaction. Journal of Family Issues, 20(6), 771788. Tassier, T., Menczer, F., 2001. Emerging small-world referral networks in evolutionary labor markets. IEEE Transactions on Evolutionary Computation, 5(5), 482-492. Tassier, T., Menczer, F., 2008. Social network structure, segregation, and equality in a labor market with referral hiring. Journal of Economic Behavior & Organization, 66(3-4), 514-528. Toivonen, R., Kovanen, L., Kivelä, M., Onnela, J.-P., Saramäki, J., Kaski, K., 2009. A comparative study of social network models: Network evolution models and nodal attribute models. Social Networks, 31(4), 240-254. AUTHOR’S BIOGRAPHY Alexander Tarvid is a PhD student at the University of Latvia. His research focuses on higher education policy, choice of field of higher education and its impact on the individual’s position in the labour market, in particular, on unemployment, shadow employment, and overeducation. He also does research on the application of agent-based simulations to the modelling of dynamics in the labour market, including various effects of social networks. In 2010-11, he participated in the World Bank project Multi-Country Policy Study of Unregistered Employment and the Shadow Economy as research assistant.

165

SIMULTANEOUS SCHEDULING OF MACHINES AND OPERATORS IN A MULTIRESOURCE COINSTRAINED JOB-SHOP SCENARIO Lorenzo Tiacci(a), Stefano Saetta(b) (a) (b)

Dipartimento di Ingegneria Industriale – Università degli Studi di Perugia Via Duranti, 67 – 06125 Perugia - Italy (a)

[email protected], (b)[email protected]

2009; Manikas and Chang 2008); Tabu search (Zhang, Li, Guan and Rao 2007)). Complexity increases in dual resource constrained problems, and extending these often quite complex heuristics to more realistic scenarios is usually not straightforward. Dauzère-Pérès, Roux and Lasserre (1998) developed a disjunctive graph representation of the multi-resource problem and proposed a connected neighborhood structure, which can be used to apply a local search algorithm such as tabu search. Matie and Xie (2008) developed a greedy heuristic guided by a genetic algorithm for the multi-resource constrained problem. However, in most real-world environments, scheduling is an ongoing reactive process where the presence of a variety of unexpected disruptions is usually inevitable and continually forces reconsideration and/or revision of pre-established schedules (Ouelhadj and Petrovic 2009). Most of the above-mentioned approaches have been developed to solve the problem of static scheduling and are often impractical in real-world environments, because the near-optimal schedules with respect to the estimated data may become obsolete when they are released to the shop floor. As a result, Cowling and Johansson (2002) addressed an important gap between scheduling theory and practice, and stated that scheduling models and algorithms are unable to make use of real-time information. A quick, intuitive, and easy to be implemented method for dynamic scheduling is utilizing priority (or dispatching) rules. The application of priority rules gives raise to a completely reactive scheduling, where no firm schedule is generated in advance and decisions are made locally in real-time. A priority rule is used to select the next job with highest priority to be assigned to a resource. This is done each time the resource gets idle and there are jobs waiting. The priority of a job is determined based on job, machine or in general resources attributes. Priority-scheduling rules have been developed and analyzed for many years (Haupt 1989, Blackstone, Philips and Hogg 1982, Rajendran and Holthaus 1999, Geiger, Uzsoy and Aytu 2006, Geiger and Uzsoy, 2006). Although priority rules have also been applied to

ABSTRACT In the paper the simultaneous scheduling of different types of resources is considered. The scenario is constrained by machines and human resources, and its complexity is increased by the presence of two types of human resources, namely the equipper, that performs only an initial action of each task (the ‘setup’), and the normal operator, that loads and unloads each piece from machines. A conceptual model of the shop is build in order to simultaneously handle priority rules for each one of the three types of resources considered (machine, equipper and operator). A simulation model has been implemented and a simulation experiment performed in order to explore the effect on mean flow factor reduction of different combination of priority rules. Keywords: dual-resource constraints, multi-resource constraints, priority rule scheduling, job shop control.

1.

INTRODUCTION

Job shop scheduling has attracted researchers for many decades, and still now is one of the most studied subjects in literature related to industrial problems. However, multi or dual resource constrained scheduling problems are significantly less analyzed, although being more realistic (Scholz-Reiter, Heger and Hildebrandt 2009). ElMaraghy, Patel and Abdallah (2000) defined the machine/worker/job scheduling problem as: “Given process plans for each part, a shop capacity constrained by machines and workers, where the number of workers is less than the number of machines in the system, and workers are capable of operating more than one machine, the objective is to find a feasible schedule for a set of job orders such that a given performance criteria is optimized”. Optimal solution are difficult to find also for the single resource scheduling problem, so that many heuristics approaches have been used in literature to find good but non optimal solutions for the machine constrained problem. These approaches include: Simulating annealing (Laarhoven, Aarts and Lenstra 1992); Genetic algorithms (Zhou, Cheung and Leung

166

x

dual-resource constrained problems (Scholz-Reiter, Heger and Hildebrandt, 2009), there are no studies in literature that deal with the presence of different types of human resources, each one competent to perform a specific action of the job cycle. In fact, resources heterogeneity is usually considered just in terms of different work efficiency of resources on different tasks. In this work we analyze a multi-resource constrained job-shop scenario in which scheduling is constrained by machines and by two types of human resources, namely ‘equippers’ and ‘operators’. Equippers and operators do not perform the same action with different efficiency, but are assigned to completely different and non-overlapping actions related to the job cycle. A conceptual model of the company’s shops is built in order to simultaneously handle priority rules for each one of the three types of resources considered (machine, equipper and operator). A simulation model has been built and a simulation experiment performed in order to explore the efficacy on flow factor reduction of different combination of priority rules. The paper is organized as follows. The job shop scenario is described in section 2. In section 3 the conceptual model of the shops is illustrated. Section 4 deals with the implementation of the simulation model, while in section 5 the simulation experiment is described and results are discussed. In section 6 conclusions are drawn. 2.

x x

 Task sequence Task 1 Turning        2 Turning        3 Milling        4 Milling        5 Control

THE JOB-SHOP SCENARIO

The scenario is representative of a real case study of a manufacturing company in the field of precision metal and mechanical processing. The company is specialized in the production of very complex components for industrial, aeronautical and aerospace applications. In the aerospace and aeronautical fields, the company produces 1/A class components such as, for example, actuators, stabilizers, worm gears, landing devices, turbine’s bearing rings and axle rotors. In the industrial sector, the company produces high quality components for machine tools and laser cutting.



 

JOB1(lotsize:9) AREA AREA1 ACTION TIME(min) SetͲup 0.85 Load 2.5 Run 5 Unload 2.5 Inspection 5 AREA1 ACTION TIME(min) SetͲup 131.45 Load 2.5 Run 3 Unload 2.5 Inspection 5 AREA2 ACTION TIME(min) SetͲup 203.92 Load 2.5 Run 45 Unload 2.5 Inspection 5 AREA2 ACTION TIME(min) SetͲup 202.93 Load 2.5 Run 45 Unload 2.5 Inspection 5 CONTROLAREA  ACTION TIME(min) Control 15

Figure 1: Example of data representing a job. 2.3. Machines In each area of interest (areas 1,2 and 3) there are computer numerical control (CNC) machines. Each machine can be equipped with a variable set of tools that allow completing the run with no interruptions. Every time a job is changed, the set of tools have to be changed depending on the new task requirements, and the controlling software has to be appropriately programmed. Then each piece belonging to the job has to be loaded, processed (run), and unloaded. After the first piece of a job has finished its run and has been unloaded, it must be inspected before that the remaining pieces of the lot can start being processed (see Figure 2).

2.1. Areas The company is organized in different areas, in which there are homogeneous machines. Every job assigned to a certain area can be processed indifferently in one of the machine belonging to that area. There are 5 areas: the cutting area, area 1 (turning), area 2 (milling), area 3 (drilling), and a control area. The cutting area and the control area are not critic for the scheduling problem, because resources assigned to these areas do not constrain the solution. However, they have been considered in our model in order to get a realistic representation of the flow time of each job.

2.4. Equippers The machine set-up is performed by the equipper operator at the beginning of each task, before processing the first piece of the lot. The inspection action on the first piece is also performed by the equipper, that controls if the run has been properly executed. If everything is ok, the other pieces of the lot can start to be processed, and load and unload operations are then carried out by the normal operator,

2.2. Jobs Each job is represented by (see Fig. 1): x

a set of tasks that have to be performed on each piece of the job, and the associated area; the sequence of tasks that have to be performed; the processing times of actions connected to each task.

a quantity of pieces that have to be processed (lot size);

167

without the need of the participation of the equipper. The equippers assigned to a certain area are able to equip all the machines inside that area.

2.7. The company’s scheduling process Production scheduling is performed through a commercial software that considers different types of resources, such as machines, operators, equipments, transporters etc. However, the scheduling is done primarily only considering the machines as limiting resource. Then, the resulting schedule is verified, checking if the capacity constraints related to the other resources are respected. In case of capacity shortages, the solution can be manually modified, for example by introducing overtime, or re-calculated, by modifying jobs due dates. This approach however does not allow a real simultaneous scheduling of machines and human resources, and the ‘trial and error’ nature of the procedure make it quite rigid, and unsuited to reacting to uncertainties of the environment. Furthermore, in the specific case study, the equipper (the most specialized human resource) is not always available during the day (see Table 1) and this makes it the primary candidate for being the limiting resource in many situations. In the next paragraphs, we describe an alternative way to approach the scheduling problem, based on dispatching rules. The approach developed is based on the fact that, besides the rules considered to assign jobs to machines, it is necessary to simultaneously consider the rules to assign operators and equippers to jobs.

2.5. Operators The normal operator performs loading and unloading of the pieces of a job. Processing starts directly after the machines are loaded. Unloading begins after processing, but if there is no operator available for unloading, the machine stays idle. The operators are not needed during processing and can work on other machines in that time period. The operators assigned to a certain area are able to load/unload all the machines inside that area.

Equipper Operator Machine the job starts setup

setup Load

1st piece

Load

run

unload inspection

unload inspection

Load

2nd piece

Load

time

run

unload

unload

Load

Load

3nd piece

3.

The systems has been modeled through a series of queues, some of which are ordered following different possible priority rules, through which decide the pickup order of the elements. The logical flow of entities in the systems is depicted in Fig. 3 (where AREA 2 is considered). There are 4 types of queues in each area, namely: PQ1, PQ2, VQ1 and VQ2. When a new job arrives, it tries to enter the area corresponding to the task that has to be performed. If all the machines in that area are busy, the job has to wait in a queue of type PQ1. When one of the machines in the area is free, the job is assigned to that machine, and the lot is divided into a number of pieces equal to the lot size. Pieces are then allocated in the PQ2 queue of the assigned machine (input buffer). Pieces in PQ2 queues may claim an equipper or an operator, depending on the action needed to complete their current task. If they claim an equipper, they also enter the virtual queue VQ1, which is served from the equippers of the area; if they claim an operator, they also enter the VQ2, which is served from the operators of the area. When an equipper or an operator is available, pieces are removed from VQ1 or VQ2, and the required action on the pieces are performed. When a task of a job has been completed, i.e. the last piece of the lot has been unloaded from the machine, the machine is released, and the job tries to enter the area corresponding to the next task of its processing sequence.

run

unload

unload





Figure 2: Actions involving the different resource types. 2.6. Shifts An important feature of the scenario is that shifts are different between Equippers and Operators. The operators work is organized in three shifts per day, each shift during 8 hours. Equippers work on a single shift per day, from 8.00 to 17.00, with an interval of one hour between 13.00 and 14.00. Thus, while operators (as machines) are available during all the 24 hours, equippers are available only in the central part of the day. The number of resources per shift in each area is reported in Table 1. Table 1: Resources Availability

Machines 24h/day

Area 1 Area 2 Area 3

4 4 3

Equippers 8.00-13.00 14.00-17.00 2 1 2

THE CONCEPTUAL MODEL

Operators 24h/day 3 1 2

168

Considering each area, we classify the queues into physical and virtual queues.

Note that while rules 1 to 5 are related to a local characteristic of the queue, rules LMQL and SMQL are taken on the basis of the length of PQ2 queues.

3.1. Pysical Queues (PQ) x PQ1. The first physical queue is related to jobs that are waiting for entering an area. The queue is physical because we can associate a job to the lot that is waiting (in a trolley for example) in a certain part of the shop. Jobs are picked up from this queue as soon as one of the machines of the area is available to work. Each area has one queue of type PQ1. x PQ2. The second physical queue is related to pieces of jobs that are waiting to be processed by a machine, i.e. they belong to a job that has already been assigned to a machine, and are waiting in the input buffer of the machine. Each machine has one queue of type PQ2.

AREA 2 P

E

P

VQ1

M P

P

P

P

P

P

PQ2

M P

P

P

P

PQ2 J

J

J

M

PQ1

3.2. Virtual queues (VQ) x VQ1. The first virtual queue is related to pieces of jobs that are waiting for an equipper, i.e., the first pieces of a job that have been already assigned to an available machine and are waiting or for the setup action on the machine, or for the inspection action. Each area has one queue of type VQ1. x VQ2. The second virtual queue is related to pieces of jobs that are waiting for an operator, i.e., pieces belonging to jobs already initiated, and waiting for loading or unloading actions. Each area has one queue of type VQ1. It is noteworthy that VQ1 and VQ2 are not physical queues. In particular, elements waiting in VQ1 can be physically in the input buffer of a machine (the first piece of a job waiting for the setup action) or in the machine itself (the first piece of a job waiting for the inspection action). Similarly, elements waiting in VQ2 can be physically in the input buffer of a machine (waiting for load action) or inside the machines (waiting for unload action).

P

P

P

P

P

P

PQ2

M P

P

P

P

PQ2

P

P

O

VQ2

Figure 3: Queues and logical flow (J = job; P = piece)

4.

THE SIMULATION MODEL

The simulation model has been built with Arena 11, using basic process, advanced process, advanced transfer, calendar schedules, and animation tools. In the next section, the main part of the model implementation is described. 4.1. The simulation model implementation The basic entity is one piece of a job. Pieces are batched, and batches move through the system when they move between different areas. A batch is separated into pieces when the job enters an area and pieces have to be processed in the machines. Machines, Operators and Equippers are modeled as Resources. The number of operators and equippers in each shift (see Table 1) is modeled using the Calendar Schedule utility of Arena, through which the detailed time patterns and the capacity of each resource can be set up.

3.3. Priority rules Priority rules are defined to select elements waiting in PQ1, VQ1 and VQ2 queues. These queues are the ones to which priority rules are applied, because elements of these queues claim a resource: PQ1 - machines, VQ1 equippers and VQ2 - operators. The priority rules for PQ1 (machines) are: 1. FIFO (First-IN, First-OUT); 2. LIFO (Last-IN, First-OUT); 3. SPT (Shortest Processing Time); 4. LPT (Longest Processing Time); 5. RANDOM; In addiction to these 5 rules, we considered two additional decision rules for VQ1 (equippers) and VQ2 (operators): 6. LMQL (Longest Machine Queue Length); 7. SMQL (Shortest Machine Queue Length).

Creation

Attributes

Read job data

Batch

0 0

Figure 4: Pieces and job creation

169

A piece is created and a series of attributes are assigned to it (e.g. the job code, the current time). Then other attributes related to the job (lot size, tasks, actions times, tasks sequence etc.) are read from file. Pieces are batched according to the lot size of the job, and the batch proceeds to the area of destination, according to its task sequence. The Route and Station modules of the Advanced Transfer tools of Arena have been utilized to perform entities movements throughout the model. The system is modeled through different sets of Station: Area Stations, Machine Stations, Equipper and Operator Stations, Actions Stations.

0 MACHINE M1

0

0

Counter

Assign First Piece of the Job Attribute

Assign Machine M1

False

0 PQ2 M1

Assign number of pieces in PQ2 M1

First Piece

0

True

GO T O EQUIPPER

False

RESET COUNTER

GO T O OPERAT OR

Figure 6: Machine Stations. When a piece exits the PQ2 queue, an attribute describing the length of that queue is assigned. This will allow ordering the succeeding VQ1 or VQ2 queues on the basis of this attribute, in order to implement rules 6 and 7. After that, another Decide module sends the entity to the Equipper or the Operator Station modules, depending on the ‘First Piece’ attribute. 4.1.3. Equipper and Operator Stations There are one Equipper Station and one Operator Station for each area of the shop floor. When a piece arrives to the Equipper Station of an area (Figure 7) it enters the Seize module corresponding to the VQ1 queue of the area. The order of the queue can be set according to one of the priority rules 1-7. The Seize module is associated to the Equipper resource of the area, whose capacity is set up through the Calendar Schedule utility. If at least one equipper is available, the piece exits the queue and seizes the equipper, which will be then released by the same entity when the action performed by the equipper (set up or inspection) has been performed. The Decide module directs the entity to the Station corresponding to the next action to be performed on the machine. This is possible thanks to the ‘current machine’ attribute (previously assigned to the entity), and another entity attribute, which is updated during the model execution, that describes the next action that the entity has to perform.

GO TO MACHINE M1

GO TO MACHINE M2

PQ1 AREA 3

True

First Piece?

4.1.1. Area Stations An Area Station is associated to each area of the shop floor. When a batch arrives at the Area Station (e.g. Area 3 in Figure 5), it enters PQ1 queue of the area. The order of the queue can be set according to one of the priority rules (1-5) defined in section 3.3. A batch can leave the queue according to a ‘scan for condition’ rule. The condition is verified when at least one of the PQ2 queues of the machines in the area is empty. The decide module directs the batch to the Station related to the available machine.

AREA 3

Separate

Decide GO TO MACHINE M3 E lse

Dispose

Figure 5: Area Stations. 4.1.2. Machine Stations A Machine Station is associated to each machine of an area. When the batch arrives to the machine (e.g. machine M1 in Figure 6), it is separated into pieces. A Counter module and a Decide module allow to identify the first piece of the job (the related attribute ‘First Piece’ is associated to the entity). An attribute (named ‘Current Machine’), related to the machine to which the piece has been assigned, is also stored. Then the entity moves to the Seize module corresponding to the PQ2 queue of the machine. The seize module is associated to a virtual resource of fixed capacity equal to 1, that is seized when the entity exits the queue, and is released by the same entity when all the actions associated to its task on the machine have been performed (see later in section 4.1.4). The introduction of this virtual resource is necessary because when the entity exits the queue and seizes the resource, this means that one machine is available to work, but not that the machine will be immediately seized: in fact, the piece could have to be waiting for an equipper or an operator to be loaded in or unloaded from the machine. So, the seize module cannot be associated to the machine resource if one want to accurately evaluate the actual machine utilization.

GO TO SETUP M1

GO TO SETUP M2

EQUIPPER Station

VQ1 Area 3

GO TO SET UP M3

Action and Machine?

E lse

USO USO USO USO USO USO

M1==100& & C ON TR OLLO==0 M2==100& & C ON TR OLLO==0 M3==100& & C ON TR OLLO==0 M1==100& & C ON TR OLLO==100 M2==100& & C ON TR OLLO==100 M3==100& & C ON TR OLLO==100

GO TO INSPECTION M1

GO TO INSPECTION M2 Dispose 7

0

GO TO INSPECTION M3

Figure 7: Equipper Stations. If the piece is directed to the Operator Station (Figure 8), it follows a very similar path. Here the queue is the VQ2 queue, which can be ordered according to

170

priority rules 1-7. The associated resource that will be seized is one operator of the area.

associated to the PQ1 queue of the machine (to allow a new piece to be loaded by an equipper or an operator, see Figure 6). The opportune ‘next action’ and ‘next area’ attributes are stored, and finally the piece enters the batch module. A similar path is followed by the first piece of a job that arrives to the inspection Station. The only difference is that the entity is delayed for a time equal to the inspection time, and then releases the equipper. When all the pieces of the job have been processed, the batch is ready to be routed to the next area of destination.

GO TO LOAD M1

GO TO LOAD M2

GO TO LOAD M3 OPERATOR Station

Decide 5

VQ2 AREA 3

GO TO UNLOAD M1 Else

GO TO UNLOAD M2

GO TO UNLOAD M3

Dispose 9

0

4.2. Verification and validation During the time for the simulation model realization, many meetings with company’s managers have been organized. For the valid modelisation of the human resources (operators and equippers) and the possible scheduling logic that could be implemented, the continuous confrontation with company’s staff during the model development has been very profitable. In this way, the essential aspects of the scheduling and the production processes have been outlined by those which operate in the day by day operations activities in the company. This confrontation also brought to renounce adopting complicated approach that are often studied by a theoretical point of view, but that are scarcely applicable to real cases. This allowed also to gain the company’s management accreditation for the use of simulation for the specific purpose of searching for alternative scheduling techniques with the aim to reduce the jobs mean flow factor. The conceptual model has been validated by the operational experts of the company: they confirmed that the assumptions underlying the proposed conceptual model were correct and that the proposed simulation design elements and structure (simulation’s functions, their interactions, and outputs) would have lead to results realistic enough to meet the requirements of the application. After the implementation, the same experts, comparing the responses of the simulation with expected behaviours of the system, confirmed that those responses were sufficiently accurate for the range of intended uses of the simulation. We also verified our model through two widely adopted techniques (see Law and Kelton, 2000). The first one consists in computing exactly, when it is possible and for some combination of the input parameters, some measures of outputs, and using it for comparison. The second one, that is an extension of the first one, is to run the model under simplifying assumptions for which its true characteristics are known and, again, can easily be computed. Furthemore, in order to check the correct implementation of dispatching rules logic, the animation capability of Arena has also been exploited (see Figure 10).

Figure 8: Operator Stations. 4.1.4. Action Stations There are four Action Stations for each machine. Each station is related to a determined action performed by equippers or operators on the machine, namely: set up, load, unload and inspection (see Figure 9). A piece that arrives at the setup Station is necessarily coming from the equipper Station (Figure 7), where it had seized one equipper. The entity now seizes the machine, and is delayed by a time equal to the setup time. Then it releases the equipper (but not the machine) and is routed toward the Operator Station. A piece that arrives at the load Station it is necessarily coming from the Operator Station (Figure 8), where it had seized one operator. If the piece is the first one of the job, the machine must not be seized, because it has already been seized during the setup. The entity is here delayed for an amount of time equal to the load action, after which releases the operator (but not the machine). The succeeding delay module corresponds to the run action performed by the machine. An attribute specifying the next action to be performed (unload) is stored before the entity is routed again towards the Operator Station. SET UP M1

Seize M1

0 LOAD M1

First Job?

0

GO TO OPERATOR Station

Release EQUIPPER M1

Delay SETUP M1

T ru e

Delay LOAD M1

Release OPERATOR M1

Delay RUN M1

next action UNLOA D

GO TO OPERATOR Station

Fa ls e

Seize M 1

UNLOAD M1

Delay UNLOAD M1

Release OPERATOR M 1

0 First job?

0

INSPECTION M1

Delay INSPECTION M1

Release EQUIPPER M 1

T ru e

GO TO EQUIPPER Station

next action INS P E CTION

Fals e

Release M1

Release PQ1 M1

next area and next action

Figure 9: Actions Stations

B atch pieces

GO TO NEXT AREA Station

0

When a piece arrives to the unload Station (coming from an Operator Station, where it had seized an operator), it is delayed for unloading, and then it releases the operator. If the piece is the first of the job, it has to be inspected by an equipper: the opportune next action attribute (inspection) is stored, and the entity is routed towards the Equipper Station. Otherwise the piece has finished his task in the machine: it releases the machine and then releases the virtual resource

171

Heger and Hildebrandkt, 2009) or ‘Stretch’ (Bender, Muthukrishnan and Rajaraman, 2004) to measure the effect of scheduling on an individual job. The Flow Factor (or Stretch) of a job is the ratio of its Flow Time to its Processing Time: [C(i) - r(i)]/p(i). Flow factor is particularly suited in this case, where multiple jobs with different processing times are considered. The mean flow factor of all the jobs has been indicated by the expert personnel of the company as the measure through which compare different scheduling combinations of dispatching rules. In particular, each combination can be compared also with the scheduling decided by the company in the same period, that obtained a mean flow factor equal to PC = 27.33. 5.2. Results Figure 11 shows the simulation experiment results related to the first part (same equippers and operators rules).

Figure 10: A screenshot of the animation. 5.

THE SIMULATION EXPERIMENT Table 2: The simulation experiment results (first part).

The simulation experiment is conducted using real data provided by the company, and refers to orders arrived during 4 months, for a total number of different jobs equal to 24. Processing times performed by equippers and operators have been modeled as normal distributed. Standard deviations data were available for some of the considered jobs, those ones that had already been manufactured by the company and for which worksampling activities had already been performed. A coefficient of variation equal to 0.03 has been assumed for new jobs, accordingly to historical data related to similar jobs. The simulation experiment is divided into two parts. In the first one it is assumed that the same decision rule is assigned both to equippers and to operators. So different scenarios have been evaluated considering all the 5x7 = 35 combinations of the 5 decisions rules for machines (queue PQ1) and the 7 decision rules for operators and equippers (queues VQ1 and VQ2). The aim is to find the best rule for machines, and then to perform the second experiment maintaining the selected machines rule fixed, and exploring all the 7x7 combinations of rules for equippers and operators. This is performed in the second part of the experiment. Each scenario has been replicated 20 times.

equippers and operators rule (VQ1 and VQ2)

MEAN FLOW FACTOR

P-Value

FIFO

FIFO

30.14

1.000

FIFO

LIFO

29.54

1.000

FIFO

LMQL

32.42

1.000

FIFO

LPT

32.01

1.000

FIFO

RND

34.49

1.000

FIFO

SMQL

28.38

1.000

FIFO

SPT

28.92

1.000

LIFO

FIFO

32.39

1.000

LIFO

LIFO

32.84

1.000

LIFO

LMQL

38.04

1.000

LIFO

LPT

31.88

1.000

LIFO

RND

32.76

1.000

LIFO

SMQL

28.94

1.000

LIFO

SPT

29.55

1.000

LPT

FIFO

37.40

1.000

LPT

LIFO

35.63

1.000

LPT

LMQL

41.05

1.000

LPT

LPT

40.09

1.000

LPT

RND

37.65

1.000

LPT

SMQL

35.41

1.000 1.000

machine rule (PQ1)

5.1. Performance Measures Traditionally, the focus of performance in this type of scheduling problems has been on the Flow Time, which is defined as the amount of time that a given job spends in the system. If the i-th job arrives at time r(i), has Processing Time p(i) (that is known at the time of its arrival), and a Completion Time C(i), its flow time will be C(i) - r(i). However, Flow Time measures the time that a job is in the system regardless of the service it requests. Relying on the intuition that a job that requires a long service time must be prepared to wait longer than jobs that require small service times, practitioners and researchers have used the ‘Flow Factor’ (Scholtz-Reiter,

172

LPT

SPT

36.67

RND

FIFO

26.27

0.369

RND

LIFO

25.74

0.001

RND

LMQL

29.43

1.000

RND

LPT

28.42

1.000

RND

RND

26.22

0.278

RND

SMQL

28.21

1.000

RND

SPT

25.12

0.000

SPT

FIFO

24.39

0.000

SPT

LIFO

24.17

0.000

SPT

LMQL

26.16

0.170

SPT

LPT

25.94

0.018

SPT

RND

25.85

0.006

SPT

SMQL

22.77

0.000

SPT

SPT

22.00

0.000

processing time jobs at the end of the shift. In this way there is a higher probability that jobs starting at the end of the shift will last a reasonable amount of time, and that the successive set-up will not occur just a short time after the end of the equipper shift.

Main Effects Plot (data means) for MEAN FLOW FACTOR machine rule (PQ1)

equip. & oper. rules (VQ1 & 2)

Mean of MEAN FLOW FACTOR

37.5 35.0

Table 3. Main effects for Mean Flow Factor.

32.5 30.0

equippers rule (VQ1)

27.5

FIFO FIFO FIFO

LMQL

25.0 FIFO

LIFO

LPT

RND

SPT

FIFO

LIFO LMQL

LPT

RND SMQL

SPT

Figure 11. Main effects for Mean Flow Factor, first experiment. The table reports the average value P (over 20 replications) of the Mean Flow Factor for each scenario. The P-Value of the t-test in the last column indicates the smallest level of significance at which the null hypothesis (H0: P = PC) would be rejected in favor of the alternative hypothesis (H1: P < PC). The lowest value of mean flow factor is obtained when the Short Processing Time (SPT) rule is adopted for all the resources of the system (machines, operators and equippers). It is noteworthy that in the most part of scenarios obtaining a mean flow factor significantly lower than the one obtained by the company, the SPT rule is adopted for machines. Figure 11 reports the main effects plots for Mean Flow Factor, in which is also shown that SPT is the machine rule that performs better in combination with all the other equippers and operator rules. In the second experiment the machines rule was fixed (SPT), while all the combination of equippers and operators rules have been evaluated. Results are reported in Table 3. It is easy to see that the number of scenarios with P < PC is significantly higher. The best combination is obtained when also equippers follow the SPT rule, while operators follow the SMQL rule. Main effects plots (reported in Figure 12) confirm that these rules are the ones that on the average perform better when combined with all the other rules. Some considerations about the validity of these rules for equippers and operators can be drawn. Equippers are available only in the central part of the day (see Table 1), while machines and operators are available during all the 24 hours. An undesirable situation would be that a machine has to be set up when the equippers are not available (eg. during the night). This would cause in fact an idle time both for the machine and potentially for operators, that cannot proceed with load and unload actions on the pieces of the job. A good situation would be that the equipper set up the machine during its shift in such a way that machines and operators can continue processing the job during the night, without the need of a set-up. By giving precedence to jobs with the shortest processing time during its shift, the equipper tends to serve the longest

173

operators rule (VQ2)

MEAN FLOW FACTOR

P-Value

FIFO

24.39

0.000

LIFO

25.38

0.000

27.82

1.000

FIFO

LPT

25.19

0.000

FIFO

RND

24.47

0.000

FIFO

SMQL

23.47

0.000

FIFO

SPT

23.72

0.000

LIFO

FIFO

23.64

0.000

LIFO

LIFO

24.17

0.000

LIFO

LMQL

25.87

0.008

LIFO

LPT

25.85

0.006

LIFO

RND

0.085

LIFO

SMQL

26.08 21.78

LIFO

SPT

22.06

0.000

0.000

LMQL

FIFO

25.38

0.000

LMQL

LIFO

24.28

0.000

LMQL

LMQL

26.16

0.170

LMQL

LPT

24.59

0.000

LMQL

RND

24.91

0.000

LMQL

SMQL

23.28

0.000

LMQL

SPT

23.73

0.000

LPT

FIFO

25.79

0.003

LPT

LIFO

24.01

0.000

LPT

LMQL

27.34

1.000

LPT

LPT

25.94

0.018

LPT

RND

23.14

0.000

LPT

SMQL

24.79

0.000

LPT

SPT

25.31

0.000

RND

FIFO

25.58

0.000

RND

LIFO

23.67

0.000

RND

LMQL

28.44

1.000

RND

LPT

24.20

0.000

RND

RND

0.000

RND

SMQL

23.56 21.96

RND

SPT

22.77

0.000

SMQL

FIFO

24.64

0.000

SMQL

LIFO

25.33

0.000

SMQL

LMQL

24.65

0.000

0.000

SMQL

LPT

24.25

0.000

SMQL

RND

22.72

0.000

SMQL

SMQL

22.77

0.000

SMQL

SPT

22.59

0.000

SPT

FIFO

22.91

0.000

SPT

LIFO

23.79

0.000

SPT

LMQL

25.15

0.000

SPT

LPT

25.05

0.000

SPT

RND

0.000

SPT

SMQL

23.30 21.51

SPT

SPT

22.00

0.000

0.000

over performs the company’s scheduling. The approach allows to gain insights into priority rule performance, and to individuate a simple and implementable scheduling logic that provides a completely reactive scheduling.

Main Effects Plot (data means) for MEAN FLOW FACTOR equippers rule (VQ1)

Mean of MEAN FLOW FACTOR

27

operators rule (VQ2)

26

25

Aknolwedgments Authors thanks Dr. Cristiano Antinori for his supporting activity during the model implementation.

24

23 FIFO

LIFO LMQL

LPT

RND SMQL

SPT

FIFO

LIFO LMQL

LPT

RND SMQL

SPT

REFERENCES

Figure 12. Main effects for Mean Flow Factor, second experiment.

Bender, M.A., Muthukrishnan, S., Rajaraman, R., 2004. Approximation algorithms for average stretch scheduling. Journal of Scheduling, 7, 195-222. Blackstone, J.H., Philips, D.T., Hogg, G.L., 1982. A state-of-the-art survey of dispatching rules for manufacturing job shop operations. International Journal of Production Research, 20(1), 27-45. Cowling, P. I., Johansson, M. (2002). Using real-time information for effective dynamic scheduling. European Journal of Operational Research, 139(2), 230–244. Dauzère-Pérès, S., Roux, W., Lasserre, J.B., 1998. Multi-resource shop scheduling with resource flexibility. European Journal of Operational Research, 107(2), 289-305. ElMaraghy, H., Patel, V., Abdallah, I.B., 2000. Scheduling of manufacturing systems under dualresource constraints using genetic algorithms. Journal of Manufacturing Systems, 19(3), 186201. Geiger, C., Uzsoy, R., Aytu, H., 2006. Rapid modeling and discovery of priority dispatching rules: an autonomous learning approach. Journal of Scheduling, 9, 7-34. Geiger, D., Uzsoy, R., 2006. Learning effective dispatching rules for batch processor scheduling. International Journal of Production Research, 46(6), 1413-1454. Haupt, R., 1989. A survey of priority rule-based scheduling. OR Spectrum, 11, 3-16. Laarhoven, P.J.M.v., Aarts, E.H.L., Lenstra, J.K., 1992. Job shop scheduling by simulated annealing. Operations Research, 40(1), 113-125. Manikas, A., Chang, Y., 2009. Multi-criteria sequencedependent job shop scheduling using genetic algorithms. Computers & Industrial Engineering, 59(1), 179-185. Matie, Y., Xie, X., 2008. A genetic-search-guided greedy algorithm for multi-resource shop scheduling with resource flexibility. IIE Transactions, 40(12), 1228-1240. Ouelhadj, D., Petrovic, S., 2009. A survey of dynamic scheduling in manufacturing systems. Journal of Scheduling, 12, 417-431. Rajendran, C., Holthaus, O., 1999. A comparative study of dispatching rules in dynamic flowshops and

As far as the SMQL rule for the operators is concerned, it is noteworthy the importance to implement a rule that is not based on a job attribute, but on a machine attribute (PQ2 queue length). On the basis of the two preceding choices dictated by machine and equipper rules, the operator has only to choose among jobs that have already been assigned to a machine and set-up. In this case, serving the machine with the lowest number of pieces in the queue means to speed up the machine release, and to favor the entering of a new job. Possible improvements in the application of dispatching rules to this job shop scenario could be reached by allowing the selection of dispatching rules depending from time. For example, it would be possible to assign the SPT rule for the equipper in the first part of its shift and the LPT in the final part, in order to increase the probability that the successive set-up will not be needed just little time after the finish of the equipper shift. Analogously, dispatching rules for machines and operators could be differentiated in the case the equippers are present (the central shift of the day) with respect to when they are not.

6.

SUMMARY

In the paper a job-shop scheduling scenario is considered, derived from a case study of a manufacturing company that works for the aeronautical industry. A conceptual model of the shops has been built in order to implement a priority rules approach for the simultaneous scheduling of machines and two types of human resources: equippers and operators. The modelisation through virtual and physical queues allowed to define rules that are related both to jobs attributes and to machines attributes, and facilitated the implementation of a simulation model. Different combination of priority rules for machines, equippers and operators have been simulated and results have been compared on the basis of the mean flow factors of the considered jobs. Results have been compared to the mean flow factor obtained by the company for the same input data, and allow identifying the best rules combination that

174

jobshops. European Journal of Operational research, 116, 156-170. Scholz-Reiter, B., Heger, J., Hildebrandt, T., 2009. Analysis and comparison of dispatching rule-based scheduling in dual-resource constrainted shopfloor scenarios. Proceedings of the World Congress on Engineering and Computer Science, pp. 921-927. October 20-22, San Francisco (California, USA). Zhang, C., Li, P., Guan, Z., Rao, Y., 2007. A tabu search algorithm with a new neighborhood strucutre for the job shop schedulino problem. Computers & Operations Research, 34(11), 32293242. Zhou, H., Cheung, W., Leung, L.C., 2009. Minimizing wighted tardiness of job-shop scheduling using a hybrid genetic algorithm. European Journal of Operational Research, 194(3), 637-649.

AUTHORS BIOGRAPHY Lorenzo Tiacci. Laurea Degree in Mechanical Engineering, doctoral Degree in Industrial Engineering, he is Assistant Professor at the Department of Industrial Engineering of the University of Perugia. He is currently teaching courses of Facilities Planning & Design, Production Planning and Control, and Project Management at the University of Perugia. His research activity covers modeling and simulation of logistic and productive processes, plants design, production planning and inventory control, supply chain management, transportation problems. Stefano Saetta. Stefano SAETTA is Associate Professor at the Engineering Faculty of the University of Perugia. His research fields covers essentially the following subjects: modelling and simulation of logistic and productive processes, methods for the management of life cycle assessment, discrete event simulation, supporting decision methods, lean production. He was involved in several national and international research projects He is the organising committee and in the scientific committee of many international conferences.

175

EFFECT OF REJECT OPTION ON CLASSIFIER PERFORMANCE S. Dreiseitl(a) , M. Osl(b) (a)

Upper Austria University of Applied Sciences at Hagenberg, Austria (b) Division of Biomedical Informatics, UCSD, USA (a)

[email protected], (b) [email protected]

(Bradley 1997; Fawcett 2006) and goodness-of-fit tests (Hosmer et al. 1997; Pigeon and Heyse 1999), probability estimates can also be used for implementing a reject option. Such an option allows the system to refrain from making a decision if the predicted membership probability for both classes is around 50%, i.e., if the system cannot make a decision with a reasonable level of certainty. In addition to rejecting uncertain cases, it may also be desirable for a decision support system to make recommendations only for cases that are similar to the ones that were used for building it. This goal is more difficult to achieve, because it involves estimating how similar a new case is to a set of previously known cases. In the literature (see Section 2.), the first reject option (around probabilities of 50%) is known as ambiguity reject option, and the second (for outlier cases) as distance reject option. In this work, we investigate to which extend the ambiguity and distance reject options have an influence on the quality of a classifier’s performance, as measured by its discriminatory power and calibration. For the ambiguity reject option, we use a logistic regression model and do not classify cases for which the model output is close to 0.5. For the distance reject option, we additionally use a one-class SVM to estimate the regions in input space where most of the cases lie. The points outside these regions are then considered to be outliers and rejected from classification.

ABSTRACT Binary classifier systems that provide class membership probabilities as outputs may be augmented by a reject option to refuse classification for cases that either appear to be outliers, or for which the output probability is around 0.5. We investigated the effect of these two reject options (called “distance reject” and “ambiguity reject”, respectively) on the calibration and discriminatory power of logistic regression models. Outliers were found using one-class support vector machines. Discriminatory power was measured by the area under the ROC curve, and calibration by the Hosmer-Lemeshow goodness-offit test. Using an artificial data set and a real-world data set for diagnosing myocardial infarction, we found that ambiguity reject increased discriminatory power, while distance reject decreased it. We did not observe any influence of either reject option on the calibration of the logistic regression models. Keywords: classifier systems, reject option, performance evaluation 1. INTRODUCTION Decision support systems in biomedicine can augment a physician’s diagnostic capabilities by providing an automated second opinion. There are a number of approaches to building such systems, ranging from capturing an expert’s domain knowledge in explicit form to using machine learning methods that learn a model from given data without additional human intervention. Here, we consider only systems of the second kind, and further restrict our attention to models that distinguish between two classes (e.g., classifying cases as either healthy or diseased). Some of these systems, such as logistic regression or neural network models, provide explicit class membership probabilities, i.e., their output is a measure to which degree a case is healthy or diseased. Other machine learning models, such as support vector machines, must be explicitly augmented to provide probability outputs. The advantages of probability outputs are numerous: Besides facilitating accurate assessments of the system’s discriminatory power and calibration via ROC analysis

2. PREVIOUS WORK The idea of not classifying cases in regions of substantial class overlap, and thus class membership probabilities of around 50%, was proposed by Chow (1970), who was also the first to conduct an investigation into the benefits of the reject rule from a theoretical point of view. Dubuisson and Masson (1993) were the first to consider distance rejection, using nearest neighbor distances to decide whether a point is too far from the remainder of a data set. The particular case of a reject option for nearest neighbor classifiers had been studied earlier by Hellman (1970). Muzzolini et al. (1998) noted that rejection thresholds have to be adjusted to the covariance structure of mixture models to be unbiased, and proposed a method for performing this adjustment. The work of Landgrebe

176

et al. (2004, 2006) focused on the distance reject option when the classification task is ill-defined in the sense that one clearly defined target class is to be distinguished from another poorly defined class in the presence of an unknown third outlier class. Tax and Duin (2008) proposed a novel method for performing classification with a distance reject option by combining multiple one-class models, one for each of the individual classes. In the statistical literature, there is some theoretical research on the effects of reject options when rejection costs are different from misclassification costs. The framework of empirical risk minimization provides a theoretical background for the works of Herbei and Wegkamp (2006), Yuan and Wegkamp (2010), and Bartlett and Wegkamp (2008), who derived an SVM classifier with a reject option.

dimensions. A recent addition to the machine learning arsenal allows us to address this problem without regard to data dimensionality (Schölkopf et al. 2001; Schölkopf and Smola 2002) . One-class support vector machines extend standard support vector machine (SVM) methodology to the case of estimating a given fraction (1 − ν) of the support of a data set; the remaining fraction ν are considered outliers. As with other support vector methods, the one-class SVM algorithm projects the data into a different feature space F using a nonlinear mapping Φ : Rn → F . Without a second class, the aim is then to separate the projected data from the origin by as wide a margin ρ as possible. The use of kernel functions k to replace projections and dot product operations is similar to other SVM algorithms. We used a Gaussian kernel   k(xi , xj ) = exp − γ xi − xj 2

3. METHODS In this section, we first describe the algorithms we used for building machine-learning models, and then the methods for evaluating the performance of these algorithms.

with inverse variance parameter γ. Values of γ were chosen in such a way that the proportion of outliers was close to ν. One-class SVMs estimate the data distribution by solving the constrained optimization problem

3.1. Machine learning algorithms We consider dichotomous classification problems as specified by an n-element data set of m-dimensional input vectors x1 , . . . , xn and corresponding class labels y1 , . . . , yn ∈ {−1, 1}. For logistic regression, we assume there is an additional constant 1 at the first position of the xi in order to simplify the notation below; these augmented data points are thus (m + 1)-dimensional. In a logistic regression model, the optimal values for the (m + 1)-dimensional parameter vector β are determined by minimizing a negative log-likelihood function. We additionally consider L2 -regularization of logistic regression models by calculating the maximum likelihood estimate βML as βML = arg min β

n 

w · Φ(xi ) ≥ ρ − ξi , ξi ≥ 0 ∀i = 1, . . . , n

subject to

where w is the parametrization of the separating hyperplane in F , and the ξi are slack variables. The dual problem is

  T log 1 + e−yi β ·xi + λβ T β .

min α i ∈ Rn

n 1  αi αj k(xi , xj ) 2 i,j=1

subject to

0 ≤ αi ≤ n 

i=1

The regularization parameter λ is usually chosen by cross-validation. The model predictions for new cases x are then given as class-membership probabilities P (y = +1 | x, βML ) =

n 1  1 ξi − ρ w2 + 2 nν i=1

min w∈F, ξi ∈Rm, ρ∈R

1 nν

∀i = 1, . . . , n

αi = 1 .

i=1

Support vectors are those data points xi for which the 1 . Outliers are corresponding αi satisfies 0 < αi < mν those points for which the decision function

1 . T 1 + e−βML·x

f (x) =

n 

αi k(xi , x) − ρ

i=1

The model outputs are thus logistic transformations of T βML · x, i.e., values proportional to the distance of x from the hyperplane parameterized by βML . Ambiguity rejection can therefore be seen to refuse classification for those cases that are within a certain distance from the separating hyperplane. For distance rejection, we need to estimate the regions of input space in which data are more dense than in others. Standard parametric and non-parametric density estimation algorithms are susceptible to the curse of dimensionality, and therefore not easily applicable in high

is negative. 3.2. Evaluation metrics We used the area under the ROC curve (AUC) as measure of a classifier’s discriminatory power, and computed an estimator θˆ of the AUC via its equivalence to a MannWhitney U-statistic as θˆ =

177

n2  n1    1   1  1 p−i < p+j + 1 p−i = p+j . n1 · n2 i=1 j=1 2

Figure 2: Sample of the artificial data set showing two normally distributed classes with the logistic regression discriminatory line. The points for which no classification is made based on distance rejection are located at the outer edge of the data set, and shown in light grey.

Figure 1: Sample of the artificial data set showing two normally distributed classes with the logistic regression discriminatory line. The points for which no classification is made based on ambiguity rejection are located close to the discriminatory line, and shown in light grey.

two classes. The parameters were chosen to achieve an AUC of about 0.9. A sample of the data with the separating line as determined by logistic regression, along with points in the ambiguity reject region, is shown in Figure 1. The same data, now with points rejected based on distance highlighted, is presented in Figure 2. The myocardial infarction data set consists of information collected from 1253 patients presenting at the emergency department of the Edinburgh Royal Infirmary in Scotland with symptoms of acute myocardial infarction (AMI). A total of 39 features were recorded, comprising patient data (smoker, diabetes, . . . ), clinical information (location of pain, sensation of pain, hypertension, . . . ), and results of an ECG test (LBB, abnormal T wave, . . . ). To increase diagnostic difficulty, we removed data about ECG measurements, retaining a total of 33 features. The gold standard diagnosis was made by expert physicians based on a combination of blood serum tests with clinical and ECG data. Of the 1253 patients, 274 were diagnosed with AMI, and 979 patients were either declared healthy or to be suffering from other ailments.

+ Here, p− i and pj are the classifier outputs for cases from classes −1 and +1, respectively, and 1 is the Boolean indicator function. The calibration of a classifier is usually assessed with the Hosmer-Lemeshow C-test (Hosmer and Lemeshow 1980). Although often critizised for a number of drawbacks (Bertolini et al. 2000), it is nevertheless the de-facto standard for determining the goodness-of-fit of a model. As a Pearson chi-squared test, it computes the test statistic G  (Oi − Ei )2 , C= i Ei (1 − E ni ) i=1

as the sum of standardized squared differences between the number of observed cases Oi and expected cases Ei for a grouping of classifier outputs into G groups, each with ni cases. By definition, G = 10 and the data is grouped is by sorted classifier outputs. Hosmer and Lemeshow (1980) observed that C has an approximate chi-squared distribution with G − 2 degrees of freedom. 4. EXPERIMENTS Our experiments on the effect of the reject option on classifier performance utilized two data sets, one simple artificial toy problem, and one real-world data set from the domain of predicting acute myocardial infarction.

4.2. Results Our experiments were carried out using MATLAB (MathWorks, Natick, MA), with our own implementation of logistic regression models and the libsvm implementation of one-class SVMs (Chang and Lin 2001). For both the artificial as well as the myocardial infarction data set, we trained the logistic regression models using 60% of the data, with the remaining 40% reserved for testing. All data features were normalized to zero mean and unit variance. The experiments were performed 50 times, each

4.1. Data sets For the artificial data set, we generated 500 data points each from two multivariate normal distributions with diagonal covariance matrices, each representing one of the 178

Table 1: Discrimination and classification of logistic regression models on the artificial data set. The results for ambiguity and distance reject options are listed by varying fractions of rejected cases. HL denotes the HosmerLemeshow test statistic; the critical value for α = 0.05 is 15.51.

logistic regression (baseline) ambiguity reject τ = 0.1 τ = 0.2 τ = 0.3 τ = 0.4 distance reject ν = 0.05 ν = 0.1 ν = 0.2

AUC mean std

HL mean std

0.880

15.07

0.897 0.913 0.925 0.936

0.871 0.862 0.851

0.01

0.01 0.01 0.01 0.01

0.01 0.02 0.02

15.22 13.93 14.73 14.82

13.93 13.10 9.65

Table 2: Discrimination and classification of logistic regression models on the myocardial infarction data set. The results for ambiguity and distance reject options are listed by varying fractions of rejected cases. HL denotes the Hosmer-Lemeshow test statistic; the critical value for α = 0.05 is 15.51. AUC mean std

HL mean std

0.842

0.12

14.51

9.03

8.58 9.35 7.82 9.61

ambiguity reject τ = 0.1 τ = 0.2 τ = 0.3 τ = 0.4

0.851 0.860 0.874 0.886

0.12 0.13 0.13 0.13

14.10 14.29 13.14 12.19

11.47 10.98 10.10 8.76

6.49 5.99 3.95

distance reject ν = 0.05 ν = 0.1 ν = 0.2

0.841 0.838 0.834

0.12 0.12 0.12

19.38 22.61 34.81

20.39 19.29 21.23

7.16

logistic regression (baseline)

with different random allocations of data to the training and test sets. All results are reported as averages and standard deviations on the test set over these 50 runs. The parameters of our experiments were ν, the fraction of outliers for the distance reject option, and τ , the fraction of cases for the ambiguity reject option. The ambiguity reject cases were the proportion τ of cases for which the model output (class membership probability) was closest to 0.5. A kernel parameter of γ = 0.001 gave a number of outliers within 10% of the desired value, as specified by ν. The number of support vectors was slightly larger. This is in concordance with theory, which states that ν is an upper bound on the fraction of outliers, but a lower bound on the fraction of support vectors (Schölkopf et al. 2001). The results of our experiments on the artificial data set are summarized in Table 1. One can observe that ambiguity rejection had a positive effect on AUC. This is to be expected, because ambiguity rejection removes those cases for which most misclassification errors occur. Furthermore, it is also reasonable that the increase in AUC is not as pronounced for larger values of τ , because fewer and fewer ambiguous cases get removed. On the other hand, there seemed to be no effect of τ on the value of the Hosmer-Lemeshow (HL) test statistic. As is known from the literature (Bertolini et al. 2000), this test dependents strongly on the particular grouping of data points, and showed high volatility in our experiments as well (as indicated by the large standard deviations). As for distance rejection, the AUC value decreased with increasing numbers of rejected cases, while the HL test statistic showed better model fit. A possible explanation for the first phenomenon is the fact that the rejected points were almost all correctly classified by the model

(because they were on the correct side of the discrimination line). Removing them therefore had a detrimental effect on the AUC. The second phenomenon may just be a fluke observation, as evidenced by the high standard deviations and the fact that it was not present in the realworld data. Table 2 provides the same information for the myocardial infarction data set. Again, we find that ambiguity rejection increased AUC without a clear effect on the HL test statistic (which exhibits even higher standard deviations than on the artificial data set). And again, we observed that distance rejection had a negative effect on AUC. The effect on the HL test statistic was in the opposite direction of the effect it had on the artificial data set. Distance rejection is also not beneficial if it is performed before training, as suggested by Landgrebe et al. (2006). In this case, one rejects data from the training set, and not from the test set. The reasoning for this is that the model may be a better representation of the underlying data generator when outliers are removed prior to model building. Table 3 shows that this is not the case: There was no difference in AUC for the artificial data set, and an even larger negative effect for the real-world data set. There were no discernible effects on the HL test statistics. 5. CONCLUSION We investigated the effect of the ambiguity and distance reject options on performance of a logistic regression model on an artificial and a real-world data set. We observed ambiguity rejection to increase AUC, and distance rejection to decrease it. Both reject options did not have an effect on classifier calibration. 179

Hosmer, D. and Lemeshow, S. (1980). A goodness-of-fit test for the multiple logistic regression model. Communications in Statistics, A10:1043–1069.

Table 3: Discrimination and classification of logistic regression models on the artificial and on the myocardial infarction data set, when a fraction ν of cases are removed by distance rejection from the training set prior to model building. HL denotes the Hosmer-Lemeshow test statistic; the critical value for α = 0.05 is 15.51.

artificial data logistic regression (baseline) ν = 0.1 ν = 0.2 myocard. inf. data logistic regression (baseline) ν = 0.1 ν = 0.2

AUC mean std

HL mean std

0.880

0.01

15.07

7.16

0.880 0.880

0.01 0.01

14.48 15.61

6.92 7.73

0.842

0.12

14.51

9.03

0.828 0.818

0.12 0.12

15.20 14.28

11.75 10.40

Hosmer, D. W., Hosmer, T., le Cessie, S., and Lemeshow, S. (1997). A comparison of goodness–of–fit tests for the logistic regression model. Statistics in Medicine, 16:965– 980. Landgrebe, T., Tax, D., Paclík, P., and Duin, R. (2006). The interaction between classification and reject performance for distance-based reject-option classifiers. Pattern Recognition Letters, 27(8):908–917. Landgrebe, T., Tax, D., Paclík, P., Duin, R., and Andrew, C. (2004). A combining strategy for ill-defined problems. In Proceedings of the 15th Annual Symposium of the Pattern Recognition Association of South Africa, pages 57–62. Muzzolini, R., Yang, Y.-H., and Pierson, R. (1998). Classifier design with incomplete knowledge. Pattern Recognition, 31(4):345–369. Pigeon, J. G. and Heyse, J. F. (1999). An improved goodness of fit statistic for probability prediction models. Biometrical Journal, 41(1):71–82.

ACKNOWLEDGEMENTS This work was funded in part by the Austrian Genome Program (GEN-AU), project Bioinformatics Integration Network (BIN) and the National Library of Medicine (R01LM009520).

Schölkopf, B., Platt, J. C., Shawe-Taylor, J., Smola, A., and Williamson, R. C. (2001). Estimating the support of a high-dimensional distribution. Neural Computation, 13:1443–1471. Schölkopf, B. and Smola, A. (2002). Learning with Kernels. MIT Press, Cambridge, MA.

REFERENCES Bartlett, P. and Wegkamp, M. (2008). Classification with a reject option using a hinge loss. Journal of Machine Learning Research, 9(Aug):1823–1840.

Tax, D. and Duin, R. (2008). Growing a multi-class classifier with a reject option. Pattern Recognition Letters, 29(10):1565–1570.

Bertolini, G., D’Amico, R., Nardi, D., Tinazzi, A., and Apolone, G. (2000). One model, several results: the paradox of the Hosmer-Lemeshow goodness-of-fit test for the logistic regression model. Journal of Epidemiology and Biostatistics, 5(4):251–253.

Yuan, M. and Wegkamp, M. (2010). Classification methods with reject option based on convex risk minimization. Journal of Machine Learning Research, 11(Jan):111–130.

AUTHOR BIOGRAPHIES

Bradley, A. (1997). The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recognition, 30:1145–1159.

STEPHAN DREISEITL received his MSc and PhD degrees from the University of Linz, Austria, in 1993 and 1997, respectively. He worked as a visiting researcher at the Decision Systems Group/Harvard Medical School before accepting a post as professor at the Upper Austria University of Applied Sciences in Hagenberg, Austria, in 2000. He is also an adjunct professor at the University of Health Sciences, Medical Informatics and Technology in Hall, Austria. His research interests lie in the development of machine learning models and their application as decision support tools in biomedicine. MELANIE OSL received her Ph.D. degree from the University of Medical Informatics and Technology (UMIT) in Hall, Austria in 2007. She is currently a postdoc fellow at the Division of Biomedical Informatics at the University of California, San Diego. Her research interests include knowledge discovery and data mining in biomedicine, clinical bioinformatics and machine learning.

Chang, C.-C. and Lin, C.-J. (2001). LIBSMV: a library for support vector machines. Software available at http://www.csie.ntu.edu.tw/∼cjlin/libsvm. Chow, C. (1970). On optimum error and reject tradeoff. IEEE Transactions on Information Theory, IT-16(1):41–46. Dubuisson, B. and Masson, M. (1993). A statistical decision rule with incomplete knowledge about classes. Pattern Recognition, 26(1):155–165. Fawcett, T. (2006). An introduction to ROC analysis. Pattern Recognition Letters, 27(8):861–874. Hellman, M. (1970). The nearest neighbor classification rule with a reject option. IEEE Transactions on Systems Science and Cybernetics, SSC-6(3):179–185. Herbei, R. and Wegkamp, M. (2006). Classification with reject option. Canadian Journal of Statistics, 4(4):709–721.

180

NEW DISCRETE TOPOLOGY OPTIMIZATION METHOD FOR INDUSTRIAL TASKS Sierk Fiebig(a) (a)

Volkswagen Braunschweig, Braunschweig, Germany (a)

[email protected]

assigned to intermediate density values [Edw07]. SIMP is combined to a gradient algorithm, e.g. the method of moving asymptotes [Svan87]. Since 1992 another important approach has been developed. The evolutionary structural optimization (ESO) is focused to remove unnecessary material from too conservative designed parts [Que00]. To ESO, it is only possible to remove material and uses a discrete element modeling in comparison to SIMP [HuaXie10]. To enable the opposite, Querin introduces the additive evolutionary structural optimization method, called AESO [Que00b]. AESO adds material to the highest stressed points in order to become an optimal structure. The combination of ESO and AESO is the bidirectional evolutionary structural optimization [BESO] method [Que00] [HuaXie10]. The main idea behind ESO, AESO and BESO is to remove under stressed elements and to add material to higher stressed areas. To designate these elements, two reference levels are defined. During the optimization these levels are adapted to the optimization progress. All elements under a reference level are removed and all elements above a second level are added. BESO uses here - depending on the individual approach - direct, gradient or interpolated information about material properties to change the structure [HuaXie10]. For industrial usage the SIMP method in combination with gradient algorithm has a large distribution. One main reason for the success of the approach is the integration of manufacturing restrictions. Without these restrictions, it isn’t possible, in most cases to get a feasible design for real life problems. At the moment no proposals for the integration of manufacturing restriction to BESO are offered.

ABSTRACT Nowadays the development of mechanical components is driven by ambitious targets. Engineers have to fulfill technical requirements under the restrictions of reducing costs and weights simultaneously. Therefore in the last years optimization methods have been integrated in the development process of industrial companies. Today, especially topology optimization methods, have gained in importance and are standard for developing casting parts. Stress or strain-energy information is used for sensitivities in all topology optimization methods. The method SIMP, today’s standard in industry, uses continuous material modeling and gradient algorithms. ESO/BESO use discrete modeling and specific algorithms depending on the individual approaches. The new Topology Optimization method uses a discrete modeling, too. The number of modified elements is controlled by the progress of the constraint. For solving tasks in the industrial development process, a topology optimization method must enable an easy and fast usage and must support manufacturing restrictions. Keywords: topology optimization, mechanical components, discrete modeling of material 1. INTRODUCTION Today several approaches exist for topology optimization. The starting point of FEA based topology optimization was at the end of the eighties [Roz01]. Bendsøe introduced first his homogenization method [Ben89]. Parallel to the homogenization method, Bendsøe presented the SIMP approach (Solid Isotropic Microstructure with Penalization) [BenSig03]. This method has become popular, because other researchers use it [Roz92]. Today the SIMP approach is one of the standard methods for topology optimization. For example, the commercial tool Tosca® from FE-Design is based on SIMP. SIMP uses continues design variables. Here the density is used as design variable. The coupled Young-Modulus transfers the modifications of the optimization to the structure results. At the end of each topology optimization, a clear discrete distribution for interpreting the results is needed. Due to this, the SIMP approach penalizes intermediate density values using a penalization factor as a power. In this way, low stiffness values are

2.

THE NEW APPROACH FOR TOPOLOGY OPTIMIZATION The motivation for the new approach is based on three reasons. The main focus is the usage of the method in industry. The overall interest of industry is to recognize parts with lower weight and cost compared to the older reference structure. In contrast to optimization from a mathematical or theoretical view, the task of optimization isn’t to find the absolute optimum. In the opinion of engineering and praxis, optimization means the improvement of the result.

181

iteration. The reason is that the new Topology Optimization method adds only at the highest stressed elements( often called hotspots) material. Starting optimization with infeasible solutions forces the optimization method to add material first to the structure. The BESO method finds the same solution in this case as an optimization run starting from full design space [Que00b]. The new approach offers in this case different solutions, because the process is controlled by the constraint limit. For industrial purposes this behavior is more powerful in later development phases, e.g. when load conditions must be changed to new requirements, the engineer wants to find a new feasible and as light as possible design but with a minimum of changes in the part.

This mean, that better optimization results are the first motivation for the new method. To achieve this and to improve the universal usage, linear and nonlinear FEA analysis should be possible with the new Topology Optimization method. Nonlinear effects are for example plastic behavior of material, nonlinear behavior in bushing and contact problems. Finally the last point, manufacturing requirements should be fulfilled. 2.1. Basic functionalities similar to ESO/BESO Using stress or strain-energy information for sensitivities are the basic ideas in all topology optimization methods, see [BenSig03] [HuaXie10] [Mat94]. Depending on this main idea, the new approach uses the stress-values for reducing or adding discrete material in the design space. Another important similarity is the discrete modeling of material. Following ESO and BESO, the lowest stressed elements are removed from the structure. This is a simple but effective method. This mechanism has also an analogy in nature, during the development and growth of plants [Mat94] and is rooted from experience for solving problems in engineering also. Due to the fact of discrete material, new elements can only be added to the borders of an existing structure. Without interpolation information the new material is placed to the areas with the highest stress levels, see the AESO method of Querin [Que00]. Depending on the discrete modeling, both methods are possible to handle linear and nonlinear effects in the FEA analysis. The only difference to a regular FEA simulation lies in the surface of the FEA model. Up to now the models from topology optimizations with discrete modeling is not as smooth as a model from a regular simulation.

2.3. Main process of the new Topology Optimization method The flow chart in figure 1 illustrates the main steps of the new Topology Optimization method. The step size controller calculates first a basic rate. Depending on this basic rate, the number of removing and adding elements is defined. After the controller the necessary elements are inserted. In this way, hotspot areas are corrected. After this correction process, the lowest stress elements according to the reduction rate are removed. After adding and removing elements, it is important to check if the structure is connected. All force transmission points must be connected to the supports. If this check fails, the controller modifies the correction and reduction rate in order to produce a feasible structure. In the heuristic steps, non connecting elements are removed from the structure. 2SWLPL]HU QR 6WHSVL]HFRQWUROOHU %DVLFUDWH

2.2. Main differences to ESO/BESO The new Topology Optimization has beside some similarities clear differences to the ESO/BESO methods. The main idea of ESO/BESO is a full stressed design, means all elements receive the same stress level. For this method the compliance-volume product can be assumed as an objective function [Edw07]. Opposite to this, the new Topology Optimization method uses only the volume as target or object function. For the optimization the new method needs constraints. Remembering the motivation, the new Topology Optimization allows several constraints, e.g. displacement or reaction force. Also the combination of all constraints is possible. Normally a min-max formulation is used. But also other mathematical operators are possible to use, e.g. weighed or distance formulations. In the original approach of BESO, no constraints are used. Only the stress levels are important. With the main focus to a normalized stress level, BESO adds material by comparing each element stress level to a reference level. Comparing this to the new Topology Optimization method more elements are added in each

&RUUHFWLRQ UDWH

$GGLQJHOHPHQWV

5HPRYLQJ HOHPHQWV

6WUXFWXUH FRQQHFWHG

\HV

+HXULVWLFV HJGHOHWLQJ XQFRQQHFWHGHOHPHQWV

5HGXFWLRQ UDWH

,QWHUIDFH

5HDGLQ UHVXOWV

5HVXOW SRVWSURFHVVLQJ

:ULWH2XW )($PRGHO

)($DQDO\VLV VLPXODWLRQ

Figure 1: Flow chart of New Topology Optimization Method The necessary interfaces to the FEA solver are integrated in the optimizer. After finishing all changes and checks, the optimizer writes the element input decks. After the FEA analysis, the result postprocessing evaluates all target functions and constraints. The read in process transfers this information back to the controller.

182

2.4. Integration into the industrial development process Several steps are needed for the procedure of a topology optimization. Normally a topology optimization is based on FEA analysis. Due to this, the topology optimization must be coupled with a FEA solver. One basic idea of this new approach is the integration in the standard development process, especially simulation process. Through this, a external FEA solver should be used. This demands interfaces to read and write the special formatted input decks of the solver. The optimizer supports two FEA solvers: Abaqus from Simulia® and Nastran.ND® from MSC®. Other FEA solvers can be integrated. Only the necessary interfaces have to be programmed in C++. To minimize the complexity of the development, the new Topology Optimization doesn’t manage the process of the topology optimization. The workflow is controlled through an external program, such as Optimus® from Noesis®.

Figure 3: Interface between optimizer and FEA solver 2.4.3. Optimization Workflow For the workflow Optimus® is used. The optimizer is integrated as User Algorithm. The optimizer writes the FEA input decks on his own. To optimus a reference to this file is transferred. During the process, optimus transfers files, starts the FEA solver and the postprecessing scripts to evaluate the stress values and the constraints. The process is shown in figure 4.

2.4.1. Preprocessing The preprocessing can be divided into two parts. First the normal FEA preprocessing has to be done. In figure 2, the two steps: meshing the part and define loads and structure supports are illustrated.

Figure 2: Preprocessing After this, for the optimization preprocessing, different files in ASCII® format are used. The design areas, the constraint and the optimization parameter are chosen. All is flexible and can be adapted to the specific problem.

Figure 4: Preprocessing 2.4.4.

Interface between the FEA solver and the optimizer The figure 5 illustrates the process between FEA postprocessing and the interface of the optimizer.

2.4.2.

Interface between the optimizer and the FEA solver The optimizer works internal with a data grid, see figure 3. The information form the internal data grid, called “matrix”, can be transferred to the FEA model. On the initial run of the optimizer the elements are mapped to the matrix. This mapping is fixed over the whole optimization. The optimizer changes the status of the matrix. Status 0 means no material, Status 1 means material. After finishing the optimization steps, this information is mapped to the FEA model. Only the elements with status 1 are written to the FEA data file. No other elements are available for the FEA solver. Therefore, the results of the FEA analysis, apart from the aliasing effects at the border, have the same result quality as a normal analysis. Figure 5: Postprocessing and interface to the optimizer

183

The result of the postprocessing is a list of all elements. Each line of this list represents one element. The line begins with the element id. Second entry is the stress value of the element. The first step of the interface maps the stress value of each element to the internal element list. After this, the stress values of the elements are mapped to the internal data matrix.

Figure 7: Forging

2.5. Integration of manufacturing restriction For a feasible industrial casting part design, numerous manufacturing restrictions have to be fulfilled. Besides minimum and maximum material strength, normally a forming direction has to be taken into account. Special production processes - especially forging - need closed structures, at best, without any holes. Additionally it is sometimes necessary to design symmetric parts, maybe for using the same part on the left and right side of a car. As well as a minimum strength restriction, casting directions, forging and symmetry restrictions are implemented.

2.5.3. Symmetry At the moment, only a plane symmetry is implemented. But other symmetries, like point symmetry, are easy to add. The mechanism for the plane symmetry can be directly transferred to them. For a plane symmetry, all elements are divided into two groups. The first group allows modifications. After adding and removing elements in this group, the changes are mapped to the second group, illustrated in figure 8.

2.5.1. Casting direction Due to a casting direction, no material inside the structure can be deleted. In this way, no undercuts exists. The figure 4 illustrates the differences between a part with active and non active casting restrictions. Without casting restriction, all elements can be removed from the structure. The optimization starts with the lowest stressed elements. With casting restriction only the visible elements can be removed from the structure. After removing one element, the next element becomes visible. In each step, the current lowest element is deleted. Figure 8: Plane symmetry

The red line in figure 6 shows the maximum element number in one row and which is possible to remove each iteration. Due to this, the algorithm can repair too large cuts in the next iteration.

2.5.4. Minimum Strength To avoid too small structures, which aren’t possible to manufacture, normally filters are used. To this reason, discrete element modeling and using the half length of the minimum material strength for the elements length, the structure has a natural material strength. If the strength of the structure is smaller, the risk of collapsing in the FEA analysis is too high. This is demonstrated for example in figure 9.

Figure 6: Casting direction 2.5.2. Forging This restriction avoids parts with holes in the structure. To implement the mechanism, the last elements in one row are blocked for adding them to the visible group. Without getting visible, theses elements can’t be removed from the structure, see figure 7.

Figure 9: Minimum material strength

184

2.6. Example: Cantilever example with plastic material One classical problem for testing topology optimization is the cantilever problem. In this case, it should directly demonstrate, how the new Topology Optimization method works using a nonlinear FEA analysis with plastic material. The material characteristic in this example has the specification of steel. On the left, two fixed supporting elements form the boundary as indicated in figure 10. In the middle of the plate on the right an enforced displacement of 20 mm at an angle of 90° to the main describes the load. In FEA simulation nonlinear geometry is activated and a real flow curve is used. As constraint function the reaction force at the node where the enforced displacement is applied, is used. The part should be optimized to a level of 10 kN. The target function is the minimization of weight, measured in elements.

Figure 11: Optimization process of a cantilever problem with plastic material behavior described in Figure 10

Figure 10: Cantilever problem with plastic material behavior

Figure 12: Detailed optimization process of a cantilever problem with plastic material behavior from Figure 10

The optimization starts with a full design space with 125000 elements and an inertial basic reduction rate of 0.1. The general optimization process can be divided into three phases. The first phase is described by large reductions of elements up to the moment, where the constraint function rises strongly. In this second phase the constraint rises to the point where the constraint limit is reached. At this point two following iterations violate the limit. The optimization makes a cutback caused by the control mechanism and adds nearly 20% of elements to the structure. As a result the constraint offers the possibility to reduce the elements once again. Up to iteration 36 the optimization run reaches the point, where the cutback was made. Now the optimization control function is under the constraint limit. So the optimization progress enters the third phase. Characterized by slow step sizes, the optimization run offers improvements in detail. Through the oscillation around the constrain limit the last unnecessary elements are removed. In figure 11 and 12 the optimization process is described through the number of elements and a normalized constraint. The value is the reaction force through the constraint limit of 10 kN. At ~25000 elements the cutback level is reached. After the second phase the optimization minimizes the weight to 20.35% of the starting value. In iteration 100 the optimization ends with a final value of 19.75%.

The changes of the structure and the stress plots during the optimizer are illustrated in figure 13.

Figure 13: changes in optimization process of a cantilever problem with plastic material behavior from Figure 10

185

compared with conventional gradient based topology optimizations, especially in cases with nonlinear effects, e.g. plastic material behavior, in the FEA simulation. Finding a satisfying solution in topology optimization reduces the necessary development time in a development department. The first designs in CAD based on the optimization runs indicate very competitive weight and fulfill immediately the technical requirements and manufacturing restrictions. So development loops and development costs can be saved. New target and constraint functions will increase the usage and more problems can be solved in less time.

Figure 14 demonstrates the result quality of the new approach. Using nonlinear FEA analysis during the optimization run, it is possible to dimension all areas in the structure correctly. Due to this, the result shows a very good utilization of material in nearly all areas. Another positive aspect is the simplicity of the structure. No ramifications are proposed. This structure is easier to be constructed and manufactured.

REFERENCES G.I.N. Rozvany: Aims, scope, methods, history and unified terminology of computer-aided topology optimization in structural mechanics, Struct Multidisc Optim 21, 90–108, Springer-Verlag, Berlin, 2001 M.P Bendsøe: Optimal shape design as a material distribution problem, Struct. Optim. 1, 193–202, 1989 M.P. Bendsoe and O. Sigmund: Topology Optimization: Theory, Methods and Applications, Springer-Verlag, Berlin, 2003. G.I.N. Rozvany, M. Zhou, T. Birker: Generalized shape optimization without homogenization. Struct. Optim. 4, 250–254, 1992 C. S. Edwards, H. A. Kim, C. J. Budd: An evaluative study on ESO and SIMP for optimising a cantilever tie–beam, Struct Multidisc Optim (2007) 34:403–414, Springer-Verlag, Berlin, 2007 K. Svanberg: The method of moving asymptotes-a new method for structural optimization, Int J Numer Methods Eng 24:359–373, 1987 Q.M. Querin, G.P. Steven, Y.M. Xie: Evolutionary structural optimisation using an additive algorithm, Finite Elements in Analysis and Design 34 (2000 291-308, ELSEVIER, 2000 X Huang, Y.M.Xie: Evolutionary Topology Optimization of Continuum Structures, Wiley, Chichester, 2010 Q.M. Querin, V. Young, G.P. Steven, Y.M. Xie. : Computational efficiency and validation of bi-directional evolutionary structural optimization, Comput. Methods Appl. Mech. Engrg. 189 (2000) 559-573, ELSEVIER, 2000 C. Mattheck: Design in der Natur, Rombach Verlag, Freiburg im Breisgau, 1997

Figure 14: Nonlinear FEA analysis of optimization results from the example in Figure 10 The figure 15 shows the reaction force of the final iteration. In the figure is the progress of the reaction over the displacement illustrated. The final value at 20mm displacement is 10008,8 N. This demonstrates, that the optimizer has the ability to deliver a result exactly to the necessary constraint limit.

Figure 15: Reaction force of final iteration from the example in Figure 10 3. 6. CONCLUSION In this paper an approach for a new Topology Optimization method is proposed. The method is developed based on requirements from the automotive industry. The main focus is the combination of finding a minimum weight and the best material distribution in one optimization run. To fulfill this task a discrete approach to include all nonlinear effects, e.g. plastic material behavior was chosen. The example demonstrates the quality of the new optimization method. For the cantilever problem the new approach shows an advantage of 30% compared with a conventional industrial gradient based topology optimization methods. The new developed method shows the usability for real life development problems. The quality of the results is significantly increased

AUTHORS BIOGRAPHY Sierk Fiebig studied mechanical engineering at the TU Braunschweig. After his studies, he started research work for his Ph.D.-thesis at Volkswagen in optimization methods for the development of mechanical optimization methods. Since joining the company he has worked as a simulation engineer in the chassis development at Volkswagen Braunschweig.

186

3G MOBILE NETWORK PLANNING BASED ON A TRAFFIC SIMULATION MODEL AND A COST-BENEFIT MODEL TO SERVICE LOS CABOS INTERNATIONAL AIRPORT Aida Huerta Barrientos (a), Mayra Elizondo Cortés (b). (a) (b) (a)

National Autonomous University of Mexico (UNAM) National Autonomous University of Mexico (UNAM)

[email protected] ,(b)[email protected].

specific service demands from network users and operators. The first generation cellular mobile (1G) technology enabled the human communication via voice. While the second generation cellular mobile (2G) technology enabled the human communication via voice and text messages, and the third generation (3G) technology enabled the human communication via voice, text messages, data and video. The same forces that fed the development of new services and the entrance of new players also saw margins grow slimmer for most services as well as significant customer churn as competitors offered alternative choices. So for network operators has been important to give mobile cellular service indoor and outdoor along cities and rural communities. To give the service, the network operators need, by first time, to plan the network. The network planning is a process on which operators take in consideration aspects such as demand of services and applications from users, service area based on the potential user locations, service rates, quality of service, environmental morphology and return of investment, for example. By one hand, in today´s extremely challenging business environment, many telecommunications operators and carriers are measuring their success by the size and growth of their profit margins. As a result, operators are under intense pressure to reduce or eliminate the major threats to these slim margins including revenue leakages and frauds, churn, inefficient network usage, and least-cost routing plans. These competitive and market pressures are also making the telecommunications industry reassess its business model and redefining the path that will return it to competitiveness and profitability (Pareek 2007). By another hand, under competitive conditions, the customer becomes the central focus of the carrier´s activities. Customer requirements not only determine service offerings, but also shape the network. In this sense, we propose a network design to service Los Cabos International Airport, which is located in Mexico country, based on user´s requirements and using 3G mobile cellular technology specifications. This proposal is based on a traffic simulation model and a cost-benefit model and the main objective is to get the cell radius that maximizes the net profit percentage of the network.

ABSTRACT We propose an application approach to planning a thirdgeneration mobile network based on a traffic simulation model and a cost-benefit model to service the interior of an airport. This proposal represents an alternative to the mobile network planning traditional process. We developed a network traffic simulation model in terms of service transmission rates of applications such as voice over IP, video phone, FTP file transfer and high definition video-phone. The simulations are executed using ARENA software. From the results of the traffic simulation model, we obtained the network capacity in terms of the cell radius. Based on the cost-benefit model and on the network capacity, we got the cell radius that maximizes the net profit percentage of the network and satisfies the user´s requirements. Under this approach is not taken into consideration some of technical aspects, but rather it is to highlight the economic aspect of the network planning process. Keywords: 3G mobile network, traffic simulation, network planning, economic model 1. INTRODUCTION The telecommunications industry has been expanded substantially since the past decade. Technology advancement along with the liberation of once closed markets and privatization of government-held monopolies changed the nature of the industry in the 1990s and continues to shake up the industry every now and then. In early 2000, the industry scaled new highs with respect capitalization. Both business and technology disruptions have introduced significant expansion and innovation. The global mobile cellular subscription was closed to 5.3 billion by the end of 2010. That is equivalent to 77 percent of the world population. So 90 percent of the world now lives in a place with access to a mobile network. For people living in rural communities this is lower at 80 percent, according to the estimation of the International Telecommunication Union. The mobile cellular technology, as happens in others technology fields, has had an innovation process in which has been defined and implemented successive generations. Each of these generations has responded to

187

So we obtain the cell number and their corresponding configuration to cover the airport total physical area. The study is done considering the operator´s view and we hope it can help to support the decision making in the telecommunication industry, specifically in the network design area.

3.

MOBILE PLANNING

CELLULAR

NETWORK

3.1. The process The network planning process consists of ten specific activities (Mishra 2007) which are carried out by different technical teams (see Fig. 2). It starts with the network design, which may or not be based on field measurements of an existing network. The next activity is the locations acquisition. The locations acquisition means that networks operators rent a physical space where sites will be built. After the acquisitions and building, the equipment is implemented through installation and commissioning. The commissioning is the activity on which the engineer field sets the correct value of the equipment parameters. This activity includes the integration of the new site to the total network according to the operator criteria. Once the commissioning is without technical and operational problems, the site is put into service. It means that phone calls can be processed by the site through the network. The activity which closed the process is the optimization. In order to optimize a new site and its integration to the telecommunication network, we need to collect measurements about the level and the quality of the signal and then to eliminate signal interferences in order to improve indicators such as the dropped call rate for example.

2. MOBILE CELLULAR TECHNOLOGY In order to have access in the mobile cellular telecommunication networks, users must have a terminal device based on a specific technology and operators must implement a network based on a specific technology too. Mobile cellular technology has evolved over generations. Thus, we have the so-called first generation (1G) technology, characterized by analog devices, while the so-called second generation (2G) technology, characterized by digital devices and based on standards as GSM (Global System Mobile) and CDMA (Code Division Multiple Access). And the socalled third generation (3G) technology based on standards such as UMTS (Universal Mobile Telecommunications System) (see Fig. 1).

Figure 1: Mobile cellular technology evolution In accordance with the UMTS specifications (3GPP 2002), using this technology makes possible to increase network bandwidth in order to get a major and better data transmission. In fact, this situation has encouraged to operators to offer a wider range of mobile services and applications, through introducing a new technology platform over their actual network (located physically between the user terminal device and the network controllers). So, using UMTS technology, the network must be planned in accordance with criteria evaluated by operators in order to satisfy customer requirements but customer requirements depend of customer communications needs, in certain places and in certain time. The operators must know the customer communications needs in order to be a very good option in a competitive market as is nowadays the mobile cellular telecommunication market.

Figure 2: The network planning process 3.2. Different type of cells The network design, the first activity in the network planning process, is carried out based on certain criteria which must be specified by the network operator in accordance with specific customer requirements. Normally, the network operator must adapt these requirements to his business model. As noted, the network design is an activity which may or may not be based on field measurements. To get field measurements is not always possible because for that, it is necessary that exists a mobile telecommunication network, does not matter if the existing mobile telecommunication network works with a different technology, the most important thing is to get the field measurements. If does not exist a network, the design is

188

We do not say that the cost-benefit criteria is not taken in account when a network operator design his own network based on macro cells, instead, we say that in many cases is more convenient for operators, in terms of trade, to base the network design on the service area criteria for macro cells and for the network design based on pico cells is more convenient for operators to be ensure the availability of the network in terms of services, applications and transmission channels.

made using simulation models based on theoretical assumptions. At the end of the network design, we can get the different types of cells, which will be implemented, as well as their capacity, both in terms of the number of transmission channels. The different types of cells depend on the physical area to which the cell needs to give the communication service. Thus, pico cell can be designed to provide service within buildings and are characterized because theirs dimensions are small and their transmission power is low. Technically, a one location correspond one kind of this cell. While micro cells are designed to provide service to shopping centers, open parks and business centers, for example, and are characterized as well as the pico cell by a low transmission power. We can also implement macro cells, which reach 2000 meters as radio service area with a transmission rate of 144 kb/s. If the applications require a better transmission rate, the network can be adapted to a major one, but the radio service area must decrease until to reach almost 500 meters (ETSI 1997). This type of cells is characterized by a big transmission power (see Fig. 3).

4. ANOTHER APPROACHES There are many approaches to design a mobile telecommunications networks and it depends on the priorities of the network operator. Once operators have defined their design criteria, it is necessary to select the technical design tools to be used. One of the most used tools in these cases is the simulation. At the present time, there are many kinds of simulators for UMTS telecommunications networks design which are configured in accordance with the parameters that we need to modeling and analyzing (Alonso y Lopez 2005). Simulators permit to analyze the network at different levels. The advantages for working with network simulators are the increased productivity in the network development, less time to market so total cost is decreased. Some examples of simulators used in the network planning are the follows: x x x x x

Figure 3: Different types of cells In order to be most competitive, in some cases the operators design a telecommunications network based on the service area criteria. This criteria works well in the cases of macro cells, it means, to cover a big area such as a community, because it helps operators to attract new customers. But in the case of a network designed to service inside of a building with pico cells, is very important to consider the cost-benefit approach and the specific user requirements because the customer will move around all the building and need to get the service in any point inside the building and the service operator need to be provided in all point inside the building in order to get customer satisfaction which will be converted in revenue for operators. So network operators must implement a network that first satisfies customer service needs and then, guarantees revenue for him.

Network simulator (System level), is possible to analyze the traffic, QoS, handover, admission control, Link simulator (Link level), controls the transmission errors, Physical layer simulator, is used to evaluate coverage area, power transmission, cells and interferences, Protocols simulator, verifies, analyzes and optimizes protocols, Integrated simulators, this kind of simulators has many functions integrated as the name indicates.

Many studies are development each year in accordance with the specific criteria of the network operators. The criteria could be technical, economic or both of them. 5. THE PROPOSAL In this paper, we propose an application approach to design a third-generation mobile network based on a traffic simulation model and a cost-benefit model, to provide 3G mobile telecommunication service at the interior of an airport. This approach consists of five general steps (see Fig. 4). 1. First, we develop a network traffic simulation model in terms of transmission services and third-generation applications, 2. Then, we perform the simulation experiments using the software ARENATM,

189

3. From the results of the experiments simulation, we obtain the network capacity measured in Mbits/s depending on the cell radius, 4. Based on the cost-benefit model and the network capacity, both as functions of the radio cell, we get the cell radius that maximizes the net profit percentage of the network, 5. So we obtain the network configuration, number of cells and radius cells, which satisfy the operator and users requirements. Figure 5: The international airport of Los Cabos, BCS By one hand, in accordance with Ferreira and Velez (2005), an application is defined as a work which requires the communication between one or more information flow, between two or more parts geographically distributed. The applications are characterized by the service attributes, the communication and the traffic characteristics. One essential data to the analyses purposes is the utilization rate for each application. By another hand, the services and applications distribution requirements can be over real time and over no real time as follows: x Real time, in this case, the applications need the information distributed for immediate use, x No real time, in this case, the information is stored in specific reception points to be used later.

Figure 4: An approach for design a 3G mobile network based on cost-benefit This approach represents an alternative to the mobile network design traditional process and we propose it to be used by network operators because this approach take account the users requirements and the operator cost-benefit, which are operation basic criteria in a competitive industry as actually is the telecommunication industry.

5.3. The network traffic model The most important traffic parameters included in a network design are transmission rate and average duration. The transmission rate is an average number of bits which are transferred between two devices, in each unit time. A possible measurement unit is kb/s. While the average duration represents the duration which each user uses a specific application, for own purposes. In accordance with Ferreira, Gomez and Velez (2003), a traffic generation model can be used to measure and to describe the traffic over the network only if it is based on the population density and on the service insight, so we can be able to know the call rate for each service. The potential services are grouped as follows:

5.1. The scenario The scenario considered is the international airport of Los Cabos, Baja California Sur, Mexico. Currently, this airport is ranked as the seventh most important in Mexico. In 2005, the airport received a total of 2,466,733 passengers, of whom 431,724 were domestic and 2.035,009 were international, in accordance with GAP (Grupo Aeropuertuario del Pacifico). The physical dimension of the terminal is 8440 meters square and it receives, on average, 730 passengers per hour every day (see Fig. 5).

x x x x

5.2. The 3G mobile services and applications According to the ITU (International Telecommunications Union), the network services and applications based on technology 3G are grouped as follows: x x

Interactive (Conversation, messages information download). Distribution (emission and cyclic).

Sound, Multimedia, Narrow band, Wide band.

For each service, we select one application in order to get a traffic model (see Table 1). The utilization figures are approximations proposed in this study and that can be tailored for particular cases.

and

190

Table 1: Services and applications Service Application Utilization Sound

Voice over IP

50%

Multimedia

Video-telephone

22%

Narrow band

File Transfer Protocol High Definition Video-Telephone

16%

Wide band

12%

The transmission rate and the duration statistic distribution, which characterize the applications selected, are included in Table 2 (Antoniou, Vassiliou and Jacovides 2003). Table 2: Technical specifications for applications Application Duration Transmission statistic rate distribution Voice over IP EXP/3 min 12 kb/s Video-telephone

EXP/3 min

128 kb/s

File Transfer Protocol High Definition Video-Telephone

EXP/0.1 min EXP/30 min

384 kb/s 1920 kb/s

5.4. The simulation study Now, we need to simulate the traffic model proposed in order to be able to analyze the traffic as function of the network capacity and the applications. Since the early days of simulation, people have constantly looked for new and better ways to model a system, as well as novel ways to use existing computer hardware and software in simulation (Law and Kelton 2000). For the purposes of this work, we use discrete-event simulation to model and analyze the network traffic. The steps that will compose a simulation study are showed in Figure 6. On the next sub-sections, we conduct the simulation study according to the steps in Figure 6.

Figure 6: The steps for a simulation study

The periods considered in one day are: x x x

8:00 hrs. – 12:00 hrs., 14:00 hrs. – 16:00 hrs., 17:00 hrs-18:00 hrs.

800 700 600 500 400 300 200 100 0 0:00 - 0:30 1:30 - 2:00 3:00-3:30 4:30 - 5:00 6:00- 6:30 7:30 - 8:00 9:00-9:30 10:30-11:00 12:00-12:30 13:30-14:00 15:00-15:30 16:30-17:00 18:00-18:30 19:30-20:00 21:00-21:30 22:30-23:00

Processed calls

5.4.1. Data collected The data which characterized the scenario were described in section 5.1 and the data which characterized the traffic were described in section 5.3.We propose the call arrival based on the statistics from the user arrival peak times at the airport. We considered that the periods on which more call are processed are when the users arrive to the airport in peak times, turn on the cellular phone and make a call.

Hour

Figure 7: Peak processed calls

191

In this study, we consider six configurations which take into account a different network capacity. The network capacity is based on the theoretical cell radius, we considered omnidirectional cells. Also, we considered that each cell is composed by seven traffic channels. For a radius cell of 30 meters corresponds a total capacity of 21 traffic channels. While for a radius cell of 21 meters corresponds a total capacity of 42 traffic channels, and so on until reaching a cell whose radius is less than 14 meters, which account for a total capacity exceeding the 98 traffic channels.

references, documentation and compatibility software and, by another side the second phase considers the problem characteristics as follows. We select ARENA software to this simulation study, so some of the questions are answered about this software. Phase one

5.4.2. The conceptual model defined The conceptual model about the traffic generated in a network is described in Figure 8.

x

There is a user manual? Yes,

x

The language code is compatible with actual computers? Yes,

x

The software has enough documentation and error diagnosis? Yes,

x

The language is known and easy to be learned? Yes,

x

The software is compatible with another kind of software? Yes.

Phase two x

What kind of real problems can be analyzed by the software? Process simulation, business simulation, supply chain simulation, logistic simulation.

x

Is easy to store and modify the system data? Is easy through modules.

x

Is easy to include user subroutines? Yes, users can create theirs owns modules and add to the software in order to create new systems.

5.4.4. Computer program Based on the conceptual model and on the ARENA software modules, we develop the computer program (see Figure 9).

Figure 8: Conceptual model Once a call starts, the network verifies if a circuit is available. For case negative, the call is rejected and the counter for this kind of calls is incremented. For case positive, the call is accepted and then is classified according to the type of call. There are four types of calls, each type corresponds a one application considered in the traffic model: 1. 2. 3. 4.

Voice over IP, Video-telephone, File Transfer Protocol, High Definition Video-Telephone.

Once a call is classified, it is processed in accordance with their duration statistic distribution specified in Table 2. Then the counter is incremented and when the call is finished, the circuit is free to be used by another call. 5.4.3. Software selection Elizondo and Flores de la Mota (2006) suggest a process which is based on some questions, in order to select a software simulation. The process is divided in two phases. By one side, the first phase is related with

Figure 9: The computer program using ARENA software

192

5.4.5. Verification & validation Once the model has been implemented, it is necessary to verify its performance. Based on the technique number 4 suggested by Elizondo and Flores de la Mota (2006), we verified the computer program. This technique consisted in the simulation execution considering many different scenarios so it was necessary to make many changes in the program parameters. After the executions, we checked the consistency of the results in accordance with the expert´s opinion. As results, we found that when the network capacity was increased, the number of rejected calls decreased and the number of call processed increased. This result corresponded with a real situation over a telecommunications network because a major network capacity correspond a major number of call processed and a minor number of calls rejected.

5.4.8. Use of the results At this point, we use the total number of calls processed in the cost-benefit model, in order to obtain the annual net profit as a function of cell radius for the different network configurations considered. 5.5. The cost-benefit model As stated in Gavish and Sridhar (1995), the economic aspect for a telecommunication network can be analyzed by four different approaches: user´s approach, service provider´s approach, regulator´s approach and manufacturer´s approach. By one side, for users, the most important aspects about the network are the service quality and cost. By another side, for service providers represented by networks operators the most important aspect is to find the network configuration which maximizes the expected revenues. While for the regulator, the most important aspect is the social welfare, to promote competition between network operators and to manage the frequency spectrum. Then, for manufacturers, the most important aspect about a telecommunication network is the equipment cost. So, each approach has their necessities to satisfy. The approach used in this contribution is the service provider. Because of the service provider business model, in this study we will take in account the impact of the other tree approaches. Cabral et al. (2005) suggest that the total annual cost of a radio network is determined by a fixed cost and a cost proportional to the number of cells required to service the area required by the operator. For this particular case, the fixed cost includes the cost of an operating license that the operator must ask the Mexican government.While the cost of each cell (node) is determined by the cost of equipment, the cost of installation and cost of operation and the maintenance as in (1).

5.4.6. Design of simulation experiments We carried out 17 simulation experiments, in each experiment we increased the number of traffic days. So, in the first experiment we simulated 5 traffic days, in the second experiment we simulated 10 traffic days, and so on, until in the experiment 17 we simulated 80 traffic days. 5.4.7. Analysis of output data 1652 total calls were processed for the network configuration with capacity of 21 traffic channels, 3155 total calls were processed for the network configuration of 42 traffic channels, 3505 total calls were processed for the network configuration with capacity of 48 traffic channels, 4146 total calls were processed for setting network of 70 traffic channels, 4173 total calls were processed for the network configuration with capacity of 84 traffic channels, and finally, 4171 total calls were processed for the network configuration with a capacity greater than 98 traffic channels (see fig. 10).

—ƒŽ…‘•–ሾ̈́ሿൌ ‹š‡†…‘•–൅ሺ‡ŽŽƒ—ƒŽ…‘•–ሻȗሺ‡ŽŽሻ(1)

On another side, the net income diary is obtained as a function of traffic carried across the network as in (2), i.e. the total number of calls processed by all cells in the network. It means taking into account the traffic flow at peak hours the network, as in the rest of the hours of operation of the network, calls are processed very few reaching sometimes be invalided.

Total processed calls

2500 2000 1500

—ƒŽ‡–‹…‘‡ሾ̈́ሿൌ‡–‹…‘‡†‹ƒ”›ȗ”ƒˆˆ‹…ƒ›•(2)

1000 500 0 0, f> u I k . Dynamics at a junction is obtained solving RPs.

or, alternatively, f İk ȡk , ȝk

­İȡk  (1  İ ) ȝk , 0 d ȝk d ȡk , ® ȡk d ȝk d ȝkmax , ¯ ȡk ,

Definition A Riemann Solver (RS) for the junction P with n incoming sub-chains and m outgoing ones (of n u m type) is a map that associates to a Riemann data ( ȡ0 , ȝ0 ) ( ȡ1, 0 , ȝ1, 0 ,..., ȡn  m , 0 , ȝn  m , 0 ) at P a vector

(3)

where ȡkmax and ȝkmax are, respectively, the maximum density and processing rate. From now on, we assume

( ȡˆ 0 , ȝˆ 0 ) ( ȡˆ1 , ȝˆ1 ,..., ȡˆ n  m , ȝˆ n  m ) so that the solution is given by the waves ( ȡi , 0 , ȡˆ i ) and ( ȝi , 0 , ȝˆ i ) on the sub-

202

chain I i , i 1,..., n and by the waves ( ȡˆ j , ȡ j , 0 ) on the Ij, j

sub-chain

n  1,..., n  m .

We

consistency condition RS ( RS (( ȡ0 , ȝ0 )))

require

Lemma On the incoming sub-chains, only waves of the first family may be produced, while on the outgoing sub-chains only waves of the second family may be produced.

the

RS (( ȡ0 , ȝ0 )).

From such Lemma, given the initial datum, for every RS it follows that:

2.1. Riemann Solvers for suppliers We discuss RSs for two types of nodes, according to the real case we examine here (for more detail refer to Bretti et al. 2007 and D’Apice et al. 2009):

1.

a node with two incoming sub-chains and one outgoing one ( 2 u 1 );

2.

a node with one incoming sub-chain and two outgoing ones ( 1 u 2 ).

ȡˆ i ȝˆ j

ij( ȝˆ i ), i 1,..., n, ȝ j,0 , j n  1,..., n  m,

where the function ij(˜) describes the first family curve through ( ȡk , 0 , ȝk , 0 ) as function of ȝˆ k :

For a given arc I k , (1) is a system of conservation laws in the variables U ( ȡ, ȝ) , namely: U t  F (U ) x

(6)

0,

ij ȝˆ k

with flux function ( f ( ȡ, ȝ),  ȝ ) .

F (U )

(7)

Eigenvalues and eigenvectors are:

Ȝ1 ( ȡ, ȝ) { 1,

Ȝ2 ( ȡ, ȝ)

­1, ® ¯İ,

r1 ( ȡ, ȝ)

ȡ  ȝ, ȡ ! ȝ,

­ § 0· ° ¨ ¸, ° ©1¹ ® 1 İ °§  1 İ · , °¨ 1 ¸ ¹ ¯© r2 ( ȡ, ȝ )

ȡ  ȝ,

2 ȝmax }.

(9)

We define two different RSs at a junction to represent two different routing algorithms: RA1. We assume that:

2 . 1 İ

(14)

(A) the flow from incoming sub-chains is distributed on outgoing ones according to fixed coefficients; (B) respecting (A), the processor chooses to process goods in order to maximize fluxes (i.e., the number of goods which are processed) on incoming sub-chains. RA2. We assume that the number of goods through the junction is maximized both over incoming and outgoing sub-chains. For both routing algorithms we can maximize the flux of goods considering one of the two additional rules: SC2. The objects are processed in order to maximize the flux with the minimal value of the processing rate. SC3. The objects are processed in order to maximize the flux. If a solution with only waves in the density ȡ exists, then such solution is taken, otherwise the minimal ȝ wave is produced. To define RPs according to rules RA1 and RA2, we introduce the notation:

(10)

Observe that:

ȝmax

ȡk , 0 d ȝk , 0 , ȡk , 0 ! ȝk , 0 .

ȝk

ȡ ! ȝ,

§1· ¨ ¸. ©0¹

­° ȡk , 0 , ® 1 İ 1 İ °¯ 2 ȡk , 0  2 ȝk , 0 ,

(8)

D {( ȡ, ȝ) : 0 d ȡ d ȡmax , 0 d ȝ d ȝmax ,

ȡmax

ȝˆ k t ȝk , ­ ȝk , ° ° ° İ  1 ȝˆ k  2 ȡk ,0 ® , ȝˆ k  ȝk , ȡk ,0 d ȝk ,0 , 1 İ ° ° İ  1 ȝˆ  ȝ k k ,0 °  ȡk ,0 , ȝˆ k  ȝk , ȡk ,0 ! ȝk ,0 , 1 İ ¯ (13)

where ȝk is the point at which the first family curve changes:

Hence the Hugoniot curves for the first family are vertical lines above the secant ȡ ȝ and lines with slope close to 1/ 2 below the same secant. The Hugoniot curves for the second family are just horizontal lines. Since we consider positive and bounded values for the variables, we fix the invariant region:

0 d (1  İ ) ȡ  (1  İ ) ȝ d (1  İ ) ȡmax

(12)

(11)

We consider a node P of n u m type and a Riemann initial datum ( ȡ1, 0 , ȝ1, 0 ,..., ȡn  m , 0 , ȝn  m , 0 ) . The following Lemma holds:

203

fk

fˆ1 f3max  f 2max , fˆ2 f 2max , fˆ3 f3max when A1 holds and A2 is not satisfied; fˆ1 f1max , fˆ2 f3max  f1max , fˆ3 f3max when x A1 is not satisfied and A2 holds. The case of both A1 and A2 false is not possible, since it would be f3max ! *inc . Now, we compute ȡˆ k and ȝˆ k , k 1, 2,3 . On the incoming sub-chains i, i 1, 2, we have to distinguish two subcases. If fˆi f i max , according to rules SC2 and SC3, we get:

(15)

f ( ȡk , ȝk ),

x

and define the maximum flux that can be obtained by a wave solution on each production sub-chain: k ­ ȝk , °  ȡ ȝ ® max max  ȝk , 0 ), k ° ȝk , 0  İ ( ȡmax  ȝk ,0 ȝmax ¯

f kmax

1,..., n, n  1,..., n  m.

(16) It is possible to prove that a necessary and sufficient condition for the solvability of RPs at nodes is

¦ fi min d i 1

ª

nm

n

¦ «ȝ

j n 1

¬

j,0

 İ ( ȡmax  ȝ j ,0

º ȡmax  ȝmax  ȝ j, 0 )» , ȝmax ¼

SC 2 :

(17)

ȡˆ i ȝˆ i

ȝi , ȝi ,

SC 3 :

ȡˆi ȝˆ i

ȝi , max{ ȝi , ȝi , 0 }.

(21)

If fˆi  f i max , for both SC2 and SC3 rules, we get that ȝˆ i , i 1, 2, solves the equation:

where

f i min ȡ0 , ȝ0

­ 2İ ȡ0 d ȝ0 , °°1  İ ȡ0 , ® °İȡ  İ 1  İ ȝ , ȡ ! ȝ . 0 0 0 °¯ 0 1 İ

ȝˆ i  İ (ij( ȝˆ i )  ȝˆ i )

(18)

fˆi ,

(22)

while ȡˆ i

2.1.1. One outgoing sub-chain In this case, algorithms RA1 and RA2 coincide since there is only one outgoing sub-chain. We fix a node P with 2 incoming arcs (labelled by 1 and 2) and 1 outgoing one (indicated by 3) and a Riemann initial datum given by ( ȡ0 , ȝ0 ) ( ȡ1, 0 , ȝ1, 0 , ȡ2, 0 , ȝ2, 0 , ȡ3, 0 , ȝ3, 0 ) . Let us denote

ij( ȝˆ i ), i 1, 2.

(23)

On the outgoing sub-chain we have, for both rules SC2 and SC3: ȝˆ 3

(24)

ȝ3, 0 ,

with ( ȡˆ , ȝˆ ) ( ȡˆ1 , ȝˆ1 , ȡˆ 2 , ȝˆ 2 , ȡˆ 3 , ȝˆ 3 ) the solution of the RP at P. We introduce a priority parameter q ]0,1[ , that indicates a level of priority at the junction of incoming sub-chains. We define:

while ȡˆ 3 is the unique value solving the equation f ( ȝ , ȡˆ ) fˆ , namely:

*

ȡˆ 3

min{*inc , * out },

3

(19)

where 2

¦f

*inc

max i

, * out

f

max 3

i 1

First, we compute fˆi , i 1, 2,3 according to rules (SC2) and (SC3). Introduce the conditions: (A1) qf3max  f1max ; *

*inc ,

max 1

f

we

get

that

fˆi

f i max ,

f

­ fˆ3 , fˆ3 d ȝ3,0 , ° ® fˆ3  ȝ3,0 °  ȝ3,0 , fˆ3 ! ȝ3,0 . İ ¯

(25)

that indicates the percentage of goods, which, from the incoming arc 1, is directed to the outgoing arc 2 (obviously, the arc 3 is interested by a percentage of goods equal to 1  Į ). We have different solutions for algorithms RA1 and RA2. In what follows, the asymptotic solution is reported only for the RA1 algorithm, since RA2 is solved as for the node with one outgoing sub-chain. As usual, we first compute the fluxes solutions. Following rules (A) and (B) of the algorithm RA1, we get that:

i 1, 2 ,

max 2

. If *  *inc , we have that: fˆ1 qf3max , fˆ2 1  q f3max , fˆ3 x 3

3

We introduce a distribution parameter Į  @0,1> ,

(A2) 1  q f3max  f 2max . If fˆ

3

2.1.2. One incoming sub-chain Consider a node P with 1 incoming arc, labelled by 1, and 2 outgoing ones, indicated by 2 and 3 and an initial datum ( ȡ0 , ȝ0 ) ( ȡ1, 0 , ȝ1, 0 , ȡ2, 0 , ȝ2, 0 , ȡ3, 0 , ȝ3, 0 ) .

(20)

.

3, 0

f3max when

A1 and A2 are both satisfied;

204

fˆ1

­ f max f max ½ min ® f1max , 2 , 3 ¾ , fˆ2 Į 1 Į ¿ ¯

Įfˆ1 , fˆ3

q 0.6 indicates that 60% of goods flow is directed from arc 1 to the outgoing one. In Table 1, numerical results for asymptotic fluxes, densities and production rates are reported while Figures 1 and 2 show the behaviour of density and production rate waves. For both rules SC2 and SC3, the results are the same, with the exception of values ȝˆ1 and ȝˆ 2 for rule SC3. For sub-chain 3, a shock wave in the density connect the initial and the asymptotic state while, for sub-chain 1 and 2, there is no waves formation (Figure 2). A similar situation happens for production rates (Figure 3): in the case SC2, only sub-chains 1 and 2 are interested by waves formation. For rule SC3, shock formations do not occurs, as all sub-chains have asymptotic states equal to the initial ones. In fact SC2 tends to make adjustments of the processing rate more than SC3.

1  Į fˆ1 . (26)

Densities and processing rates, ȡˆ i , and ȝˆ i , i 1, 2,3, are obtained as follows. If fˆ1 f1max , we get: SC 2 :

ȡˆ1 ȝˆ1

ȝ1 ,

SC 3 :

ȝ1 ,

ȡˆ1 ȝˆ1

ȝ1 ,

(27)

max{ȝ1 , ȝ1, 0 }.

If fˆ1  f1max , ȝˆ1 , either for rule SC2 or SC3, satisfies the equation: ȝˆ1  İ (ij( ȝˆ1 )  ȝˆ1 )

fˆ1 ,

Table 1: Numerical results for a node of 2 x 1 type RA1 = RA2 SC2 SC3 ˆf (0.35, 0.2, 0.55) (0.35, 0.2, 0.55)

(28)

while: ȡˆ1

ij( ȝˆ1 ).

(29)

On the outgoing sub-chain j, j SC2 and SC3, we have that: ȝˆ j

2,3, for both rules

­ fˆj , ° ® fˆj  ȝ j ,0 °  ȝ j ,0 , İ ¯

(0.35, 0.2, 0.3)

(0.95, 0.55, 0.3)

r

^

1

`

i

(0.35, 0.2, 0.6),

( ȝ1,0 , ȝ2,0 , ȝ3,0 )

(0.95, 0.55, 0.3).

x

r 2,0

m

t > 0m1 ,0

m

` m 1 m3 ,0

i

m2 ,0

m3 ,0

x

m2 ,0

` m 2

x

Figure 3: Production rates at t = 0 and t = 1 on subchains for rule SC2 3. SIMULATIONS In this section, we present some simulation results to foresee the behaviour of goods fluxes on a supply network. In particular, we study how to choose the injection times of different goods levels to increase the production.

0.8, i 1, 2,3,

( ȡ1,0 , ȡ2,0 , ȡ3,0 )

r` 2

Figure 2: Densities at t = 0 and t = 1 on sub-chains for rules SC2 and SC3

2.2. Example In what follows we report densities and production rates at the instant t 0 and after some times (at t 1 ) for different initial data using different routing algorithms. We consider a node of type 2 x 1, assuming the following data: 0.25, ȝimax

r` 3 r 3,0

r 2,0

(31)

fˆj ! ȝ j ,0 .

r

r 1,0

x

fˆj , namely:

fˆj d ȝ j ,0 ,

2

t>0

r` 1

are obtained for rules SC2 and SC3 as before.

İ

ȝˆ

r 2,0

Remark For sequential sub-chains (one incoming arc, 1, and one outgoing arc, 2), the fluxes solutions are fˆ fˆ min f max , f max while ȡˆ , and ȝˆ , i 1, 2, 2

(0.35, 0.2, 1.01)

(30)

ȝ j,0 ,

t =0 m1 ,0

1

(0.35, 0.2, 1.01)

r 1,0

while ȡˆ j solves the equation f j ( ȝ j , 0 , ȡˆ j )

ȡˆ j

t=0

ȡˆ

(32)

3.1. Numerical methods We refer to a Godunov method for a 2 u 2 system (details are in Bretti et al. 2007, Godunov 1959), which

As there is only one outgoing sub-chain, algorithms RA1 and RA2 coincide and the choice

205

is described as follows. Define a discrete grid in the

x, t ,

plane j

,



x ,t n

whose points are

j

, and indicate by

k

j'x, n't ,

U nj

k

and

P nj

't 1 d . 'x 2

,

respectively, the approximations of density and



production rate of the arc I k in the point x j , t

n

3.2. A complex network We present some simulation results for a supply network, whose topology is in Figure 4.

. An

approximation scheme for the system (1) reads as: ­ k n 1 °° U j ® ° k P n 1 °¯ j

't k n k n g U j , U j 1  g 'x 't k n k n Pj  P j 1  k P nj , 'x

k

U nj 



kU





n j 1

k

, Uj

n

,

(33) Figure 4: Network with 8 nodes and 10 arcs

where the Godunov numerical flux g is found solving RPs among the states U  , P on the left and

U , P

Such a network can model the chips production. First, potatoes are washed (arc I1 ) and then they are skinned off (arc I 2 ). Assuming that two different types of fried potatoes are produced (classical and stick, for example), node 2 is a diverging point: a percentage Į of potatoes are sent to arc I 3 for stick chips production, and a percentage 1  Į to arc I 4 for the classical potatoes production. On arcs I 5 and I 6 , potatoes are fried and on arcs I 7 and I 8 they are salted. Node 7 is a merging point: considering a certain priority level q, potatoes are directed to arc I 9 where they are put in envelope; on arc I10 , the obtained packets are sealed. The goods evolution inside the supply network is simulated in a time interval > 0,T @ , with T = 1000 min,

on the right:

g U  , P , U  , P  ­ U  ,  P  , ° °§ 1  H P  2H U ,  P · ,  ¸ °¨© 1  H  1  H  ¹ ° 1 H ®§ 1  H · °¨ 2 U   2 P  ,  P  ¸ , ¹ °© °§ 1  H · °¨ 1  H P   HP   HU  ,  P  ¸ , ¹ ¯©

U   P , U  d P , U   P , U  ! P , U  t P , P ! P , U  t P , P d P , (34)

't 1 . 'x 2 The dynamics at node 2 is solved using the RA1 algorithm. In fact, the redirection of potatoes in order to maximize the production on both incoming and outgoing sub-chains is not possible, since classical and stick potatoes have different shapes. Moreover, at node 2, we use rule SC2 and a distribution coefficient Į 0.3 for arc I 3 . At node 7, dynamics is solved using the RA1 algorithm with rule SC3 and priority level q 0.4 for arc I 7 (notice that, for such last node, algorithm RA1 and RA2 coincide). We assume that, at the beginning of the simulation (t = 0), all arcs are empty. Moreover, in Table 2, initial conditions for processing rates, maximal processing rates, lengths and processing times, are reported for each arc I k , k 1,...,10 .

with

P

(37)

using the approximation scheme (33) with

P 

1 H U  P . 2

(35)

We need to introduce the boundary data value, given by the term k U nj1 . For the first arc of the supply network, k U nj1 is defined by an assigned input profile; otherwise, k U nj1 is determined by the solution to RPs at nodes. Remark. The construction of the Godunov method is based on the exact solution to the RP in the cell º x j 1 , x j ª u º t n , t n 1 ª . To avoid the interaction of waves ¼ ¬ ¼ ¬ in two neighbouring cells before time 't , we impose a CFL condition like:

't 1 max ^ O0 , O1 ` d , 'x 2

Table 2: Parameters for the supply network Ik Pk ,0 Pk max Lk Tk

(36)

where O0 and O1 are the eigenvalues of system (1). Since, in this case, the eigenvalues are such that O0 1 , O1 d 1 , the CFL condition reads as:

206

1

10

15

15

15

2

7

10

30

30

3

7

10

20

20

4

15

20

15

15

5

5

8

20

20

6

5

10

20

20

7

12

12

20

20

8

10

10

25

25

9 10

15 10

15 10

15 10

HL 1.5

r 10 t,x

1 0.5 0 0

10 8 6

15

400 t

U1,b t , 0

800 1000

0

Figure 5: ȡ10 t , x for t

100

HL

r 10 t,x

2

10

1 0 0

8 6 400

2

600 800 1000

(38)

0

Figure 6: ȡ10 t , x for t

where t is the time instant at which the injection levels inside the supply network abruptly change. Notice that levels 30 and 5 of ȡ1,b have been chosen according to

x

4

200 t

­30, 0 d t d t , ® ¯ 5, t  t d T ,

2

600

10

Maximal densities on arcs are obtained using equation (11), where we consider İ 0.2 . Boundary data are also needed: for arc 1, it represents the amount of goods, that have to be processed inside the supply network; for arc 8, it is a sort of wished production. The input profile for arc 1 is chosen as a constant piecewise function with one discontinuity, namely a Heavyside function. In fact, during production processes, goods are injected inside supply networks at almost constant levels in different time intervals:

x

4

200

500

In Figure 7 and 8, fixing t 500 we show how the dynamics of the supply network is influenced by different choices of RSs at nodes 2 and 7.

the following criterion: when 0 d t d t , the arcs of the supply network process a great amount of goods and often reach the maximal density; when t  t d T , the arcs process goods whose density is always less than the maximal one. For arc 8, we assume a boundary datum equal to max ȡ10 16.667 , hence we require a possible wished output near to the maximal density processed by arc 10. The aim is to choose some t value, that guarantees maximal production. First we examine the behaviour of ȡ10 t , x , for t 100 and t 500 . The overall system

HL 20

m7 t,x

20

10

15

0 0

10 x

200 400 t

Figure 7: ȝ7 t , x for t

is completely influenced by t . In Figure 5, we notice one production peak at time, approximately, t 400 , but the average level of density is quite low (about 0.6). Such phenomenon is not present in Figure 6, where there is one peak production, and, after it, the production decreases slowly until it reaches a fixed constant level.

5

600 800

1000 0

500 using rule SC2 at node 2

and SC3 at node 7

HL

m7 t,x

2

20

1 0 0

15 10 x

200 400 t

Figure 8: ȝ7 t , x for t and SC2 at node 7

207

5

600 800

1000 0

500 using rule SC3 at node 2

The function f ȡ10 t , L10 , ȝ10 t , L10 , namely the

1000

flux on the last point of arc 10, in the case of rules SC2 - SC3, is depicted in Figure 9 for different choices of t to understand the final product flows. The obtained results present some interesting features: first, although different values of t are used, the flux starts to be different from zero always at the same temporal instant ( t 350 ), indicating that the input flow does not influence the production dynamics, that depends only on network characteristics (initial conditions, maximal processing rates, arcs length, and so on); second, shifts of the input flow discontinuity do not foresee f ȡ10 t , L10 , ȝ10 t , L10 . Such translations of

900

J

800 700 600 500 200

400

600 t

Figure

10:

behaviour

J t

of



1000

for

different

combinations of rules at nodes 2 and 7: SC2 – SC2 (dot dashed line); SC2 – SC3 (continuous line); SC3 – SC2 (dashed line); SC3 – SC3 (dot dot dashed line)

phenomenon indicates that, also using a conservation law with a linear function and a transport equation for the production rates, the dynamics on the whole network is strongly not linear.

We get that t max is almost insensible to rules at nodes 2 and 7 and its numerical approximation is t max 830 . The just made analysis strictly depends on the input flow characteristics and network parameters. In general, the behaviour depicted in Figure 8 is a priori unpredictable due to the non linearity of supply networks, as confirmed by other similar simulation. In Figures 11 and 12, ȡ10 t , x and ȝ10 t , x are

7 6 5

f

4 3

represented for t 830 in the case of SC2 – SC3 rules: ȡ10 t , x is higher with respect to other cases already

2 1 400

500

600

700

800

900

examined in Figure 5 and 6; ȝ10 t , x is constant and

1000

equal to 10. Such result is not surprising since, according to RSs at nodes, the production rates are kept equal to the initial ones on outgoing sub-chains.

t

Figure 9:

f ȡ10 t , L10 , ȝ10 t , L10

different values of the t 100 (continuous line), t t 500 (dot-dashed line)

evaluated for

discontinuity instant: 200 (dashed line) and

The area described by f ȡ10 t , L10 , ȝ10 t , L10 ,

HL

that can have strong variations for different t , represents the number of goods produced at the end of the simulation. In particular, we could ask if there exists a value t for which

r 10 t,x

2

10

1 0 0

8 6 400 t

2

600 800 1000

³ f ȡ t , L , ȝ t , L dt 10

10

10

10

(39)

is maximum. In Figure 10, J t



HL

m10 t,x

(precisely if t  > 200, 700@ ), until it

reaches a maximum t in a constant way.

max

830

is reported for the following

combination of rules at nodes 2 and 7: SC2 – SC2, SC2 – SC3, SC3 – SC2, and SC3 – SC3. We observe that J t almost increases linearly for a wide range of values of t

0

Figure 11: ȡ10 t , x for t

0

x

4

200

T

J

800

20 15 10 5 0 0

10 8 6 4

200

, and then it almost decreases

400 t

2

600 800 1000

Figure 12: ȝ10 t , x for t

208

0

830

x

Armbruster, D., Degond, P., Ringhofer, C., 2006. Kinetic and fluid models for supply chains supporting policy attributes, Transportation Theory Statist. Phys. Armbruster, D., Marthaler, D., Ringhofer, C., 2004. Kinetic and fluid model hierarchies for supply chains, SIAM J. on Multiscale Modeling, 2 (1), pp. 43-61. Bretti, G., D’Apice, C., Manzo, R., Piccoli, B., 2007. A continuum - discrete model for supply chains dynamics, Networks and Heterogeneous Media (NHM), 2 (4), pp. 661-694. Daganzo, C., 2003. A Theory of Supply Chains, Springer Verlag, New York, Berlin, Heidelberg. D’Apice, C., Manzo, R., 2006. A fluid dynamic model for supply chains, Networks and Heterogeneous Media (NHM), 1 (3), pp. 379-398. D’Apice, C., Manzo, R., Piccoli, B., 2009. Modelling supply networks with partial differential equations, Quarterly of Applied Mathematics, 67 (3), pp. 419-440. Helbing, D., Lammer, S., Seidel, P., Seba, T., Platkowski, T., 2004. Physics, stability and dynamics of supply networks, Physical Review E 70, 066116. Helbing, D., Lammer, S., 2005. Supply and production networks: from the bullwhip effect to business cycles, in D. Armbruster, A. S. Mikhailov, and K. Kaneko (eds.) Networks of Interacting Machines: Production Organization in Complex Industrial Systems and Biological Cells, World Scientific, Singapore, pp. 33-66. Godunov, S. K., 1959. A finite difference method for the numerical computation of discontinuous solutions of the equations of fluid dynamics, Mat. Sb. 47, pp. 271-290. Göttlich, S., Herty, M., Klar, A., 2005. Network models for supply chains, Communication in Mathematical Sciences, 3 (4), pp. 545-559. Göttlich, S., Herty, M., Klar, A., 2006. Modelling and optimization of Supply Chains on Complex Networks, Communication in Mathematical Sciences, 4 (2), pp. 315-330.

A further remark can be done on the dependence of J t by the distribution coefficient Į at node 2. In Figure 13, we represent different pictures of J t , evaluated using rules SC2 – SC3, for different values of Į . It is evident that, if Į grows, J t becomes higher but the value of t at which it attains its maximum point has no meaningful variations. 2000

J

1500

1000

500

200

400

600 t

Figure 13: behaviour of J t



800

1000

using SC2 – SC3 for

different values of Į : Į 0.1 (dashed line); Į 0.3 (continuous line); Į 0.5 (dot dot dashed line) and Į 0.8 (dot dashed line) 4. CONCLUSIONS In this paper, starting from the model proposed in D’Apice et al. 2009, goods flows on a supply network have been studied. An input flow of piecewise constant type with only one discontinuity has been chosen for simulating the behaviour of a supply network for chips production. Recent studies on experimental data seems to confirm the correctness of the assumptions underlying the model. In particular, the real flow profiles on each arc are consistent with the shapes of the flux functions. For such a network it has been proven that an accurate choice of the discontinuity point allows to maximize the total final production. The influence of the supply evolution on RSs at nodes and on the distribution parameter is analyzed. In future we aim to develop numerical schemes to solve the optimal control problem of choosing an input flow of piecewise constant type in order to obtain an expected pre-assigned network outflow. The idea is to find the minimum of a cost functional measuring the network outflow evaluating its derivative with respect to the switching times (the controls) of the input flows through the evolution of generalized tangent vectors to the control and to the solution of the supply chain model.

AUTHORS BIOGRAPHY CARMINE DE NICOLA was born in Salerno, Italy, in 1972. He graduated in Electronic Engineering in 2002 with a thesis on simulations of processor IAPX 86. He obtained a PhD in Mathematics at the University of Salerno in 2011 discussing a thesis about supply networks modelling and optimization techniques. He is actually a research assistant at the University of Salerno. His scientific interests are about fluid – dynamic models for the analysis of traffic flows on networks, operational research models in airport management, and queueing theory. His e-mail address is [email protected].

REFERENCES Armbruster, D., Degond, P., Ringhofer, C., 2006. A model for the dynamics of large queueing networks and supply chains, SIAM Journal on Applied Mathematics, 66 (3), pp. 896-920.

209

ROSANNA MANZO was born in Polla, Salerno, Italy. She graduated cum laude in Mathematics in 1996 and obtained PhD in Information Engineering in 2007 at the University of Salerno. She is a researcher in Mathematical Analysis at the Department of Electronic and Information Engineering University of Salerno. Her research areas include fluid – dynamic models for traffic flows on road, telecommunication and supply networks, optimal control, queueing theory, self – similar processes, computer aided learning. She is author of about 40 papers appeared on international journals and many publications on proceedings. Her e-mail address is [email protected]. LUIGI RARITÁ was born in Salerno, Italy, in 1981. He graduated cum laude in Electronic Engineering in 2004, with a thesis on mathematical models for telecommunication networks, in particular tandem queueing networks with negative customers and blocking. He obtained PhD in Information Engineering in 2008 at the University of Salerno discussing a thesis about control problems for flows on networks. He is actually a research assistant at the University of Salerno. His scientific interests are about numerical schemes and optimization techniques for fluid – dynamic models, and queueing theory. His e-mail address is [email protected].

210

SUGAR FACTORY BENCHMARK Rogelio Mazaeda (a), Alexander Rodríguez (b), Alejandro Merino (c), César de Prada(d), Luis F. Acebes (e) (a)(b)(c)(d)(e)

Systems and Automation Department, School of Industrial Engineering, University of Valladolid Center of Sugar Technology (CTA)

(a)

[email protected], (b) [email protected], (c)[email protected], (d)[email protected] (e) [email protected]

ABSTRACT This paper describes a simulated benchmark specifically designed for trying out and comparing different decentralized control strategies applicable to large scale complex process plants. The benchmark represents a reduced version of a typical beet sugar factory and basically corresponds to the interrelation of the Evaporation and the Sugar End sections. The underlying dynamic model is full of realistic details and has been derived from first principles, stating all the involved mass, energy and population balances. The ready to be used executable is offered to the interested parties as a standard and documented package accessible from any system implementing the widely used OPC process communication protocol.

obtained steam, with important energy content, is reused in the same section and in other departments of the plant. On the other hand, the downstream Sugar End, is a very complex installation organized in several stages (in figure 2.b the main "A" stage is depicted), with many individual batch and continuous pans performing the highly uncertain and poorly measured process of sugar crystallization. The Sugar House and the Evaporation sections (see figure 1) interacts strongly and not only due to the interchange of the stream of concentrated syrup to process. In addition, the crystallizer pans are heavy consumers of the steam which is served by the evaporators cascade. The overall management and control of these two specific departments constitutes, then, an important challenge. The difficulty of the task is very much compounded by the need to coordinate continuous and semi-batch type units.

Keywords: Sugar production, hybrid models, industrial process, hierarchical control, distributed control 1. INTRODUCTION The typical beet sugar factory consists of several sections in series, which are respectively concerned with the extraction of the sucrose out of the sliced beets by a diffusion in water process, the elimination of as many impurities as possible in a purification plant, the concentration of the resulting still impure sucrose solution in a cascade of evaporators and finally, the Sugar End or House, where the crystallization of the dissolved sucrose in batch and continuous crystallizers is carried out to deliver the white sugar grains with commercial value. The preceding cursory description is deceptively simple. Each of the above mentioned sections is a full fledged plant on its own right, hosting many process units, some of them of difficult individual operation (Poel, Schiweck and Schwartz, 1998). Additionally, each specific unit participates in a complex layout with numerous mass and energy recycles implying a tightly integrated environment that makes the overall management of the factory an arduous task. The Evaporation section (fig. 2.a), for example, is made up of several serially connected units in an arrangement that seeks to improve the efficiency of the factory by reusing the steam served by the boiler. The primary objective here is the elimination of the extra water contained in the incoming fresh juice, but the

Figure 1: Top level view of the benchmark The efficient and smooth conduction of the Sugar Factory, and this is true for most process plants, is a complex undertaking. The required solution goes well beyond the regulation of some variables to fixed setpoints at the individual unit level. Nowadays, in most factories, the needed coordination is heuristically tackled by the operators and technical personnel. Automatic solutions are, of course, possible. In the specific case discussed, where the scheduling of the different batch pans is critical, the control scheme described in Prada, Sarabia, Cristea and Mazaeda (2008) does the job. They use a simplified

211

model of the section in a coordinating central MPC controller implementing a novel control signal continuous re-parametrization of an otherwise discrete or hybrid problem, to find the optimal solution with a reasonable utilization of computer resources.

example, at the entrance of a plant of several hours processing time, would not imply too much of a difference for the control actions to be currently adopted at the output. It is then possible for the local control to work, and this would mean important advantages like the possibility of a scalable distribution of the control computer job. Additionally, in the long run, it would result in a conceptually simpler problem. A recent review of the subject is to be found in Scattolini (2009). In this context, a dynamic realistic first principle model of a reduced scaled down version of the Evaporator and Sugar End is going to be offered as a benchmark for testing and comparing different plantwide control strategies. The benchmark to be described has officially been adopted by the European Network of Excellence entitled Highly-complex and networked control systems, HYCON2, (Framework Program 7 Network of Excellence HYCON2, 2011). In section 2 a description of the process represented in the benchmark is given, section 3 briefly explains the main assumptions and characteristics of the underlying model and gives references to more detailed descriptions, while section 4, for the sake of conclusions, discusses some of the control and plantwide coordination challenges posed by the proposed simulated plant. 2. BENCHMARK DESCRIPTION The benchmark proposed consists of the interrelation between the evaporation section and the first stage of a Sugar House. The scaled down version incorporates detailed first principle models of three evaporators in series and the same number of parallel batch crystallizers, a very simplified, static representation of the A stage centrifuges and along with the necessary auxiliary equipment such as buffer tanks. The first evaporator receives the stream of technical sucrose solution, the thin juice, with a certain flow rate, purity and concentration to deliver the thick juice at a much greater concentration or Brix. In the real situation the purity and Brix would depend on the beet quality and the workings of the previous sections, but here are considered as given. The concentration of the thick juice is enforced by the PID loop controlling the vacuum pressure in the chamber of the last evaporation unit. The energy for heating the juice in the first evaporator comes from the live steam delivered by the factory boiler but the second and third units re-uses the vapors obtained from the evaporation of part of the water conforming the juice which is processed in the previous one. This reutilization scheme increases the overall factory's efficiency and is possible since each successive unit is operated at a lower pressure, and so temperature, than the upstream unit.

Figure 2: Benchmark detailed view. a) Evaporators cascade. b) Sugar End A stage In any case, there is an emergent notion in the control community stating that centralized optimizing schemes, in spite of their unquestionable mathematical properties concerning, for example, the rigorous optimality of the solution, should not be the only way to go, especially in the case of large plants. It is argued, for example, that the burden of maintaining a central good enough model is too exacting; and that the associated numerical problem scales badly with the size of the plant. On the other hand, the structure of the typical process plant, where the main product stream travels serially from installation to installation, with occasional recycle loops, seems to suggest that a decentralized but coordinated or hierarchical control arquitecture would be a reasonable, maybe suboptimal, way to go. The intuition being that the disturbances appearing, for

Steam reuse is not limited to the evaporation cascade. An important fraction of the water evaporated from the syrup is diverted to provide heating energy at other sites in the factory. The subsequent Sugar House

212

department is a particularly heavy steam consumer and the steam demand it represents on the evaporation section is very severe due to the large amounts required and to its intermittent character. The vapours obtained from the last evaporator are sunk in a barometric condenser. It is common knowledge of the sugar industry that the flowrate of this stream should be minimized since it represents wasted energy. The difference between the boiler steam pressure at the input and the one that is enforced in the last evaporator chamber by controlling the flowrate to the condenser, drives the downstream vapour flow. A real factory Sugar House has an architecture consisting in various, usually three, stages: the first or A stage is dedicated to the production of the commercial white sugar crystals and the rest to the exhaustion of the remaining syrup. The benchmark only represents the first stage. It consists of the parallel array of semi-batch crystallizers followed by a similar disposition of batch filtering centrifuges. The crystallizers receive the so called standard liquor and deliver the massecuite: a viscous slurry consisting of the grown sugar crystal population which is suspended in the resulting syrup or mother liquor. At the end of their respective strike, each pan discharges the massecuite into a common strike-receiver tank. Next the downstream centrifuges perform the required step of separating the sugar grains form the mother liquor. The purity quality requirements of the white sugar product determine the need of applying water during the centrifuging process for improved filtering. Water helps in expelling the traces of mother liquor from the crystal faces but re-dissolves part of the sugar crystal mass. As a result of this technological setup, each centrifuge offers two type of syrups classified according to their purity: first the so called poor or green syrup, with purity similar to the original mother liquor and then the rich or wash run-off syrup, of higher purity, enriched with the re-dissolved sucrose. The poor syrup gets processed in the following stages and the rich syrup is directly recycled in the A stage. The standard liquor is conformed in the melter. The main part is the thick juice coming from the evaporators but it also receives the contribution of the above mentioned A rich syrup and of the recovered sugar crystallized in the B and C stages. In the benchmark the melter is simply modelled as a level controlled open tank which assumes the instantaneous and full dissolution of the crystal streams. The interplay between the workings of the crystallizers and those of the centrifuges is very important for the efficiency of the factory: a bad strike with a not uniform population of crystals would have a poorer filtering capacity in the centrifuges, so it would demand more water with the associated negative impact of the efficiency. To keep the complexity of the benchmark under check, the mentioned compromise is not modelled. In any case, the simplified model of the

centrifuge, allows, as a perturbation, the introduction of more water, always a prerogative of the human operator, to simulate the mentioned effect. It is to be noted that the efficiency impact associated with this action, should be understood also in the sense that the obtained syrups are more diluted, so it would imply and extra evaporation effort in the crystallizers and more processing time to each cycle or alternatively a greater demand of steam from the evaporation section. The differences in rhythm between continuous and batch operated equipment determines the existence of buffer tanks of the appropriate size in the flowsheet of the plant. The standard liquor tank, in particular, which serves the feed syrup to the pans, should accommodate the peaks in demands from the crystallizers with the continuous supply of standard liquor from the melter. In the typical plant, the operator schedules the workings of the parallel array of pans to keep the syrup inventory in the container between safe limits. Observe that the long and uncertain processing times of the batch units make the task difficult. A similar situation exist in the strike receiver, whose level is controlled by modifying the working rhythm of the batch centrifuges, and so their throughput, but the problem here is easier due to the short cycle time and predictability of their time controlled cycles. 2.1. Evaporator The evaporators considered are of the Robert type. Each unit has two chambers. The heating chamber or calandria encloses a set of vertical tubes that contains the boiling juice. The heating steam enters the shell of the calandria and the energy needed for boiling is transferred to the juice inside the tubes. The heating steam condenses on the shell walls and finally leaves as condensate water. The juice to concentrate enters at the bottom of the chamber, is heated and rises in the interior of the tubes driven by the resulting vigorous bubbling effect produced while boiling. The more concentrated syrup goes over the rim of the tubes and falls into the central downtake and to the output of the unit. The steam produced from the water evaporation emerges from the juice phase, and reaches the upper space containing the vapour. There is a pipe at the top which leads the vapour resulting from evaporation to the calandria of the next station or to the condenser. There is a complex steam delivery circuit that distributes the vapour stream to other units, especially to the batch crystallizers in the Sugar House. The juice level in the central downtake must be regulated to assure that the generated vapours are safely sealed in the upper part of the unit and to guarantee the height difference with the input of the downstream evaporator which is needed to drive the juice flow. 2.2. Batch crystallizers The process of crystallization is possible when the concentration of the solute to crystallize, sucrose in this case, exceeds the concentration defining the solubility

213

of the substance at the given temperature. The supersaturation, which is conventionally defined in the sugar industry as the ratio of the two mentioned concentrations, should be greater than unity for crystallization to occur. For moderate values of supersaturation, a metastable zone is defined where the growing of existing crystals is possible but the probability of the creation of new grains out of the solution, is negligible. If the supersaturation, however, is increased so that it trespasses the fuzzy defined labile zone frontier, then the nucleation phenomenon turns explosive, a situation to be avoided in industrial crystallization processes. The solubility of the sucrose increases with temperature and with the presence of impurities. In sugar industry supersaturation can be obtained by cooling or by evaporation of the water in excess. In the present benchmark the batch crystallizers use the latter mechanism. The batch sugar industrial crystallizer serves the purpose of creating the conditions of supersaturation of the technical solution of sucrose, so that a tiny initial population of sugar crystals may steadily grow until it achieves the commercial average size. It is an important technological requirement that the spread of the distribution of sizes in the population is kept as narrow as possible and this implies that supersaturation should be always kept at moderate values in the so called metastable region. The supersaturation conditions are created by striking the right balance between the rates of water evaporation and of the standard liquor which is supplied to replenish the solution of the sucrose that has migrated to the faces of the crystals. The evaporation is carried out at low, vacuum pressures so as to keep the temperature of the mass at reduced values so avoiding the quality impairing caramelization of sugar. The process is conducted in a semi-batch fashion, and this means that the prevalent conditions are continuously varying along the strike. The impurities, for example, get accumulated along the cycle, so the mother liquor purity gets progressively lower and this implies a greater difficulty in keeping the right supersaturation. The amount of massecuite in the pan grows from an initial value of roughly half to the full capacity of the pan at the end; and this fact has an important negative impact on the circulation of the mass and on the the heat transfer efficiency of the unit. The difficulties in the conduction of the pan are compounded by the absence of important on line measurements. The supersaturation, for example, that is critical, depends of other variables like the concentration, the temperature and the purity of the solution. Purity is not measured on-line but it is periodically reported by the factory laboratory. The online measurement of the concentration of the solution or Brix is problematic and the temperature of the mass is not homogeneous, especially at the end of the strike because of poor circulation, so its determination is also uncertain.

Figure 3: Vacuum pan crystallizer The vacuum pan is constructed as a cylinder with a diameter and height of comparable dimensions. At the bottom, there is a floating calandria type of heat exchanger. The mass circulates inside the tubes and the central downtake, and the heating steam goes inside the shell. The calandria is designed so as to bolster the circulation of the mass. The massecuite rises in the tubes as the water component is evaporated. The steam rises through the existing mass and emerges at the surface to enter the vapour occupied phase. The massecuite, which is driven over the tubes rim by the rising bubbles, returns to the bottom via the central downtake. In order to intensify the circulation, a mechanical stirrer is placed at the bottom of the downtake. The stirrer contribution is important mostly at the end of the strike, when the achievable evaporation is lower and the natural agitation provided by the bubbles are probably not enough. The pan is provided with the necessary valves to regulate the heating steam input to the calandria, the feed syrup to the chamber and the seed magma containing crystal initial population. There are valves that allow the discharge of the mass at the end of the strike, and the control of the evaporated steam out of the camber to the condenser. There is also a cleaning valve for inputting steam after the product evacuation with the purpose of removing the traces of massecuite. Note that in the specific unit modelled (fig. 3), it is allowed to individually choose the evaporator effect which is going to be the source of heating steam. Normal practice dictates the use of lower pressure steam (II effect) at the beginning of the cycle when the heating process is more efficient and then, to switch to higher pressure steam (effect I) as the cycle progresses and the mass transfer coefficient rapidly diminishes. The instrumentation used in the pan allows to measure the following variables: chamber mass temperature and steam space pressure, the level attained by the mass and the steam pressure in the calandria. The unit has a radio frequency (RF) sensor whose on-line readings can be calibrated to somehow represent the

214

concentration of the slurry: solution plus growing crystals. The electrical current which is drawn by the stirrer motor, could be taken as a indication of the consistency of the mass, and this fact is put to use at the end of the strike, when the RF transmitter measurements are less reliable.

the evolution of the strike. So, the feed syrup input flowrate is controlled in the stage by the scheme shown in fig. 4.b, where the setpoint for mass concentration is given by a Brix vs. level curve that should be adjusted by the operator to reflect the changes in standard liquor purity. The considerable reduction of the heat transfer coefficient in the heat exchanger, taking place as the mass level rises, is compensated, in some degree, by modifying the setpoints of the calandria and of chamber steam pressure regulators. 5. Tightening Up: The purpose here is to increase the consistency of the massecuite in preparation for the discharge. There is no further introduction of syrup but the evaporation continues until the electric intensity consumed by the stirrer motor attains a configurable value. The setpoint of the calandria pressure controller is raised to accelerate the process. 6. Discharge: The heating steam input is shut down and the vacuum is broken in the chamber by opening the cleaning valve. When the pressure reaches an appropriate high value, the discharge output gates are opened. 7. Cleaning Up: Cleaning valve is kept opened with discharge gates closed to get rid of the traces of massecuite which remain contaminating the interior walls.

2.2.1. Batch crystallizer program Each batch crystallizer cycle follows a recipe implemented by sequential program (fig. 4.a) whose main stages are the following: 1. Loading: A high capacity valve is fully opened to start the introduction of standard liquor in the chamber. The objective is to load enough syrup so as to completely cover the calandria in such a way as to maximize the circulation process of the mass and improve the heat transfer coefficient. The two PID based loops controlling the pressures in the chamber and in the calandria are both put in automatic mode. 2. Concentration: The heating of the mass continues to concentrate the mass with the purpose of reaching the required supersaturation. The heat transfer coefficient and the steam consumption are both, at this moment, very high. The stage ends when the syrup Brix reaches a value, which in view of the most recent standard liquor purity report from the laboratory, would correspond to the right supersaturation. 3. Seeding: The seeding stage proceeds by automatically introducing the amount of seed magma which is considered adequate for obtaining the final correct average size. 4. Growing of the grain: This is the longest and most important stage of the cycle. The population of crystals in the seed should be made to grow as the sucrose in the solution migrates to their faces along the strike. The supersaturation would tend to decrease as the impurities accumulate in the solution, so standard liquor should be introduced in a controlled way to add the dissolved sucrose needed for compensating this effect. The amount of syrup to introduce would depend on the rate of evaporation, and on the purity currently existing in the pan. In the absence of on-line supersaturation estimation, the stage is conducted by establishing a curve of massecuite Brixes that gives, at each instant in the evolution of the process, the total mass concentration that should be enforced to obtain the right supersaturation. It should be noted, that the readings of the RF sensor takes into account not only the dissolved substances but also the mass of growing crystals. The level attained by the mass in the chamber, which should grow from the value initially loaded to the pan full capacity, is taken as a measure of

Figure 4: Crystallizer program. a) Main stages. b) Brix controlling loop in growing state In figures 5 and 6 the evolution along one cycle of the level of the mass and of the mass concentration of crystals are respectively shown, highlighting in each case the instant of activation of some important events.

215

being accessed by the trainee, the models should meticulously state all the involved mass and energy balances. In the case of the crystallizers the evolution of the mass of grains is tracked by means of the formalism of the population balance equations. The need to be able to simulate whole large factories determined the use of globalized models, wherever possible, to facilitate the numerical integration effort. There is an ample use of non dimensional relations used in the general chemical engineering literature and in sugar studies, and mass and energy transfer rates are put in relation with the characteristics of the processed streams. This provides reasonable starting values for the physically related parameters that can be further tuned to adapt to each specific case. The physico-chemical properties of the main products such as syrup or juice and massecuite had been taken from the existing specialized literature (Bubnik, Kadlec, Urban, and Bruhns, 1995). The characteristics of typical utilities such as liquid water and steam are readily available. The capacity of the generic model to faithfully reproduce real plant data and to meet the informed qualitative demands of sugar experts had been described elsewhere (Mazaeda, 2010; Merino, 2008; Mazaeda, Prada, Merino and Acebes, 2012). The assembled model here proposed does not exactly emulate any existing plant; but the deployed individual units are the ones that have been calibrated and validated with real data. The Sugar Benchmark has a very different purpose from the one that originally motivated the design of the underlying OO library. In any case, the model proposed, with its abundance of realistic details, full of special situations, which are needed for training, but that that would be considered as non-essential in almost any other type of application, would stand with respect to the designer of the overall control strategies, in a situation approximately similar to the one he/she would encounter when facing a real world problem. Detailed descriptions of the evaporator and the vacuum pan crystallizer models can be found in Merino, Alves and Acebes, (2005) and Mazaeda and Prada (2007) respectively. The centrifuges model has been specifically created for the benchmark. It is less involved that the previous models and has a static character. It simply consists of the necessary balances to each component, taking into consideration the dissolution provoked by the amount of water introduced. The exact composition of the poor and rich syrup streams and the humidity of the separated sugar grain product are decided by adjustable parameters.

Figure 5: Level evolution

Figure 6: Crystal content evolution 3. BENCHMARK MODEL The model has been created assembling objects instantiated from a previously existing library of sugar factory components using the Object Oriented (OO) concepts implemented by the EcosimPro modelling and simulation tool (ESA International, 2008). The original purpose of the library was to provide the elements for constructing realistic dynamic simulators specially dedicated to the training of the beet sugar factory control room operators. The library contains a representation of the main units to be found in the factory. It obviously include classes representing evaporators and batch evaporative crystallizers but it also hosts the auxiliary equipment needed in any process industry such as valves, pumps, tanks and even PID regulators (Merino, Acebes, Mazaeda and Prada, 2009). These ancillary elements are used to create the topology of the specific plant but are also deployed in the definition of the main process units to define its internal structure. All the models are derived from first principles and are coded in a generic way, exposing numerous parameters, making possible the adaptation of the resulting overall instantiated model to the specific situation at hand. The models exhibit a hybrid character. They are predominantly made of continuous time dynamical equations but must also respond correctly to discrete events fired, for example, as the program of the crystallizers move form on stage to the next. The original motivation as a training aid has permeated the modelling effort, guiding the election of the general assumptions of the library. Since all variables and actuators were in principle susceptible of

4. THE BENCHMARK CHALLENGE The benchmark main purpose is to serve as a testing platform to develop high level strategies with the capacity of guaranteeing an optimal economical behaviour simultaneously dealing with the management of the continuous processes and the scheduling of the batch units. The solutions proposed should be able to cope with the perturbations and uncertainties

216

x

represented by the variability of the working conditions and the lack of a complete knowledge of the state of the process. More specifically, the problem of conducting the whole plant in a smooth way is complicated due to the uncertainty in the batch crystallizers processing times which are very dependent on the characteristics of standard liquor. As an example, in figure 7, the effect that a modification of the feed purity, keeping the Brix vs. level curve fixed, has on the cycle time and on the evolution of other important variables, is shown. A smoothly managed plant should schedule the workings of the batch pans in such a way as to guarantee the consumption of all the syrup delivered by the upstream continuous section. Figure 8 shows the evolution of the levels in the feed syrup tank and in the strike receiver in this ideal situation.

x

Guarantee the processing of all the incoming syrup. The inventories of buffer units, strike receiver and, fundamentally, the standard liquor tank should be kept between safe limits. The quality standards of the produced white sugar should be met.

Figure 8: Level in buffer tanks over several cycles

Figure 7: Effect of feed syrup purity on the performance of a crystallizer. a) level. b) Sugar content. c) Brix of mother liquor. d) Average crystal size. e) Supersaturation. f) Purity of mother liquor Figure 9: Level in buffer tanks with problems

In figure 9, a decrease of the flow rate of thick syrup or an increase of purity could lead to the depicted situation, where the level in the feed syrup tank is clearly diminishing with the risk of violating the safety restriction on the required inventories. It should be noted that the availability of steam and the crystallizer cycle time are both very dependent on the Brix enforced at the output of the evaporation section. On the other hand, the amount of water to the centrifuges alters the concentration of feed syrup and also its purity. More formally, the objectives to be achieved in conducting the plant simulated in the benchmark are the following: x

Table 1: Allowed range for perturbation variables Parameter Description Range Wevap Mass flow rate into 10-16 kg/s evaporation Bmm_evap Brix of juice into 40-45 % evaporation Pmm_evap Purity of juice into 91-95 % evaporation W2mc Mass flows rate of 0.02-0.025 water to massecuite in centrifuges

Minimize the consumption of boil steam while serving the demands of the crystallizers.

217

Table 2: Degrees of freedom for plant-wide control Variable Description SPevap_Brix Setpoint of Brix control loop at Evaporation output Pboiler Pressure form boiler to Evaporation effect I LoadVP[k] k= 1-3. Load command for each vacuum pan Valvcent Valve controlling centrifuge throughput The objectives must be attained in the presence of perturbations, whose range of variation is shown in table 1. The minimum set of variables which are at the disposal of the design solution for the management of the plant are described in table 2. The variables that can be read form the benchmark are the same ones which are typically sensed in the real factory, namely: x x x x x x

Figure 9: Cycle time acceleration by increasing evaporation rate. a) Vacuum pressure setpoint. b) Calandria pressure setpoint. c) Supersaturation. d) Aggregated area of crystal population An example of the complex issues involved in applying the second, lower level strategy can be discussed analyzing the figure 9. The setpoints of the controllers of heating pressure to the calandria and of vacuum in the chamber influences the duration of the strike. So it would be legitimate to consider the use of these references as additional degrees of freedom for achieving the plant-wide objectives. It should be bore in mind, however, that an immoderate use of this kind of acceleration is somehow artificial and is limited by the purity of the feed liquor and so should be performed with caution. An unreasonable increase in the calandria pressure would imply an excessive increase of the supersaturation. The reason being, that as the crystallization kinetics is basically unaffected, the existing aggregated crystal area of the sugar population had not reached the value that would be able to sustain a flowrate of crystallization capable of compensating the new rhythm of water evaporation. As a consequence, supersaturation gets inside de labile region with the consequent prejudice to the quality of the product: a wider and reduced average size population. It goes without saying that the new setpoint pressure to fix would also depend of the availability of steam from the evaporation section. The benchmark gives a suitable platform for testing many other types of interesting applications like, for example, the hybrid identification of the discretecontinuous units, or the explicit handling of the uncertainty by embedding the numerical optimization procedure in a stochastic framework.

The levels of all evaporators, vacuum pans and tanks. The temperatures of all evaporators, vacuum pans and tanks. The Brixes of all syrups involved. The evolution of the Brix of the massecuite in each vacuum pan. The purity of the syrups involved. The values of all the pressures involved: in the chambers of the evaporators and vacuum pans and in the heat exchangers.

It is possible to consider the problem proposed at several levels. In the simpler approach, the crystallizers could be considered to be reasonably well controlled and the plant-wide coordinator should be simply concerned with guaranteeing the availability of steam and syrup as demanded. The third objective of keeping up with the quality requirements of the sugar product would be considered as automatically enforced by the pan's program if the steam demand is served. But it is also possible a more involved strategy, which deals directly with the control of each crystallizer. This would imply the handling of the values of the Brix vs. level curve, of the setpoints for the pressure of the calandria and of the chamber, among other details. Of course, this other approach would surely be able to achieve a more efficient solution, but would also imply a greater responsibility concerning the quality of the end product.

218

operators training. In: 19th European Modeling and Simulation Symposium (Simulation in Industry), EMSS 2007. Mazaeda, R., Prada, C. d., Merino, A., Acebes, L.F., 2011. Librería de modelos orientada a objetos para la simulación del Cuarto de azúcar: Cristalizador continuo por evaporación al vacío. Revista Iberoamericana de Control Automático, 8, 100-111. Merino, A., 2008. Librería de modelos del cuarto de remolacha de la industria azucarera para entrenamiento de operarios. Ph.D.Thesis, University of Valladolid. Merino, A., Acebes, L.F., Mazaeda, R., Prada, C.d., 2009. Modelado y simulación del proceso de producción de azúcar. Revista Iberoamericana de Control Automático, 6, 54-60. Merino, A., Alves, R., Acebes, L.F., 2005. A training simulator for the evaporation section of a beet sugar production process. In: European Simulation Multiconference. ESM05. Poel, P.V.d., Schiweck, H., Schwartz, T., 1998. Sugar Technology. Beet and Cane Sugar Manufacture. Dr. Albert Bartens. Prada, C. d., Sarabia, D., Cristea, S., Mazaeda, R., 2008. Plant-wide control of a hybrid process. International Journal of Adaptive Control and Signal Processing, 22, 124-141. Scattolini, R., 2009. Architectures of distributed and hierarchical Model Predictive Control-a review. Journal of process control, 19, 723-731. Zamarreño, J., 2010. Acceso a datos mediante OPC. Editorial Andavira.

Figure 10: HMI interface to the sugar benchmark Finally, it should be said that the simulated plant is going to be made available as an executable (Alves, Normey-Rico, Merino, Acebes and Prada , 2005) which implements the OPC protocol (Iwanitz and Lange, 2002; Zamarreño, 2010) and with a graphic interface (Alves, Normey-Rico, Merino, Acebes and Prada, 2006) making possible its standalone operation. In figure 10 a screenshot of the user interface is shown. The use of OPC will additionally facilitate the access to the simulated data from any of the many clients currently supporting that widely adopted standard. ACKNOWLEDGMENTS The research leading to these results has received funding from the European Union Seventh Framework Programme [FP7/2007-2013] under grant agreement n°257462 HYCON2 Network of excellence. REFERENCES Alves, R., Normey-Rico ,J., Merino, A., Acebes, L.F., Prada, C.d., 2006. Edusca (educational scada): Features and applications. In: 7th IFAC Symposiumin Advances in Control Education. Alves, R.A., Normey-Rico, J., Merino, A., Acebes, L.F., Prada, C.d., 2005. OPC based distributed real time simulation of complex continuous processes. Simulation Modelling Practiceand Theory, 13, 525-549. Bubnik, Z., Kadlec, P., Urban, D., Bruhns, M.,1995. Sugar Technologist Manual. Chemical and Physical Data Manufacturers and Users. Dr. Albert Bartens. ESA International, 2008. EcosimPro User Manual, EL Modelling Guide. EA International and ESA. Framework Program 7 Network of Excellence HYCON2, 2011. Highly-complex and networked con-trol systems. www.hycon2.eu. Iwanitz, F., Lange,J., 2002. OPC: fundamentals, implementation and application. Hüthig. Mazaeda, R., 2010.Librería de modelos del Cuarto de azúcar de la industria azucarera para entrenamiento de operarios. Ph.D. Thesis, University of Valladolid. Mazaeda, R., Prada, C. d., 2007. Dynamic simulation of a sucrose batch evaporative crystallizer for

219

CENTER TITLE HE DESIGNING AND IMPLEMENTING A MODEL TO EXAMINE R&D SECTION´S CAPABILITIES WITH EMPHASIS ON REVERSED ENGINEERING IN CHEMICAL FACTORY Neda Khadem Geraili(a), Mona Benhari(b) (a)

Department of Technology management,Faculty of Management & Economics,I.A.U Science & Research Branch,Tehran,Iran (b) Computer Engineering Graduated student (a)

[email protected], (b)[email protected],

mechanisms. In competitive conditions of the current industrial world, doing research and development activities is one of the most effective things that managers of economic enterprises can take on and in industrialized countries of the world, the costs allocated to these activities are being increased day by day, on the other hand referring to the research and development without creating the necessary infrastructure units such as: organizational structure, human resources characteristics and management of these organizations causes the failure of research and development process of economic institutions.(Kheradmand 2007) Process of R & D includes: basic research, applied research and development. By using basic research, the scientific findings are presented in the form of hypothesis, theories and general rules and in the next stage, it would be the applied research which determines the possible applications for basic research findings. The stage of development, scientific knowledge obtained from basic research and applied research is applied in order to provide new and developed products or processes and innovation can occur. The main centers of the R & D process consists of industrial research institutes, academic research institutions and government R & D centers. Reverse engineering is extraction and development of technical information from available products. This method unlike the direct designing process meant to production according to customer requirements and the initial idea is based on the engineering analysis. (Amiri and others 2008) Many activities in this area are done with different goals daily. Extraction of technical knowledge designed by R & D units in different industries, providing technical documentation of industrial equipment and probably copying the products are some examples in this field. Developing countries to access complicated technologies require a method that fills the technology gap between these countries and the developed countries at the right time and among different methods of having access to technology, reverse engineering is the most appropriate method.

ABSTRACT Today, research and development and related activities to access new technologies in the industrial world have been challenging activities. R&D units were known as Technology pillars and are unique resources of innovation, creating R&D units and institutes or developing old ones to new effective ones for developing countries are inevitable issues. There are various methods to access the Technologies and one of the most important ones, especially among developing countries is Reverse engineering which is Consciousness method taken from exist Technology. This paper also reviews research background of (R & D) and reverse engineering, implements simulation software modeling to achieve the most effective parameters and their relationships and offer strategies for optimal per unit R & D and access to technologies of modern examined. This study contains R & D department of a chemical factory as selected population which is one of the important industrial factories in Iran. In conclusion some key effective factors have been extracted through this simulation. The company can consider them to develop R&D unit and improve the product quality. Keywords: Research and Development (R&D), Reverse engineering, Technology Transfer, System dynamic. 1. INTRODUCTION Research, development and management are the pillars of creation, development and utilization of technology and they are the needs of economic and social development in every country, in the developing countries which are seeking industrial and economic self-reliance, policy and planning research programs and development and management of practices impact on these activities are considered as priorities for national, industrial and manufacturing activities. Expansion and enhance of R & D activities, particularly "industrial research ", requires understanding the factors affecting research and development process and designing policies and effective activities of such

220

This research was done in the context of systemic thought and to achieve research objectives and answer of research questions, System Dynamics modeling has been used. Modeling dynamic systems is an essential tool in systemic thought that in order to better understand the behavior of the system; imaging and strengthening the conceptive models and displaying the behavior of system are used. Since diversity of effective factors is very significant on issues, in order to obtain acceptable solutions, the least important factors are omitted by using modeling in the study. The general category type of research will also be an applied research. This study utilizes a model (simulation), which includes relations with variables methods in reverse engineering in R & D units which will be done. This model can be considered as a new process in popular and applied industries. And the places where there is no new scientific and systematic approach in these projects, so this research and simulations can greatly help managers in making decisions and having reasonable controls. Thus, at first there have been gathered interviews and questionnaires with key members collaboration and make a systematic relation between variables and suitable parameters base on model and achieved the desired data will be collected and model validation is done and CLD related to basic parameters can be drawn. Then modeling using Vensim simulation software based on raw data and descriptive statistics are done and the parameter sensitivity analyses also take place. Finally, based on the model output and parameters sensitivity analysis software, necessary suggestions and performed conclusions are provided.

2.1. Subheadings Initial Caps, bold, flush left. Use Times New Roman Font and 10 points in size. Start the text on the next line. Please use the “Heading 2” style. 2.1.1. Secondary Subheadings Initial Caps, bold, indented of 0.7 cm. Use Times New Roman Font and 10 points in size. Start the text on the next time. Please use the “Heading 3” style. 3.

SELECTING APPROPRIATE TECHNOLOGY AND REASERCH AND DEVELOPMENT REQUIRED ACTIVITIES

The process of selecting technology and appropriate product includes all actions and activities which are contiguous with the objectives, conditions and specifications and technological needs to determine the most appropriate technology requirements and also the most appropriate strategy to achieve the goals is done by considering the circumstances and technical, economic and legal relations. (Akhbari and others 2008), (Allahyari 2009), (Ebrahimi), (Fort collins and kaufman 1989) This process typically includes steps such follows: 1-Information on market needs consumer preferences and market traction for new products. 2-Information about status of competition, situation of required innovation, research and development activities, products and environmental threats. 3-Having Information about the required global situation of technology and explanations of their pattern in some period of times by using technology forecasting techniques. 4-Identification of technical and scientific facilities of country and possibility of access to materials, energy, and production processes. 5- Analysis of investment company status among others in terms of scientific, technical, economic support and technical ability, communication and marketing. 6-Strategic planning to determine and select the required technology, according to data collected. 7-Designing strategies to achieve product and selecting the most appropriate strategy (according to the results of technical and economic feasibility studies later we will pay attention to this issue.) Marketing efforts, policy and strategic planning and selecting the required technology to achieve the set of activities that make up except those involved with research and development activities and research and development, so reaching them is not possible except with effective management of information and technical experiences of engineers. The sensitivity of this action is so much in a way if investments policy and managements not to be done accurately and according to practical techniques and if a comprehensive analysis not to be done, the rest innovation activities will be overwhelmed and they eventually might lead to investment plan failure.

2.

MAJOR RESEARCH AND DEVELOPMENT ACTIVITIES IN THE STAGES OF REVERSE ENGINEERING PROCESS The experiences of advanced countries suggests that technological progress in these countries owe R & D activities and technological infrastructure more than anything and without implementing such activities there is no way to achieve the desired technology for that, though by buying the product or purchase process, the required parts can be obtained, but even successful technology business by buying technology (research and development activities necessary for building or research strategies to produce copy that is necessary) also requires research activities and industrial engineering services. For example, before attempting to determine how to access to technology, knowledge and action for defining technology needed to meet demands and needs of development projects such countries need some activities which won’t be fulfilled without collaboration of involved ones engineering services directly.(Colin Bradley 1998), (Jokar 2008), (Book of technology comercialization)

221

considered with an approach of reverse engineering in order to strengthen R & D unit.(Fort collins and kaufman 1989),(Roussel and saad 1991),(Chiesa and masella 1996) First stage: gathering effective factors on research and determining the primary assumptions Second stage: determining the most important factors affecting research and inserting them in the initial model and draw CLD to determine relationships between factors and variables for the basic model. Third stage: extending the primary model and finding datum and formulas for model and model validation Fourth stage: testing Software Model (simulated). (Sterman 2000),( www.system dynamics.org/conferences 1998)

4.

SELECTING APPROPRIATE TECHNOLOGY AND REASERCH AND DEVELOPMENT REQUIRED ACTIVITIES (Ghani 2009)], (Houshangnia 2009), (Laghvi 2010),(Mardi 2011) 1 - Controlling the possible technical and economic studies, making plans to copy a subset of product or product requirements inside the factory. 2 - Researching for understanding mechanisms of functioning components for realization of working mission for products and discovering the relationships between components during operation. 3 - Departing the components that we have decided to copy them and registering charts for the products and its details during operation. 4 - Providing maps and photos, or piece or pieces that they're going to be copied. 5 - Identifying, testing and determining the raw materials used to manufacture any component of product. 6 - Identifying, testing and diagnosis process for production of any component of product. 7 - Identifying, testing and diagnosis of complementary operation performed to achieve mechanical, physical, metallurgical, chemical properties… which are required. 8 - Detecting required machineries and tools for manufacturing and assembly. 9 - Compiling technical knowledge for production and assembly and its components. 10 - Preparing production drawings, maps, templates, models and tools needed for production and control stages. 11 - Monitoring the prototyping operation specifically or under direct supervision. 12 - Controlling samples made and matching them with the profile of desired standards. 13 - Assessing and analyzing test results and in case of necessary, revision of sizes and tolerances. 14 - Designing production and estimating required machinery and equipment. 15 - Planning of factory and designing factory production and assembly line.

5.1. . Introducing the basic components in capability of research and development (R&D) of the chemical Factory: (Fadaei 2010), (Fort collins and kaufman 1989),( Chiesa and masella 1996) This section is based on offering presumptions that simply realize modeling process and it is attempted to present the model which can be implemented in Vensim software. Assumptions made are extracted based on the information of a Chemical factory and interviews with senior managers and library studies and the most effective variables and essential factors are mentioned bellow: 1 - Professional and skilled personnel (manpower) 2 - Identification of risk factors 3 - Market and customers needs 4 - If necessary interaction with industries and marketing and production units 5 - System inputs include raw and new materials (new requirements) 6 - Technical Knowledge 7 - Sufficient capital (cash flow) 8 - Cost market 9 - Planning 10 - Economic blockade 11 - Project Manager 12 - Skillful and risky management of R & D

5. METHODS This paper aims to achieve a model to evaluate the capabilities and competencies of R & D of a chemical factory based on modeling system. This model is provided to managers and decision makers as a tool and it plays the role of easing in decision process and opting organizational strategies. So according to this view, this article is considered in the category of applied research. On the other hand the type of approach used in the paper is in a way that by dynamic modeling, the effective components on planning present a new model and it solves the problems that current planning methods are facing with. In this study it is tried that modeling (simulation) and all factors and important variables and influencing patterns, trends empowerment of R & D unit for the chemical industry to be

5.2. Model validation test In order to demonstrate the model validity, we use real tests K Square test to explain credibility. Since based on calculations done on questionnaires and interviews, the most effective factors and variables in the model were determined, and information relevant to the years 2001 to 2011 are available. So by putting this information in the model, the information of next year is predicted and they are measured with realities so the credit degree of model could be provided. It should be mentioned that information and figures of influential factors, including system inputs (raw materials, equipment, technical knowledge), human resources, including (motivation, working

222

group, spontaneous, education, good environment, planning and project control, management empowerment and the number of projects done in R & D through the questionnaire were determined. Also system inputs, planning and control of projects, skillful management of Research and Development and human resources were considered as independent variables and the numbers of projects R & D were considered as the dependent variables. The aim is that according to the four-term factors: 1) system inputs, 2) Planning and Project control, 3) strong management, 4) Human Resources with a number of projects to reach an acceptable quality. (Kheradmand 2007), (Rabelo 2004),( Sterman 2000)

According to figures contained in the questionnaire, the average number of variables and factors extracted from 2001-2011, they have been mentioned in above table and then for credibility and accuracy of model operation, the K Square test was measured by the following formula: X2 (n-1) = And the figures of 2 X was extracted according to the above table; 2 X which was provided according to K Square test is 2X 0.05,9 =16.92 and since 2X is four variables of model and is less than 16.92, so the modeling assumption is accurate and according to the figures and information listed above, the model is accepted with 0.95 confidence.

Table1:Model validation test by using of real data test & square of KAY :

Hum an Reso urces

Strong manag ement of R&D unit

Plan ning & Proj ect Cont rol

Syst em Inp ut

6.45

5.6

7.75

6

Num ber of ende d proj ect in R& D unit 2

6.75

5.92

7

6.16

5

6.8

6.08

7.75

6.4

7

6.55

6

8

6.75

2

6.45

5.92

7.5

6.66

1

6.4

5.33

7.25

6.5

1

6.2

5.41

7.25

6.33

3

5.9

5.5

6.75

6.41

0

5.6

5.33

6.75

6.08

0

4.95

5.1

6.5

5.91

7

7.6

7

8

7.6

5

7.6

2.949

7

2.872

8

0.983

7.6

2.23 7

5

X2 (n2 1) =x

5.3. C.L.D draw After extracting the effective parameters on model, in the second step of modeling, the causal relations between the parameters must be considered, so for this purpose, we extract C.L.D. It is noteworthy that due to the volume of datum and the high number of parameters affecting the issue, single-step extraction of C.L.D is not feasible.therefore for the purpose of understanding it easier, we divided the casual model of the issue into five stages and in each stage a loop of causal model has been drawn; and then finally according to the defined relations, we link each stage to the other stages and each component to the other components to achieve the ultimate causal model of the problem. Finally, the effective parameters of the model simulation, sensitivity analysis and the related results are also presented.

Yea r

200 1 200 2 200 3 200 4 200 5 200 6 200 7 200 8 200 9 201 0 201 1 )fit ( Mo del EIX2

Proje ct time

Inspectio n& raising of obstacles

Data (new needs)

Skillful manager of r&d

Human resource

Price of end product

Planning & project control

Enough capital

Input system

Techni cal Knowl edge

Market need Relatio n of R&D with market ing

Quality of manageme nt of R&D

Exami nation of equip ment

To find risk factors Econo mic blocka de

9,0/05

Data (new needs) Project numbe

Chart 1: CLD(casual loop diagram),(Model)

223

(Amiri,farzad and others 2008), (Colin Bradley 1998),( Mardi 2011)

5.5. Multiple Graphs relevant to the main parameters of the model

and Input and Human Resource and Plan

5.4. Draw of chart of stock &flow by vensim software: (Craig W.kirkwood 1998), (Rabelo and Helal 2004) Formula tray relation of Regression for Number of ended project in R&D unit: Number = _0.61 Input+0.07 Planning control+0.72 Managenent of R&D_0.34 Human Resources. Multiple Graphs relevant to the main parameters of the model (Kheradmand 2007), (Rabelo 2004)

8

4

5.5

3 2

4 4 3 5 2 5 3 2 5

4 3 5

4 3 5

2 2

3 0.5

1

4 3 2

5 3 4 5 3 4 5 3 4 2 2 2

1 1

1

1 1 -2 1 1 1 2011201220132014201520162017201820192020 Time (Year) Stock & Flow Diagram_R&D ability Model

Planning & Control Project rate

Number : Current 1 1 1 1 1 1 1 Input : Current 2 2 2 2 2 2 2 2 Human Resource : 3Current 3 3 3 3 3 3 "Planning & Control Project" 4 : Current 4 4 4 4 4 "Management of R&D" 5 : Current 5 5 5 5 5 5

Planning & Control Project

chart3: Multiple Charts of the basic model variable 5.6. Statistical analysis (sensitivity analysis): (Houshangnia 2009), (Rabelo 2004),( Sterman 2000)

Number

Input Input rate

Number 200

150

100

Performance of Human Resource rate

Human Resource 50

Management of R&D

0 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 Time (Year)

Management rate

Number : Current

Chart 2: stack and flow

Input 6

3

0

-3

-6 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 Time (Year) Input : Current

224

Year

Input

8.5

7

5.5

2005

2007 2009 2011 2013 Time (Year)

2015

2017 2019

Human Resource 8

Input : Current

Year

Optimistic state for system input

Project numbers in optimistic conditions 2011 0.6 -1.3 -6.6 2020 -4.2 -7.8 15.23 Chart5: Pessimistic level conditions for system input, Coefficient change from -0.61 to -0.80

10

4 2001 2003

Primary state for system input

Primary state for system input

5.995

Optimistic state for system input

Project numbers in optimistic conditions 2011 0.6 7.7 53.8 2020 -4.2 9.23 142.4 Chart4: Optimistic level conditions for System Input, Coefficient change from -0.61 to +0.1

3.99

1.985

-0.02 2001

2003

2005

2007

2009 2011 2013 Time (Year)

2015

2017

2019

2015

2017

2019

Human Resource : Current

Human Resource 20

16.5

Input 13

8

9.5

4 6 2001

0

2003

2005

2007

2009 2011 2013 Time (Year)

Human Resource : Current

-4

-8 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 Time (Year)

Year

Primary state for human resource

Optimistic condition for human resource

2011 2020

3.05 -0.01

9.4 12.15

Input : Current

Number 20

10

Project numbers in optimistic conditions 60.9 171.8

Chart 6: Optimistic level conditions for Human Resource, Coefficient change from -0.34 to +0.3

0

-10

Human Resource 8

-20 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 Time (Year)

4

Number : Current 0

Number 200

-4

150

-8 2001

2003

2005

2007

100

Human Resource : Current 50

0 2001

2003

Number : Current

2005

2007

2009 2011 Time (Year)

2013

2015

2017

2019

225

2009 2011 2013 Time (Year)

2015

2017

2019

Project numbers in optimistic conditions 2011 3.3 4.6 10.99 2020 1.23 3.69 46. 63 Chart8: Optimistic level conditions for Management of R&D, Coefficient change from 0.72 to 0.85 year

Number 4

0.5

-3

-6.5

-10 2001

Optimistic state of R&D

Primary state of R&D

Management of R&D 2003

2005

2007

2009 2011 2013 Time (Year)

2015

2017

6

2019

Number : Current

3

Year

Primary state for human resource

Optimistic condition for human resource

Project numbers in optimistic conditions 2011 3.05 1.45 -9.36 2020 -0.01 -3.05 2.69 Chart7: Pessimistic level conditions for Human Resource, Coefficient change from -0.34 to -0.5

0

-3

-6 2001

2003

2005

2007

2009 2011 2013 Time (Year)

2015

2017

2019

"Management of R&D" : Current

Number

Management of R&D

2

6 -3.5

4.5 -9

3 -14.5

1.5 -20 2001 2003

0 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 Time (Year)

2005

2007

2009 2011 2013 Time (Year)

2015

2017

2019

Number : Current

year

Primary state for R&D Management

Pessimistic conditions for R&D management

Project numbers in pessimistic conditions 2011 3.3 1.6 -11.88 2020 1.23 -2 -8.3 Chart9: Pessimistic level conditions for Management of R&D, Coefficient change from 0.72 to 0.55

"Management of R&D" : Current

Management of R&D 6

5

4

3

Planning & Control Project

2 2001

8 2003

2005

2007

2009 2011 2013 Time (Year)

2015

2017

2019

4

"Management of R&D" : Current

0

Number 60

-4 45

-8 2001

30

2003

2005

2007

2009 2011 2013 Time (Year)

"Planning & Control Project" : Current 15

0 2001

2003

2005

2007

2009 2011 2013 Time (Year)

2015

2017

2019

Number : Current

226

2015

2017

2019

Planning & Control Project

Number

8

40

5.5

29

3

18

0.5

7

-2 2001

2003

2005

2007

2009 2011 2013 Time (Year)

2015

2017

-4 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 Time (Year)

2019

"Planning & Control Project" : Current

Number : Current

Number

Year

Primary Pessimistic Project state for state for numbers in planning & planning & pessimistic project project conditions control of control of R&D R&D 2011 2.39 2.05 -0.7 2020 -2.5 -3.08 20.29 Chart11: Pessimistic level conditions for Planning & Project control of R&D, Coefficient change from 0.07 to 0.04

20

14.5

9

3.5

-2 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 Time (Year) Number : Current

6. CONCLUSIONS As noted, the purpose of this paper is presenting a local model to consider capabilities and abilities of Research and Development unit by emphasizing on reverse engineering in a Chemical Factory. Based on the pattern provided, indicators and parameters to measure Features and capabilities of R & D unit are available. In this model, some of the basic parameters which were measured, analyzed based on sensitivity analysis include: Number of Projects done in R & D unit, system inputs such as raw materials and new requirements, professional and skilled manpower, skilled management of R & D, Project Planning and Control of research and development project in analyzing the sensitivity of above indices and parameters, some results were obtained. By eliminating restrictions of indices, it means in terms of optimism for the variables, we can mostly observe that growth is very rapid rate for parameters and variables and also features R & D unit and ultimately the growth of parameters lead to significant increase in the rate of the number of projects conducted by R & D unit. Also by considering the conditions of reverse condition it means, the sensitivity test of variables in terms of increase of barriers in the parameters, we can mostly witness decrease of parameters and number of projects done in R & D. Finally by reviewing the outputs of model software graphs and sensitivity test of parameters and effective factors for model, in this article the capabilities and abilities of Research and development (R & D) of chemical factory have been determined. We can use these indices and variables to achieve specific goals specific and by comparing them relatively with other industries of R & D and or by comparing them with

Year

Primary Optimistic Project state for state for numbers in planning & planning & optimistic project project conditions control control 2011 2.39 2.65 0.99 2020 -2.5 -1.9 19.05 Chart10: Optimistic level conditions for Planning & Project control of R&D, Coefficient change from 0.07 to 0.10 Planning & Control Project 8

4

0

-4

-8 2001

2003

2005

2007

2009 2011 2013 Time (Year)

2015

2017

2019

"Planning & Control Project" : Current

227

x

research and development units of industrial plants located in other countries, the efficiency and effectiveness of R & D unit in the chemical factory to be measured. Finally, this model can be used to identify desired features and capabilities of R & D and to use required managements to reach them and they be used to successfully control the programs. The model has the ability to generalize and use in most application areas for selection of strategies, policies and research and development projects and access to new technology and we can observe the basic parameters of the model by its effect on the capabilities and abilities of research and development unit. By observing all tables and charts of the basic parameters we can conclude that in spite of reduced volatility variables and factors in the model during the years 2001-2010, the increase of projects number in R & D unit of chemical factory has had a good condition. This increase is relevant to the urgency and necessity of unfinished projects which are being done by the important factor of R & D Management. According to the relations of all the basic parameters and doing regression and the corresponding formula model, we can conclude that the most effective factor on the number of projects is the factor of management, planning and control of projects and R & D unit of the chemical factory has a good condition in terms of R & D management, and the reason of this issue in addition to observing relations of variables and formulas coefficients, in addition to the information and graphs is due to necessity and emergency point of unfinished projects and coordination of the main parameters by skillful of Research and development (R & D).

x x x

x x x x

x x x

x

REFERENCES x Akhbari,mohsen and others,2008,article of process of developmevt of new products,tadbir journal number 184 x Allahyari,parinaz,2009, designing a model for organizational R&D units of generator companies, MS thesis. x Amiri,farzad and others,2008, article of john doe reel of development of new product in environment of generation of globalization level, third international conference R&D x Colin Bradley,1998, The application of reverse engineering in rapid product development, ISI Technical paper x Craig W.Kirkwood,1998, System Dynamics Methods: A Quick introduction, college of Business Arizona State University x Ebrahimi,abdolhamid,conference of development of new products, http://www.irmmc.com/index1.htm x Fadaei,marjan, 2010,examination of effects of R&D unit on efficiency of industries, MS thesis

x

228

Fort Collins,H.R.Kaufman,1989,Book of R&D Tactics Ghani,asgar,2009, application of reverse engineering in achieve to technology of wrapped products, MS thesis Houshangnia,amir pasha,2009, simulation of role of R&D in development process, MS thesis Jokar,mohamad sadegh,2008, article of processes of structural for development of new product,(models of innovation in creation technology) Kheradmand,kamran,2007, examination effect of R&D on profitable of industries in Iran, MS thesis Laghvi,reza,2010, designing of R&D system and execution of it in weaving factories, MS thesis Mardi,asghar,2011, designing of system structure for R&D units in generator companies, MS thesis P.A.Roussel K.N.saad ,T.J.Erickson,1991,Book of Third Generatino on R&D.Harvard Bussiness school press , Boston Rabelo,L. and Helal,M.2004,Analysis of Supply chains using system dynamics, Neural Nets and Eigenvalues Sterman,j,2000, Bussiness DynamicsModeling & system thinking for a complex world, McGraw Hill, New York,USA BOOK of Technology comercializationi, the 5Stage R&D Commercialization Process http://www.1000ventures.com/technology_tran sfer/tech_commercialization_main.html. Vittorio chiesa,Christina Masella,1996,Paper of Searching for an effective measure of R&D performance, ISI Technical paper www.systemdynamics.org/conferences/1998/P ROCEED/abstracts.pdf

ON THE USE OF MINIMUM-BIAS COMPUTER EXPERIMENTAL DESIGNS Husam Hamad Electronic Engineering Department Hijjawi College of Engineering Technology Yarmouk University, Jordan [email protected]

Space-filling designs such as the Latin hypercube designs (Mckay, Beckman, and Conover 1979) aim at uniformly scattering the points over the design variables space. Different system response complexities require different metamodel types in order to adequately accommodate the underlying behavior and reduce bias errors. Hence, different metamodel types exist depending on the underlying response. Response surface models and kriging metamodel types receive much coverage in the current literature on the design and analysis of computer experiments. Other types, considered to be equally competitive in current usage according to (Simpson et al. 2008) include multivariate adaptive splines, radial basis functions, neural networks, and support vector regressors. The study in (Chen et al. 2006) concludes that no one metamodel type stands out. A similar conclusion is made in (Wang 2003), stating that no one metamodel type is definitely superior to others. According to (Goel et. al 2007), the consensus among researchers is that no single metamodel type can be considered the most effective for all responses. Nonetheless, (Wang 2003) also concludes that kriging and second-order polynomial response surfaces are the most intensively investigated metamodels. Based on a Google Scholar search, the work in (Simpson et al. 2008) concludes that response surface models are the favorite methods in structural optimization disciplines. While (Viana and Haftka 2008) conclude in a Google Scholar search that the distinction between metamodels diminished after an initial popularity for response surface and artificial neural networks techniques; they nonetheless acknowledge that response surface models are the favorite techniques in structural optimization. The review in (Chen et. al, 2006) on the design and modeling of computer experiments investigated experimental design methods and their relation to the various types of metamodels used in computer experiments. The review presented conclusions from attempts by many researches to determine the most appropriate experimental design for the selected metamodel type. Based on their own computational study tests on the available options, (Chen et. al, 2006) conclude that response surface model designs such as

ABSTRACT Computer experimental designs are used to generate data in metamodeling of multiresponse engineering systems. Metamodels, which are also called surrogate models, offer more efficient prediction of system responses but add errors when used as surrogates for the simulators. Error sizes depend on computer experimental designs. Only bias errors are incurred in deterministic computer experiments; however, the majority of experiments reported in the literature are not optimized for minimum bias. Box and Draper–the pioneers of the response surface methodology– originated the work on minimum bias designs in the late 1950's. Space-filling designs such as the Latin hypercubes are mainly in current use; sometimes even in response surface models. This work is a practical study via a number of analytical and electronic circuit examples on the use of minimum bias designs for response surface metamodels. Some minimum bias designs in hypercuboidal spaces are also introduced. Keywords: Experimental design, minimum bias design, Latin hypercube design, response surface models 1. INTRODUCTION Computer experimental designs are sampling techniques used to determine combinations of design variables to generate metamodels (also known as surrogate models) of complex engineering systems responses. Different sampling techniques are used to generate metamodels using simulation output for system responses for points in the experimental design. For deterministic simulations, errors introduced by the metamodels are systematic, or bias, errors caused by the deficiency of the metamodel in truly representing the response. Contrary to data in practical experiments, no variance-related error components are present in computer experiments. Experimental designs can be categorized into classical designs and the more recent space-filling designs (Chen et. al 2006). Classical designs such as factorial designs (Myers and Montgomery 1995) and central composite designs (Box and Wilson 1951) are primarily used in response surface modeling methods.

229

x

the central composite designs and the Box-Behnken designs are "good only" for response surface models, while all other experimental designs (space-filling designs such as the Latin hypercube samples) are appropriate for all metamodels other than the response surface models. It is noteworthy to mention here that minimum bias designs were not included in the review by (Chen et. al 2006). A Google Scholar search similar to those in (Simpson et al. 2008) and (Viana and Hafka 2008) was conducted in this work in April, 2011. The results are shown in Table 1 for response surface models and in Table 2 for kriging metamodels. Table 1: Search Results Related to Response Surface Models Using Google Scholar Number of Publications Search Phrase 200020052011 2011 approximation OR metamodel OR 12800 9090 surrogate AND "response surface" "experimental design" AND 15400 12400 "response surface" "minimum bias design" AND 24 14 "response surface" "Latin hypercube" AND 2210 1720 "response surface"

There are two main objectives for the work presented in this paper: x

x Table 2: Search Results Related to Kriging Metamodels Using Google Scholar Number of Publications Search Phrase 200020052011 2011 approximation OR metamodel OR 9360 6790 surrogate AND "kriging" "experimental design" AND 1830 1340 "kriging" "Latin hypercube" AND "kriging" 1370 1130

To show that minimum bias computer experimental designs can potentially give more accurate response surface models than the widely used space-filling designs of comparable size. This is demonstrated via analytical functions and electronic circuits. To introduce some minimum bias computer experimental designs for hypercuboidal spaces of dimensions 2 to 6.

The remainder of this paper is organized as follows: section 2 demonstrates through analytical examples the motives for using minimum bias designs. Section 3 deals with error types due to variance and bias, presenting basis which are subsequently applied to cuboidal design spaces to construct minimum bias designs. Some of these designs are then used in the electronic circuit examples of section 4. Conclusions are given in section 5.

The results in Tables 1-2 lead to the following general possible interpretations with regard to experimental designs and metamodeling methods: x

From the other tables entries, of the 15,400 papers since 2000 having the phrase "response surface" AND "experimental design", only 24 papers mention "minimum bias designs" while 2,210 papers talk about "Latin hypercube" designs. What are the reasons for the unpopularity of minimum bias computer designs? During the times minimum bias designs were presented in articles in the late 1950's and early 1960's (Box and Draper 1959; Draper and Lawrence 1965), experiments were conducted in the laboratories, and hence the reasons for avoiding large experimental designs are obvious. However, the recent space-filling designs used in computer experiments of today can have larger sizes than most of minimum bias designs, so reasons attributed to size for ignoring these designs are ruled out.

2. MOTIVATION Sample points in a minimum bias experimental design are located in the design region such that the design's moments satisfy certain conditions as outlined in the next section. In this section, analytic examples are used to demonstrate the superiority vis-à-vis prediction accuracy of metamodels based on minimum bias designs (MBD) in comparison to models derived using other experimental designs such as the Latin hypercube (LHC) designs. Figure 1 shows four experimental designs used to derive a first-order response surface for the response given in Equation (1):

Response surface models of the 1950's still compete with the more recent metamodels such as the kriging type. As seen in Table 1, the number of publications with "response surface" in combination with any of the words approximation, metamodel, or surrogate since 2000 is about 12,800. The majority of these publications (9,090) appeared in the last half of the last decade from 2005 to 2011. The corresponding statistics for "kriging" metamodels are 9,360 for the period 20002011, with 6,790 of these publications appearing in the period 2005-2011.

y 5  2 x1  x 2  0.5 x1  3x1 x 2  x 2 ; x  [-1,1] (1) 2

230

2

1 x2

this is not always the case. To illustrate, the above metamodeling activities are repeated for the response in Equation (2) (see Figure (3) for function plot):

1 x2

0

0

9

-1 -1

0 x1 (a)

-1 -1

1

1 x2

0 x1 (b)

1

x2

0 x1 (c)

3.2446 u 10 6 , 3.5757 u 10

a9

0

0 x1 (d)

(2)

where a1 659.23 , a2 190.22 , a3 17.802 , a4 0.82691 , a5 0.01885 , a6 0.0003463 , a7

-1 -1

1

; x  [905,995]

11

a8

1.6606 u 108 ,

and

.

200

1

(a) 1 00

y

-1 -1

i 1

i

i 1

1

0

¦ a x  900

y

0

Figure 1: Experimental Designs (a) MBD 1 (b) MBD 2 (c) FAC (d) LHC

-1 00 900

Two MBDs are shown in parts (a) and (b); part (c) depicts the standard response surface model design known as factorial design (FAC), and a LHC design is shown in part (d). Metamodels built using these designs are validated using a 21x21 sample. See Table 3.

940

960

980

1 000

x 150

(b) y

100

Table 3: Validation Results Corresponding to the Four Experimental Designs in Figure 1 Experimental Design Figure 1 Part RMSE MBD 1 a 1.160 MBD 2 b 1.160 FAC c 1.243 LHC d 1.381

50

0 915

920

925

930

935

940

945

x

Figure 3: Function Plot for y (x ) in Equation (2) for: (a) x  [905,995] (b) x  [915,945]

As shown in the table, the lowest root mean square error (RMSE) is obtained using any of the two MBDs. To demonstrate the relation between RMSEs for MBDs and LHCs, 100 metamodels are fitted using 100 different LHC samples. RMSEs for these metamodels are compared to the RMSE obtained if a MBD is used; see Figure 2. In the figure, the RMSEs shown are normalized to the RMSE for the MBD (the dotted line at 1.0).

In Figure (4) RMSEs for 100 different metamodels built using 100 different LHC samples are compared to the RMSE for the metamodel derived using a MBD (each LHC sample has the same size as the MBD). 1.5

Normalized RMSE

1.5

Normaized RMSE

920

1

1

MBD

MBD 0.5

0.5

10

20

30

40 50 60 LHC Sample No.

70

80

90

10

20

30

40 50 60 LHC Sample No.

70

80

90

100

Figure 4: Normalized RMSEs for 100 LHC Designs for the Function in Figure 3(a).

100

Figure 2: Normalized RMSEs for 100 LHC Designs. The results depicted in the figure clearly demonstrate the superiority of MBDs. Unfortunately,

The figure shows that RMSE for the MBD case is not always lower (a few points are below the dotted line

231

at 1.0 corresponding to the normalized RMSE for the MBD). However, for most of the 100 LHC samples, their RMSEs are worse by comparison to the MBD sample. The reason for the discrepancy between results of the similar metamodeling activities summarized by Figures 2 and 4 are attributed to the underlying response being modeled. As it will be shown, MBDs result in least errors provided the underlying assumptions for deriving MBDs are satisfied. Usually, the derivation assumes that the complexity of the response is such that it is higher than the response surface metamodel that fits it by one order; e.g., the response follows a third-order polynomial if the metamodel fitted is a second-order polynomial. Obviously, as the design variables space narrows down, such assumption about orders becomes more valid (see Figure 3). This is demonstrated in Figure 5, which is similar to Figure 4 except now the design space for the response is narrowed down to x  [915,945] from x  [905,995] in Equation (2).

conditions for MBD derivation in terms of the so-called design moments. (Draper and Lawrence 1965) applied the above mathematical conditions in (Box and Draper 1959) to derive MBDs for cuboidal regions. They used parameterized experimental design sets to build first and second-order MBDs. However, many of the tabulated results involve sets with parameters outside the assumed coded design space boundaries. This may be inappropriate in many practical engineering system design problems; for example, negative transistor widths cannot be implemented in practice. Our work (also for cuboidal design spaces) involves the parameterized experimental design sets mentioned shortly later on in this section. However, solutions for the parameters resulting in practical MBD sets (i.e., with none of the parameters outside the design space) are taken when the mathematical conditions related to design moments are applied. The sets are used to construct second and third-order MBDs. Consider a k -dimensional space with design (input) variables x1 , x 2 ,..., x k . It is assumed that the space is coded such that 1 d x1 , x 2 ,..., x k d 1 . Second and third-order MBDs in our work are constructed using combinations of the following sets: C (0 k ) ,

Normalized RMSE

1.5

1

F (D k ) , and S (D a , E k  a ) . Explanation for this notation is provided in Table 4.

MBD

0.5

10

20

30

40 50 60 LHC Sample No.

70

80

90

100

Figure 5: Validation Results for the Function in Figure 3(b).

Notation

3. MINIMUM BIAS DESIGNS There are two sources for errors in metamodels: (i) noise in the experimental design data used to fit the metamodel; and (ii) inadequacy of the metamodel. Accordingly, errors are categorized as: (i) variance, and (ii) bias errors, respectively. In practical experiments, variance errors are the assumed error source while bias errors are the only source of errors in computer experiments. Standard response surface designs such as the central composite designs are derived ignoring bias errors; i.e., derivations in this case assume that the fitted metamodel adequately represents the response. In minimum bias designs, however, it is customary to assume that the true response is a one-order higher polynomial than the metamodel. Thus, if the metamodel is a second-order polynomial then the MBD is derived assuming a third-order polynomial response. There is no intention in this paper to provide rigorous mathematical treatment for MBD derivations. Such derivations originated in the pioneering work by Box and Drapper in 1950's (Box and Draper 1959), with more recent treatment in (Goel et. al 2008) and (Abdelbasit and Butler 2006). The results are presented in terms of satisfying the necessary and sufficient

F (D k )

C (0 ) k

Table 4: Notation Used Meaning #points a design point at 1 the center

Notes x1

xk

factorial design

2k

See Table 5 for k 3

All k permutations of factorial designs with S (D a , E k  a ) a variables at D and k  a variables at E

k2 k

See Table 6 for k 3

Table 5: F (D k ) Factorial Design for k x1

rD

x2

rD

x3

rD

D

D

D

D

D

D

D D D D D D

D D D D D D D

D

D D D D

232

...

3

0

(a)

Table 6: S (D , E

k a

a

x1

rD

x2

rE

) Design for k

rE

x3

x1

rE

x2

rD rE

x3

M2

3 with a 1 x1

rE

x2

rE rD

x3

x1

x2

x3

x1

x2

x3

x1

x2

x3

D

E

E

E

D

E

E

E

D

D

E

E

E

D

E

E

E

D

D

E

E

E

D

E

E

E

D

D

E

E

E

D

E

E

E

D

D

E

E

E

D

E

E

E

D

D

E

E

E

D

E

E

E

D

D

E

E

E

D

E

E

E

D

D

E

E

E

D

E

E

E

D

(b)

R3

S (D , E a

k

2 3 4 5 6

k a

D

D

0.620

0.418 0.816 0.868 0.913 0.973

0.759 0.434 0.448 0.460 0.450

a

1 1 1 1 1

C Output Signal

The gain (the ratio of output signal to the input signal) Aamplifier of the amplifier, and the maximum gain A filter and bandwidth BW filter of the filter are modeled

using the appropriate MBDs in Table 7, with k 2 for the amplifier and k 5 for the filter. Figure 7(a) shows RMSE comparisons for Aamplifier for the region W1  [2,200] and W2  [2,200] , where W1 and W2 are the width of the two amplifier transistors M1 and M2 in Figure 6(a). When the space is narrowed down to W 2  [2,20] for W2 , RMSEs become worse (by comparison to RMSE for the MBD) for more of the 100 LHC samples as demonstrated in part (b) of Figure 7. This is expected as demonstrated earlier for the function in Figure (3).

)

E

R1

Figure 6: Two Electronic Circuits: (a) Amplifier (b) Filter.

Table 7: Second-Order MBDs F (D )

R2 R4

Input Signal

MBDs for k 2  6 (generated by applying the sufficient and necessary conditions in the references mentioned at the beginning of this section to the above design sets) are given in Table 7 for second-order MBDs and in Table 8 for third-order designs. Add one center point C (0 k ) for each row in the tables to complete the MBD.

k

Output Signal

M1 Input Signal

2.5 (a)

S 1 (D , E a

F (D ) k

k

2 3 4 5 6

k a

S 2 (D , E a

)

k a

Normalized RMSE

Table 8: Third-Order MBDs )

D

D

E

a

D

E

a

0.685 0.724 -

0.255 0.775 0.844 0.801 0.951

0.741 0.252 0.305 0.311 0.287

1 1 1 1 1

0.378 0.202 0.194

0.763 0.743 0.742

1 1 1

2

1.5 MBD 1

0.5

10

20

30

40 50 60 LHC Sample No.

70

80

90

100

30

40 50 60 LHC Sample No.

70

80

90

100

2.5 (b) Normalized RMSE

Note that the size of second-order MBDs in Table 7 is 1  k 2 k points for k d 5 . 4.

APPLICATION TO ELECTRONIC CIRCUIT MODELING In this section two electronic circuits are modeled using MBDs and the results are compared to LHC designs. The two circuits are the amplifier and filter in Figure 6.

2 MBD

1.5

1

0.5

10

20

Figure 7: Results for the Amplifier Circuit: (a) for the Region W1  [2,200] and W2  [2,200] (b) for the Narrower Region W1  [2,200] and W2  [2,20] .

233

Results for RMSEs for the filter circuit are shown in Figure 8. Note that while the results give advantage for the MBD for A filter as shown in part (a); however,

Box, G.E.P., Draper, N.R., 1959. A basis for the selection of a response surface design. Journal of the American Statistical association, 54, 622-654. Box, G.E.P., Wilson, K.B., 1951. On the experimental attainment of optimum conditions. Journal of the Royal Statistical Society, Series B, 13, 1-45. Chen, V., Tsut, K-L., Barton, R.R., Meckesheimer, M., 2006. A review on design, modeling and applications of computer experiments. IIE Transactions, 38, 273–291. Draper, N.R., Lawrence, W.E., 1965. Designs which minimize model inadequacy: cuboidal regions of interest. Biometrika, 52, 111-118. Goel, T., Haftka, R.T., Shyy, W., Queipo, N.V., 2007. Ensemble of surrogates. Structural and Multidiscilinary Optimization, 33, 199-216. Goel, T., Haftka, R.T., Shyy, W., Watson, L.T., 2008. Pitfalls of using a single critereon for selecting experimental designs. International Journal for Numerical Methods in Engineering, 75, 127-155. Mckay, M.D., Beckman, R.J., Conover, W.J., 1979. A comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics, 21, 239-245. Myers, R.H. and Montgomery, D.C., 1995. Response surface methodology: process and product optimization using designed experiments. New York: Wiley. Simpson, T.W., Toropov, V., Balabanov, V., Viana, F.A.C., 2008. Design and analysis of computer experiments in multidisciplinary optimization: a review of how far we have come – or not. American Institute of Aeronautics and Astronautics, 1-22. Viana, F.A.C., Haftka, R.T., 2008. Using multiple surrogates for metamodeling. Proceedings of the 7th ASMO-UK/ISSMO International Conference on Engineering Design Optimization, Bath (UK). 118. Wang, G.W., 2003. Adaptive response surface method using inherited Latin hypercube design points. Journal of Mechanical Design, 125, 210-220.

part (b) of the figure shows that RMSEs for the LHC samples are lower for BW filter . This is the worst case obtained in our work. Nonetheless, even for this case the RMSE for all 100 LHC samples is nearly 90% on average of the RMSE obtained using MBD as can be inferred from Figure 8(b). 1.5

Normalized RMSE

(a) MBD

1

0.5

10

20

30

40 50 60 LHC Sample No.

70

80

90

100

90

100

1.5

Normalized RMSE

(b)

MBD 1

0.5

10

20

30

40 50 60 LHC Sample No.

70

80

Figure 8: Results for the Filter Circuit: (a) A filter (b) BW filter . 5. CONCLUSIONS Metamodels are appropriate surrogates for simulators in the design of complex engineering systems provided that the errors incurred are acceptable. Bias errors due to metamodel inadequacy result in inaccurate metamodels when computer experimental data are used to construct these metamodels. This paper demonstrated that minimum bias computer experimental designs are potentially superior in response surfaces by comparison to space-filling designs such as the popular Latin hypercube samples. Also, the paper introduced minimum bias designs for normalized hypercuboidal spaces. The list of these designs is by no means exhaustive, and more work is needed to expand the list for higher-dimension spaces and higher-order minimum bias designs.

AUTHOR BIOGRAPHY HUSAM HAMAD is an associate professor in the Electronic Engineering Department at Yarmouk University in Jordan. He is the Vice Dean of Hijjawi College of Engineering Technology at Yarmouk. He received his B.S. in Electrical Engineering from Oklahoma State University in 1984, M.S. in Device Electronics from Louisiana State University in 1985, and PhD in Electronic Systems Engineering from the University of Essex, England, in 1995. He was a member of PHI KAPPA PHI Honor Society during his study in the U.S. His research interests include modeling, analysis, simulation and design of electronic systems and integrated circuits, metamodel validation, electronic design automation, and signal processing.

REFERENCES

Abdelbasit, K.M., Butler, N.A., 2006. Minimum bias design for generalized linear models. The Indean Journal of Statistics, 68, 587-599.

234

A PRACTICAL GUIDE FOR THE INITIALISATION OF MULTI-AGENT SYSTEMS WITH RANDOM NUMBER SEQUENCES FROM AGGREGATED CORRELATION DATA Volker Nissen(a), Danilo Saft(b) (a) (b)

Ilmenau Technical University, Faculty of Economics, Institute for Commercial Information Technology, Chair of Information Systems for Services (WI 2), Postfach 100565, 98684 Ilmenau, Germany (a)

[email protected], (b)[email protected]

different initialisation values. Researchers may want to relay data acquired from the real world (e.g. through measurement series, questionnaires or statistical archives) to initialise their agents with according parameter values for reasons of testing, prognosis, or simply for validity. In the case of social or economic simulations, an agent may possess variables such as income, reputation, job satisfaction, household size, etc. Scientists may however not in all cases be lucky enough to find realworld-data at a level that is as detailed as their desired simulation setup may require. They, therefore, may need to evade to more aggregated forms of simulation, matching the aggregation level of the empirical data available. This option can be unsatisfactory as important details of the micro-level to be simulated and/or the emerging micro-macro-links within such a simulation might need to stay unaddressed. One alternative is testing different random number distributions where detailed data is missing. This latter approach holds chances, but also challenges, such as the possible availability of empirical data that cannot give numbers on a detailed level, but gives aggregated distributions and correlations of different variables found in an empirical study. Table 1 (Oreg, 2006, p. 88) shows an instance of the results such empirical surveys yield. The values shown in table 1 were put forth by Oreg in 2006 and will serve as a data example throughout this paper. Table 1 gives descriptive statistics and correlations between parameters important to individual behaviour in the context of organisational change. The statistics were extracted from a series of questionnaires given to individuals within a company undergoing several organisational adjustments. The researches recorded variables thought to be important to an individual’s opinion formation about an organisational change. They derived a static statistical model of interdependencies of individual properties (tb. 1, variables 1 to 9), the resistance an individual develops towards an organisational change (tb. 1, variables 10 to 12) and the behavioural outcome its opinion has on its specific job (tb. 1, variables 13 to 15). From a multi-agent modelling perspective, it becomes possible to analyse the dynamic behaviour of a simulated company as a whole by modelling individuals

ABSTRACT This article describes a scalable way to initialise a simulation model with correlated random numbers. The focus is on the nontrivial issue of creating predefined multidimensional correlations amongst those numbers. A multi-agent model serves as a basis for practical demonstrations in this paper, while the method itself may be interesting for an even wider audience within the modelling and simulation community beyond the field of agent-based modelling. In particular, we show how researchers can create streams of correlated random numbers for different empirically-based model parameters when just given aggregated statistics in the form of a correlation matrix. An example initialisation procedure is demonstrated using the open source statistical computing software “R” as well as the open source multi-agent simulation software “Repast Simphony”. Keywords: MAS Parameterisation, Correlated Random Numbers, R-Project, Repast 1. INTRODUCTION The simulation of a model may sometimes require a large amount of parameters, which influence its outcome (significantly). In a subset of these cases, the parameters may be interdependent in such a way that the initialisation of a model needs two or more parameters to correlate in a predefined manner. A procedure to generate and utilise such numbers will be explained in the following. We use the example of an agent-based model, since one of our main research areas is the field of agent-based economics. In this research, we regularly find illustrative scenarios to which the concept presented in this paper is applicable. Note that while the statements here will be kept limited to agentbased models for scientific validity, these explanations can easily be transfered to the initialisation of other types of simulation models. In a variety of multi-agents systems, the model to be simulated may consist of a large number of heterogeneous agents. Heterogeneity can come in the form of different spatial positions of individual agents, different network connections, opinions, etc. In general, each of these agents possesses a set of parameters with

235

Table 1: Exemplary descriptive statistics and correlations for the variables of an empirical study by Oreg (2006, p. 88) Table 2: Exemplary descriptive statistics and correlations for the variables of an empirical study by Oreg [2, p. 88]

with heterogeneous specific opinions and the (direct or indirect) influence individuals have on each other, e.g. through social interactions and information exchange (cmp. tb. 1, variables 8 and 9) or other behaviour affecting organisational “neighbours”. In this example, the goal of researchers in the field of multi-agent-based simulation (MABS) would be to better understand the process of organisational change or analyse the effect of certain formal and informal hierarchies and network structures on company performance. In fact, such investigations are taking place within our own MABS research of which a first part has already been published (Nissen and Saft, 2010). We utilise the real-world study of “resistance to change” behaviour in organisations to initialise our own multi-agent simulation of a virtual organisation in order to better understand how resistance to change spreads and can be influenced by management. This requires (here: agent-based) modelling at a more detailed level than the aggregated statistical data in table 1 provides. To this end we can however use the values in table 1 to yield specific initialisation data for a (large) number of simulated members of a virtual organisation. To initialise each agent in our simulation model with its own personality values, it is necessary to extract sequences with specific random numbers while accounting for the multidimensional correlations given in table 1 in order to yield correct number sequences. This challenge is trivial only when one needs to generate two correlated series of random numbers, i.e. when the agents in the simulation only possess pairs of two correlated properties.

where x1 and x2 are two uncorrelated random numbers from a given distribution, ଵ and ଶ are their standard deviations, and c is the desired correlation coefficient between ଵ and ଶ ; i.e. the resulting correlated random numbers. With more than two sequences of correlated random numbers to generate, one can use a variety of mathematical approaches that are more or less difficult to go through manually. In our case, an Eigenvectordecomposition was employed as we found it to be a process that is robust and offers good performance. There also is the option of using the so-called Choleskydecomposition (Lloyd and Trefethen, 1997, pp. 172 178) which will however not be explained further here. Given a correlation matrix C (see columns “1” through “14” in table 1) one can define a matrix

  ௜  λ୧

where Ei are the eigenvectors of C and λi are the eigenvalues of C. With a matrix Iu consisting of formerly uncorrelated random numbers, we can derive a matrix ௖  ௨  ்

(1a)

ଶ   ∗ ଶ  1  ଶ ∗ ଶ ∗ ଶ

(1b)

(3)

where  ் is the transpose of  . ௖ then contains random numbers with correct correlations. Iu can consist of any number of random values where each line can represent the initialisation values for a single agent and each column stands for one of the correlated parameters of an agent. This offers great flexibility since one only needs to choose the number of rows in the original matrix Iu as big as the number of agents one wishes to instantiate/simulate. Iu should already include the general distribution properties such as – in our case - the mean and standard deviations given in table 1.

2. A LITTLE BIT OF MATH For only two numbers (parameters) to be correlated, there is a simple approach to retrieve two correlated random numbers from a set of uncorrelated random numbers: ଵ  ଵ ∗ ଵ

(2)

236

3.

A PRACTICAL GUIDE FOR THE UTILISATION OF CORRELATED RANDOM NUMBER SEQUENCES USING “R” AND “REPAST” While the pattern to generate correlated random numbers shown in section 2 can be time-consuming to do by hand, it is a very easy process once one employs supporting software. The popular and well-documented open source program “R” (Hrishikesh, 2010; Jones, 2009) is able to make the necessary calculations (for all practical purposes implied here) in just a fraction of a second. The tool, along with many additional packages, can be downloaded freely for a variety of platforms (The R Project, 2010). Below, we will present exemplary step-by-step R code to generate random number sequences for the three correlated parameters “improvement in powerprestige”, “improvement in job security”, and “trust in management” listed in table 1. The code is to generate correctly correlated random numbers for 2500 agents as virtual “employees” of a simulated organisation. Note that the code pattern is scalable to a large number of parameters and agents. The first step is to generate a matrix of uncorrelated random numbers for each agent, already taking into account the correct mean and standard deviation properties (in this case assuming the Gaussian distribution from table 1):

E  /   / /0 @6/  @/     /0 @6/  6/@ 50

//  /D  @/    @  .  & = 5 @ # // &5# 9  @/ = / ./@ 9%"$   D   6     @/ =  @/ = /= ./@ %"$ /0 @/   6/   /0 0 0   >/     9D  @  @/0  66/     @ %/  0  @/=@    >/     @/0  6/ D0  %"%  6=  / 9%"$ @/=@ 0   66   @     "$-:'5$     4(  -&C$" 2 =7 /  (  :   %"%   9%"$ //   /0 //  /  /   /0 // 66 /   /0  DA

241

%/ =&/  %  %&% B   /  ; D0 0  /0 J / !/     / @  /0 @ 60 @  D /0 /0 @  @/      0/  ./@ 

6   660  /0   / @  /   /   50   / . .@ > /0 @ /   /  @/  /   /0 6  /D @   /    / / // /0 1 /.  /0  660      4 6 /  @DA / . @ . 0  /0    /    /=  @/  50/  / . 6 / /   . D0 0      @/   /   '-    A6 /0   / .  /0  @/  7/ /0  @DA   @ /     @  @0 0 0. D /0/ /0 6 .  0  /D   /   / /0 @    @@ /  /D    / !6  /.   # /0 /  ; /0 @  6/  /0  /  "7%   #&% %$  / 50  6 //  @60 >      / 0 //

/0  /!/  #&% %$ # /0 /   ; @/0 /   @6//    &"%   '- @   #&% %$  / # /0     D @=6 /0 @ //   @/0   / 66/  40

@@ /   M / /  @  0 . /0 "  /       / /. @6    =  /  //    /       / /   /0 7B# 0 //      !/  //    / 6 /    /. /     /   // @/ = / ./@ # @/ = /   @/   /0 !/    /  0    

   

  F 7  $//   " /   0   =. 50  =.    /      DF

40/0 5)& 6&%%5 & 6-% 50 @/ = / @ @  / 0/D :0  / /.  /0 /  ./@ 6  1  /  / /.          /0 @6/ 6 //  #/  /0   / 660  /0 @/0@/  @ /    /0 6    / /  @      /. 50   //  / 6 / 0/ /  66 / /0  >/   /   . @ /0  / @@ 0/ /  " / 6@ / / 66! @/  / /0    /0 @@   /0  //   /0  >/  &6 //   @  / D /0     /.

  /  /0  >/   / 50/3 @A /0 @@  /    / @   /0.   /  /0  @/ 

  F %@ 2 D6 /   #@6@ //  7070 (% 6'% P  /0 @DA  /@6 /   @/  ./@ Q  4   / /  @6  /0 0  @ / @   /   / / /0 / /0/ /0 0    / / /D   50 

244

 / /  @A @   /0  . 0 >/   /0 L M  A /0  /   P.  4;  /0 ./@   % 0    4; D0 0  / 6/   6@/     0    / /@6 . @  # 6  /.  /0 /0   /3 //  /       . @  /@6 D0 / @ / @ /6  0 0  6  50   @A  / 6/ /0 . @  //  /0      /  6  6/  /0 D 0  /0 @ 50/  / .   / /0  / /0  / 6    . D0 0 /. 0 / D /0 /0 /  /0  @/  #&% %$   // //   @ %     %     4 66   /0@ 6/ / #&% %$ D0 0 @  /0 6  /D   D /0   / /@6 . @  50/ 6@ / / 66.  . /0 66 @/0  

    6/ / / /0 0     !@6 D0    D /0  0 0 1 .  

 /  /0  @/   

  @   A / 0    0  @ @ 1 .     /0 @6@ //  60 D0    !6 90  

# /0  /  D 6 /  @/0. 6@ //  /      6 //  6D /  @/ =  %"% 6 /     @   / D0 0 @ / @6! @/ =  //   #/3 . / // . @6  /0  / /  @       0 /  6 //  D / / / /    /   50

 /  

  @  / /. 6 @    /  /0   "/ /0/ D   /0 @ / D0 0    @6     66/ .  >/      . D0 0 @ /  /0  >/    66/ . / @@ "/ / / 3    / /  ! /0 @ @ 1 .  0      6 .  D0 0 1 . 0  / 6/    D0

/ // D /0  /0     / /0

//   /D   9D  /0  /  D  /   /  /0 @0 @ / 66! @/ /  /A /   /0  @/  / / @ / @  /0 !/  " D  /  A

/0 6@ D0 0   /0       / @/  / / /0 @0 @  /0 /   /0@ 5 // /0  0     @/  D 0 @ /0   / .  /  !6 B  /  ; C / DA D    / /   @DA 6/ / #&% %$ D /0  /  @ @ D0 0 6@ / . @  0      /    /0  @/ 

  )F :!@6  %/ =" / %/ =  @  D /0 B  / 5@6 /  50 6    6@ /  / @ /    /  / / D0 0  ! /0   . /@   /  @   9 D 6A / /0 1 .     !6 9/> 50   / 0D @ . / @    /3 . / !/ /0 6/  6  /0 . @  //     /3

@  /0/   /      / 6  @ @ 1 . .  D0 0 /0  @/ 

 /0   / 

  @  /  #    6@ / / 0   / /  6   /   D /0   / 1   / 6/ /0 0 0   / A6  /  @/   /0  /   D /0 /0  1 . 50/3 D0. /0 !@6  /0   ) /0 1 .  /0     1 / )>  /0   /   /   /0 @ 0   /0  @ @ 1 . 50 /0  / / @ @ /0 //   /D   # D  /  D /0 /0 6  !@6 / . /0/ /  0     @ @ 1 . 1 / 4 9> /0 @  / /0  1 . / 4 9D   /0    

  .  4  0  / / /0 / 

 . /0   /  1 . 0 0 / 4 9>  .   1 / ) 9> /  6 . . / /  0 0 1 . / 4 $ /

 . / . @ . @ . /0 1 .  

* : *   "    4 %/ = @/0   @6! 6F "  .   A 0 !    !   ; F  )O= O) ' > < =-   $68  /  /  6/   ./@ F   @8/0  : /   %  #$7 4=44=(4O= B  -   9 / &  ;     % /0/ '    &/  F #  B   ' /    % @ /      "#  $  !  %$!&'( B " 2  B %  5 44 %/ = " / 7 $ @/  F P0  /0 " /   )*$+  0 // hierarchical universe 6=F‰L‰A‰K // functional objects F = {f | f : U*o U} // global functions L = {f | f : Levelj*o Levelj} // level structures A = {f | f : Levelj*o Levelj+1}// abstractions K = {f | f: Levelj*u Levelj+1o Levelj+1} // knowledge abstractions I = 6*ˆ R //initial functions R = {r | r6*u R*o 6 u R } // transformation rules.

Without giving up anything essentially human, culture, social or natural togetherness, different approaches, humans have a lot in common: philosophic desire, comprehension of the own hierarchy in the context of the other two, free life based on understanding the necessities, constructive fear of the unknown, and especially the love for creation. Except the three cultural ways, that permanently Divide et Impera et Intellige, there is no other. (Niculiu 2008)

5. SEMIOTICS  SYNTAX u SEMANTICS Transferring an ontological approach, communication through language requires the distinction of three levels:

We need Consciousness to return intelligently to Faith People of one choice exist, in all senses of the word. They either comprehend all the alternative ways and their convergence, or, in the context of natural love for philosophy and interest for the other selectable directions, put more passion in one direction. Of the first category are temporary elected, in different convergent hierarchical modes, the social leaders, of the second, the institutional directors. Both kinds of leaders are more philosophical than their cohabitants, even if the ones master the strategic perspective given by an attained peak, while the others have the joy of the courage to climb into profoundness. The elected artists permanently reconfigure a system of laws, to be beautiful by intelligibility, true by consistence, and good by human understanding. The elected physicists, pure or from different correlated scientific domains all collaborating with mathematics and engineering, govern by research strategies with Gods Fear. The elected engineers critically construct and criticize constructively. For any social role, the elected concentrate, respectively, on Faith (mathematicians), Intelligence (physicists), and Conscience (engineers). There always exists a human, called No.1 or the Philosopher, depending on the stability of the times, cloudy or clear Sky. He will always lead directly the elected or the philosophers, who will know to educate and learn optimally the humans of all ages, including themselves. We have to start. Otherwise, it is no hurry. Intellige is to link, to understand, and to be aware. In Latin: intellego = to understand, to feel, to master, to gather in mind. Artificial has a derogatory sense; however, the root of the word is art. Arts remind of liberty, as Arts for arts. Artificial is at first sight the complement of natural. Our ideas transfer us to places that are neither natural nor artificial. Maybe artificial means something natural created by the human being and Nature is an extension of our body. However, we feel to be superior to Nature, as to our body: we can think. (Penrose 1994) Why are only humans creating arts, why do they need to know more, and why do they construct other and other natural things they have not found in the Nature? We learned the arts have to discover the Beauty, that science has to look for the Truth, and that engineering invents things to help us, caring for the Good.

1) the level of reality; 2) the level of cognitive representation of this reality; 3) the level of material representation - text, signs, images etc When we acknowledge an object in association with a certain sign, than marks are created in our brain in virtue of which the simple appearance of the same sign will evoke a thought or reference directed to this object as the impressions stored in the memory were reactivated –see Figure1 (Ogden and Richards 1930).

THOUGHT OF REFERENCE

symbolizes

refers to

SYMBOL

REFERENT stands for

Figure 1: Semiotic triangle The solid lines in this triangle are meant to represent the causal relations of symbolization (remembrance, evocation) and reference (memory, perception). Opposite to these, the dashed line signifies that the relation between symbol (word) and referent (reality), linguistically the most important, is barely imputed. The immediate conclusion is that the multiple perspectives in multilinguistic endeavors are (at best) locally, temporarily and partially resolvable. Assuming the referent (reality) is an existing entity for all the interlocutors, they may still have different thoughts (concepts) associated to the same referent, depending on their social-geographical personal universe or past experiences. It is therefore a difficult task for the translator to find the most suitable word for the most similar reference in the target language. As difficult as the above situation may seem, it can get even worse: the situation in which the referent exists in one language - and has both a reference and a symbol accordingly attached - and is inexistent in other(s). Language can here not overpass its limits, e.g., snow for the languages of the hot climate countries.

260

Arts are free, and even when they return to Reason, as mathematics, they bring results, that could before just be seen by Intuition, to send by Inspiration and Imagination to Intelligence. Physics reaches and gets conscious of Reasons limits, both by the quantum theory and by the too complex phenomena, e.g., society and human. It looks like there is no difference for the intelligence that is useful to one of the ways. An example, that confirms that they simply represent different approaches to understand and develop the (presently natural) Reality, is architecture, which we cite in each of them. To conclude: Intelligence is more than Reason, to make us feel as beings superior to Nature, what also means that we have to respect Nature more:

The principles are structured/ typified corresponding to the higher level => hierarchy types. For the classical representation problem space = (states, actions), problem solving means the process starting from an initial state to look for an operation set that leads to a result state. Solving strategies structure the process to look for the solution (goal-project-concept). Intelligent systems demand a cosimulation of the parts belonging to different domains, e.g., hardware & software, in the context of unified representation for design and verification. Unified simulation of hard-soft systems is imposed by the incompatibility or non-optimality that results by the initial partition of the system, as by the inefficiency of traversing the design-verification cycle for a fixed partition. Unified simulation methodologies eliminate the rigid partition constraint: It implies planning and learning, i.e., the possibility of communication between different levels of the knowledge hierarchy. Intelligent simulators can learn by iterative generation and validation of models, possibly interactive. The objective of the human-machine dialog is to advance toward the simulated intelligence by transmitting the knowledge between human and his mental/ physical extensions in a common language. The input dialog is oriented toward learning. Knowledge bases on a morphism that applies the behavior of the object-system on the internal model of the simulator. The output-dialog on the result specifications is oriented toward planning. The dialog can be extended to the internal unfinished zones, to maintain the integrity of the hard-soft simulation. Further, communication concordant to the humanmachine dialog principles can be also extended from assuring the interface problems between the knowledge hierarchy (planning/ learning) and similar activities corresponding to the hierarchy types that are based on simplifying abstraction forms. The three different essential ways to approach this goal have common central themes: learning and planning, knowledge representation, and functional constraints.

Spiritus Sanus in Mens sana in Corpore sano Therefore, there is something else in the Intelligence, which allows us to consider ourselves humans, human groups, peoples, beings on the Earth, or conscious beings in the physical Universe. We also feel that there is something essential beyond the physical – the metaphysical (Plato). More, there is something exterior to the human intelligence, without that we could not fight the Time to evolve. We have to feel complete, even if we need education and permanent work in communication with the other humans, of the past, the present, and the future. We need Conscience to link Faith to Intelligence We have to remember the abstractions that assisted us to go further. We said complete human to someone complete in a context, what implicitly supposes the power to go beyond the context. This is the story of the integers (integer = perfect, complete): they have a beautiful complete theory, however, do not forget to build the rational numbers to feel as close as needed to any real number. Nevertheless, they realize this is not enough, rewarded by the conscience of the continuous Reality – infinitely more powerful than the discrete/ countable one. To IR, we get by the perfect circle that is beyond the power of Reason. For example, we plan to realize artificial intelligence, to have a friend that is conscious of the problems to solve together. For the moment, there is no artificial intelligence. However, we learn to be conscious of the computer limit to process only rational numbers. This means it uses a sequence (xn)nIQ that converges to n—a (Newton), what reminds us of the density of IQ in IR.

x

x

x

6. INTELLIGENT SYSTEMS The reasoning of systems capable of reflexive abstraction, i.e., intelligent, starts by describing the problem, and is controlled by problem solving strategies; these derive from the approach principles contained by a knowledge level superior to that of the current simulation.

Concept-symbol analogy: concept representation and symbol operation try to simulate the mental processes. Structural analogy: the activity of brain is emulated by neural networks, cellular automata, genetic algorithms, membranes or quantum computing. Hierarchic-parallel analogy: thinking is considered a collective phenomenon that is produced by constitutive phenomena parallel and recurrently.

The limits of the knowledge domain for intelligence simulation are reconfigurable: learning can guide the representation - semantics and architecture of the system, and functional constraints can formalize the cognitive constraints in the spatial-temporal reasoning context.

261

The informatics extensions of the contemporary human impose the knowledge of a third language to use his artificial mental or physical extensions, next to the mother language for context integration and to the one surpass the context - nowadays, American English. The evolution of the programmable systems from punctual activities as answer to explicit orders, to autonomous activities, supposes a knowledge-based high-level symbolic object-oriented dialog, to awake the consciousness by explicit selections, and the adaptability by assisted decisions. As result of formal version of (part of) the natural language, a high-level language for intelligent dialog has to inherit: x

x x

Consequently, understanding is simulated by the evolution of the representation language in the symbol hierarchy. This approach is that of the classical artificial intelligence. Its limits proceed from restriction to logic sequential mathematical discrete reasoning, what results in the incapacity to represent conscience, intention, intuition, i.e., intelligence. The regularities of the cognitive processes are represented as inferential strategies common to the dialog partners: inference is not just deductive, but mostly inductive. CONCLUSIONS Conscience simulation demands transcending the present limits of computability, by an intensive effort on extensive research to integrate essential physical and mathematical knowledge guided by philosophical goals. Even mathematics will have to develop more philosophy-oriented to approach intuition. Simulability is computability using the power of continuum. There are positive signs for this from analog electronics, control systems, and mechatronics. Real progress towards this way of computation needs unrestricted mathematics, integrated physics and thinking by analogies. Evolution implies the separation of faith and intelligence, so we have to better understand both, integrating them to human wisdom, to be divided further to get more human. Metaphorically phrased, our searches and researches should have as axioms:

syntactical regularities (studied by computational linguistics) and semantic correspondences (studied both in language philosophy as in AI), regularities of the cognitive processes (studied by cognitive psychology and intellectics), relations with the I/O system (perception/ action) of the individual intelligent agent, and with the interactions of the individual intelligent agents (social relations) in the intelligent system.

Formalized conform to information theory, syntax and semantics offer a representation of a world preexistent to the dialog. The resulted formal system has to be correct - any formula corresponds to a fact and any formal computation to a real reasoning, and complete – any real fact corresponds to a formula and any real reasoning to a formal computation. Consequently, understanding is simulated by the evolution of the representation language in the symbol hierarchy. This approach is that of the classical artificial intelligence. Its limits proceed from restriction to logic sequential mathematical discrete reasoning, what results in the incapacity to represent conscience, intention, intuition, i.e., intelligence. The regularities of the cognitive processes are represented as inferential strategies common to the dialog partners: inference is not just deductive, but mostly inductive. The evolution of the programmable systems from punctual activities as answer to explicit orders, to autonomous activities, supposes a knowledge-based high-level symbolic object-oriented dialog, to awake the consciousness by explicit selections, and the adaptability by assisted decisions. As result of formal version of (part of) the natural language, a high-level language for intelligent dialog has to inherit: x

x x

x God is Unique. x His ways are Uncountable x His plans are Hierarchical. Philosophy is not a specialty but a human right. There have to be schools to prepare the teachers of philosophy for the other humans. These schools have to develop also respect for those that look for the Way on one of the three alternative paths that correspond to the fundamental partition (arts, science, engineering). Because recently the essential Divide et Impera do not Intellige, the only philosophers are the masters in: x Arts – especially mathematicians, and others that, aware or not, compose mathematically x Science – physicists, and those that do not forget their science is a chapter of physics x Engineering – mostly those working in domains that attain the limits of the pure Reason. Mathematics is one of the arts. The music is at least as beautiful and expressive, but mathematics does not demand an extraordinary talent, allows a reasonable dialog about it, and has well-defined reconfigurable limits of that it is aware. Mathematics has to be educated as soon as possible and has not to be confounded with its handcraft. The music gets more often out of its character. The two arts evolved together: Johann Sebastian Bach, Antonio Vivaldi, Joseph Haydn were musically gifted mathematicians, who preferred the liberty of the music to the bands of the Reason.

syntactical regularities (studied by computational linguistics) and semantic correspondences (studied both in language philosophy as in artificial intelligence), regularities of the cognitive processes (studied by cognitive psychology and intellectics), relations with the I/O system (perception/ action) of the individual intelligent agent, and with the interactions of the individual intelligent agents (social relations) in the intelligent system.

262

The Reason, as initial zone, makes mathematics more sure but less charming than the other arts that can refer directly to the Reality: literature, music and sculpture. The visual arts are too dependent of the Nature because seeing is the most used sense for the human natural being. The mathematics school is continuous, whereby sculpture, literature, and music can generate sooner higher singular peaks: Michelangelo, Shakespeare, Beethoven, by an exponential/ other highly nonlinear continuity. Arts are free. But mathematics first expressed reasonably that Reality could only be approached by Reason Physics is the Science. The other natural and social sciences are its chapters, even if they are not yet aware of it, or just try to return to their riverbed by intermediary specialties instead of integrative bridges. As any artificial system, the society is structured on natural bases, and it develops by natural laws. The modern age forced these laws towards Reason, and recently they got out of control. The social laws got also unreasonable. Physics is essential for the constructive reconfiguration of the Faith. Engineering is most frequently both art and science, and is as important as arts and sciences in the fundamental partition of the Reality needed for evolution. However, it is more dangerous than its alternative approaches, of which it has to be strictly bridled. Reasons are twofold: Its result, called technology, is defined by its complement – so it is not superior to this. It does not impose spiritual proximity between the creator and the user – so it can be applied in a complete different scope than it was generated. However, any engineering is the homonymous complement of a special science that collaborates with mathematics, therefore, integrated sciences into physics and mathematics remaining among arts solve the case.

Niculiu, T., 2008. Object-oriented Symbolic Structural Intelligent Simulation, Bucureúti: Matrix Publishers. Ogden, C., Richards I., The Meaning of Meaning, 3rd ed. New York, 1930 Penrose, R., 1994. Shadows of the Mind, Consciousness and Computability, Oxford: University Press Rabaey, J., 1997. Reconfigurable Computing, Proceedings of IEEE International Conference on Acoustics, Speech & Signal Processing, pp. 127132. March 28-31, München. Rudin, W., 1973. Functional Analysis, New York: McGraw Hill. ùtefan, G., 2010. Ethos, Pathos, Logos, Bucureúti: ALL. Traub, J.F., 1999. A Continuous Model of Computation, Physics Today, 5(5), 39-43. Zeigler, B., Praehofer, H., Kim, T., 2000. Theory of Modeling and Simulation, Oxford: Academic Press. Zhong, N., Weihrauch, K., 2003. Computability Theory of Generalized Functions, Journal of Automated Computer Machinery, 50(6), 574-583. MARIA NICULIU is MA (European Intercultural Communication Strategies) and MS (Mechanics Engineering). Her MA thesis (2011) is about Von intensiv zu nachhaltig – from intensive to sustainable. She graduated the Faculty of Literatures and Foreign Languages, German/ French/ Dutch - Bucharest University (2009) with a paper on Sprachenvielfalt und Mehrsprachigkeit, was taugt die Europäische Union als Standort der Mehrsprachigkeit - Multilinguism and what makes the European Union as model for a multilingual society. After the MS from the Faculty of Mechanics Engineering of University Politehnica of Bucharest (1987), she enriched her technical and economical background - member of the Romanian Authorized Accountants and Financial Experts (2008), and directed her interests towards a multicultural approach of the European social and economical evolutions, based on linguistic and behavior simulation. TUDOR NICULIU is Professor at the Electronics, Telecommunications, and Information Technology Faculty of the Politehnica University in Bucharest, and Senior Researcher at the Center for New Electronic Architectures of the Romanian Academy. He is looking for hierarchical integration of different domains, to understand intelligence by simulating it, and to apply it to intelligent simulation. Since 1991, he teaches and researches at the same institution, PhD 1995, MS 1985. Before, he was Senior Researcher at the R&D Institute for Electronic Components in Bucharest, researching and designing hierarchical simulation of analog integrated circuits. He studied Mathematics at University of Bucharest (MA 1994). He published 12 books, 25 journal articles, and 69 international conference papers. He is IEEE Senior Member of CAS, Computer, and SMC Societies, as Fellow of International Institute for Advanced Studies in Systems Research & Cybernetics.

REFERENCES Ageron, P., 2001. Limites inductives point par point dans les categories accessibles, Theory and Applications of Categories, 7 (1), 313-323. Hofstadter, D., 1979. Gödel, Escher, Bach - The Eternal Golden Braid, Washington DC: Vintage. Keutzer, K., et al., 2000. Orthogonalization of Concerns & Platform-based System-Level Design, IEEE Transactions on CAD of Integrated Circuits and Systems, 19 (12), 1523-1543. Lupu, C., 2004. Locality Measured by Contour Patterns - A Topographic Model. Proceedings of 15th IASTED International Conference on Modelling and Simulation, pp. 50-54. March 1-3, Marina del Rey (California, USA). Miller, R., et al., 1993. Parallel Computations on Reconfigurable Meshes, IEEE Transactions on Computers, 42(6), 678-692. Niculiu, M., Niculiu, T., 2009. European Spirit Evolution by Multicultural Harmony. Proceedings of 6th International Symposium on Personal and Spiritual Development in the World of Cultural Diversity

263

AN ASYNCHRONOUS PARALLEL HYBRID OPTIMIZATION APPROACH TO SIMULATION-BASED MIXED-INTEGER NONLINEAR PROBLEMS K.R. Fowler(a), T. Kopp(b), J. Orsini(c), J.D. Griffin(d), G.A. Gray(e), (a,b,c)

Department of Mathematics and Computer Science, Clarkson University (d) SAS (e) Sandia National Labs

(a)

[email protected], (b) [email protected], (c)[email protected], (d)[email protected], (e) [email protected]

over the last several decades on a wide range of applications (Conn, Scheinber, and Vincente 2009). Because DFO methods only rely on function values, parallelism is often straightforward and, in the case of expensive simulation calls, can make otherwise intractable problems solvable. . Hybrid DFO algorithms have emerged to overcome inherent weaknesses and exploit strengths of the methods being paired (Talbi 2004; Raidl 2006; Alba 2005). Often, the hybrid algorithms are designed to address problems that could not otherwise be solved. In this work, we focus on the parallelism of a hybrid evolutionary algorithm with a local search that was designed for simulation-based mixed-integer problems with nonlinear constraints (Griffin, Fowler, Gray, Hemker, and Parno). The performance of the hybrid was demonstrated on a suite of standard test problems and on two applications from hydrology (Gray, Fowler, and Griffin 2009; Gray, Fowler, and Griffin 2010; Griffin, Fowler, Gray, Hemker, and Parno) that were known to be challenging for a wide range of DFO methods (Fowler, Kelley et al 2004; Fowler 2008). Some of those challenges, which are not unique to environmental engineering, included discontinuous optimization landscapes, low amplitude noise, and multiple local minima. Specifically, in (Fowler, Kelley et al 2004), a comparison of derivative-free methods on the hydrology applications showed that a genetic algorithm (GA) performed well in terms of identifying the correct integer variables but then failed to achieve sufficient accuracy for the real variables. On the other hand, given a reasonable initial iterate with respect to the integer variables, the local search methods showed fast convergence. These observations motivated the pairing of the GA with a generating set search approach, referred to as Evolutionary Algorithms Guiding Local Search (EAGLS). The resulting algorithm pairs the binary mapping of the genetic algorithm to handle integer variables with asynchronous, parallel local searches on only the real variables. The new method has strong global search aspects and can still maintain high accuracy from the local search phase. Previous studies focused on the ability of EAGLS to solve a variety of MINLPs with varying difficulties

ABSTRACT To address simulation-based mixed-integer problems, a hybrid algorithm was recently proposed that combines the global search strengths and the natural capability of a genetic algorithm to handle integer variables with a local search on the real variables using an implementation of the generating set search method. Since optimization is guided only by function values, the hybrid is designed to run asynchronously on a parallel platform. The algorithm has already been shown to perform well on a variety of test problems, and this work is a first step in understanding how the parallelism and local search components influence the search phase of the algorithm. We show that the hybridization can improve the capabilities of the genetic algorithm by using less function evaluations to locate the solution and provide a speed-up analysis on a standard mixed-integer test problem with equality and inequality constraints. Keywords: genetic algorithm, pattern search, asynchronous, mixed-integer nonlinear programming 1. INTRODUCTION The need for reliable and efficient optimization algorithms that do not require derivatives is common across engineering disciplines. In general, the optimal design process requires such algorithms to work in conjunction with simulation tools, resulting in what is known as black-box optimization. For example, the simulation may require the solution to a system of partial differential equations that describes a physical phenomenon. These problems are challenging in that optimization must be guided by objective function (and possibly constraint) values that rely on a computer simulation, without any additional knowledge other than the output from the simulation itself. The simulation may be computationally expensive and add undesirable features to the underlying problem such as low amplitude numerical noise, discontinuities, or hidden constraints (i.e. when the program simply fails to return a value due to its own internal solver failure). Derivative-free optimization (DFO) methods have been developed, analyzed, and demonstrated successfully

264

in constraint formulations and problem size. In (Griffin, Fowler, Gray, Hemker, and Parno), EAGLS was able to solve a water supply hydrology application that previously could not be solved without significant parameter tuning of either of the two software tools that were merged to create the hybrid. Little work has been done to understand how the asynchronous parallelism that is inherent in the implementation impacts the search phase of the algorithm. This work is a first attempt at using parallel performance measures to understand the algorithms strengths and weaknesses. For this work, we consider objective functions of the form ݂ǣԹ௡ೝ ା௡೥ ՜ Թ and mixed-integer nonlinear optimization problems of the form ‹௣‫א‬ஐ ݂ሺ‫݌‬ሻ.

During the selection phase, better fit individuals are arranged randomly to form a mating pool on which further operations are performed. Crossover attempts to exchange information between two design points to produce a new point that preserves the best features of both ‘parent points’. Mutation is used to promote a global search and prevent stagnation at a local minimum. Termination of the algorithm is typically based on a function evaluation budget that is exhausted as the population evolves through generations. Often, GAs are criticized for their computational complexity and dependence on optimization parameter settings, which are not known a priori (Dejong and Spears 1990; Grefenstette 1986; Lobo, Lima, and Michalewicz 2007). Parameters like the population size, number of generations, as well as the probabilities and distribution indices chosen for the crossover and mutation operators affect the performance of a GA (Reed, Minsker et al. 2000; Mayer, Kelley, et al. 2002). Also, since the GA incorporates a randomness to the search phase, multiple optimizations are often useful to exhaustively search the design space. However, if the user is willing to spend a large number of function evaluations, a GA can help provide insight into the design space and locate initial points for fast, local, single search methods. The GA has many alternate forms and has been applied to a wide range of engineering design as shown in references such as (Karr and Freeman 1998). Moreover, hybrid GAs have been developed at all levels of the algorithm and with a variety of other global and local search DFO methods. See for example (Blum, Aquilera, et al. 2008; Talbi 2004; Raidl 2006) and the references therein. The EAGLS software package was created using the Non-dominated Sorting Genetic Algorithm (NSGAII) software, which is described in (Deb, Pratap et al. 2002; Zitzler, Deb and Thiele 2000; Deb 2000; Deb and Goel 2001). Although a variety of genetic algorithms exist, the NSGA-II has been applied to both single and multi-objective problems for a wide range of applications and is well supported. In particular, it is deigned to be used “off-the-shelf” which made it a good candidate for hybridization.

(1)

Here ݊௥ and ݊௭ denote the number of real and integer variables and, ‫ א ݔ‬Թ௡೥ , ‫ א ݖ‬Ժ௡೥ . In practice, ȳmay be comprised of component-wise bound constraints on the decision variable in combination with linear and nonlinear equality or inequality constraints. Often, ȳmay be further defined in terms of state variables determined by simulation output. We proceed by first reviewing the genetic algorithm, the generating set search method, and software that are hybridized to form the new algorithm. We then present numerical results and outline future directions. 2. EAGLS 2.1. Genetic Algorithms The EAGLS approach combines a genetic algorithm and a generating set search approach. GAs (Goldberg 1989; Holland 1975; Holland SIAM) are one of the most widely-used DFO methods and are part of a larger class of evolutionary algorithms called populationbased, global search, heuristic methods (Goldberg 1989). GAs are based on biological processes such as survival of the fittest, natural selection, inheritance, mutation, or reproduction. Design points are coded as “individuals” or “chromosomes”, typically as binary strings, in a population and then undergo the above operations to evolve towards a better fitness (objective function value).

2.2. Generating Set Search and APPS Asynchronous Parallel Pattern Search (APPS) (Hough and Kolda 2001; Kolda 2004) is a direct search methods which uses a predetermined pattern of points to sample a given function domain. APPS is an example of a generating set search (GSS), a class of algorithms for bound and linearly constrained optimization that obtain conforming search directions from generators of local tangent cones (Lewis, Shepherd et al. 2005; Kolda, Lewis et al. 2006). In its simplest form, the method evaluates the objective function on a stencil of points and if a better point is found, the stencil is moved to that point, otherwise the size of the stencil is reduced. Optimization is terminated either based on a function evaluation budget or when the stencil becomes sufficiently small. The basic GSS algorithm is:

A simple GA can be outlined with: 1. 2. 3.

Generate a random/seeded initial population of size ݊௣ Evaluate the fitness of individuals in initial population Iterate through the specified number of generations: a. Rank fitness of individuals b. Perform selection c. Perform crossover and mutation d. Evaluate fitness of newly-generated individuals e. Replace non-elite members of population with new individuals

265

Let ‫ݔ‬଴ be the starting point, ȟ଴ be the initial step ଶ௡ size and ࣞሼ݀௜ ሽ௜ୀଵೝ be the set of positive spanning directions. While not converged Do: 1.

Generate trial points ෩ ෩௞ ‫א‬ ܳ௞ ൌ ൛‫ݔ‬௜ ൅ ȟ௞ ݀௜ หͳ ൑ ݅ ൑ ुሽ where ȟ ሾͲǡ ȟ௞ ሿ denotes the maximum feasible step along ݀௜ .

2.

Evaluate trial points (possibly in parallel).

3.

If ‫ݔ׌‬௤ ‫ܳ א‬௞ such that ݂൫‫ݔ‬௤ ൯ െ ݂ሺ‫ݔ‬௞ ሻ ൏ ߙȟଶ௞

size is either left unchanged or increased. In contrast, when the iteration was unsuccessful, the step size is necessarily reduced. A defining difference between the basic GSS and APPS is that the APPS algorithm processes the directions independently, and each direction may have its own corresponding step size. Global convergence to locally optimal points is ensured using a sufficient decrease criterion for accepting new best points. A trial point ‫ݔ‬௞ ൅ ȟ݀௜ is considered better than the current best point if ݂ሺ‫ݔ‬௞ ൅ ȟ݀௜ ሻ െ ݂ሺ‫ݔ‬௞ ሻ ൏ ߙȟଶ , for ߙ ൐ Ͳ.

Then ‫ݔ‬௞ାଵ ൌ ‫ݔ‬௤ (successful iteration) Else ‫ݔ‬௞ାଵ ൌ ‫ݔ‬௤ (unsuccessful iteration) and ȟ௞ାଵ ൌ

୼ೖ ଶ

(2)

Because APPS processes search directions independently, it is possible that the current best point is improved before all the function evaluations associated with a set of trial points ܳ௞ have been completed. These results are referred to as orphaned points as they are no longer tied to the current search pattern and attention must be paid to ensure that the sufficient decrease criteria is applied appropriately. The support of these orphan points is a feature of the APPS algorithm which makes it naturally amenable to a hybrid optimization structure. Iterates generated by alternative algorithms can be simply be treated as orphans without the loss of favorable theoretical properties or local convergence theory of APPS.

(step size reduction)

The majority of the computational cost of pattern search methods is the ʹ݊௥ function evaluations, so parallel pattern search (PPS) techniques have been developed to perform function evaluations simultaneously on different processors (Dennis and Torczon 1991; Torczon 1992). For example, for a simple two-dimensional function, consider the illustrations in Figure 1 taken from (Gray and Fowler 2011). First, the points ݂ǡ ݃ǡ ݄ǡand ݅ in the stencil around point ܿ are evaluated. Then, since ݂ results in the smallest function value, the second picture shows a new stencil around point ݂. Finally, in the third picture, since none of the iterates in this new stencil result in a new local minima, the step size of the stencil is reduced.

2.3. Why EAGLS works EAGLS combines the NSGA-II with the APPSPACK software (Gray and Kolda 2006). APPSPACK is written in C++ and uses MPI (Gropp, Lusk et al. 1996; Gropp and Lusk 1996) for parallelism. Function evaluations are performed through system calls to an external executable which can write in any computer language. This simplifies its execution and also makes it a good candidate for inclusion in a hybrid scheme. Moreover, it should be noted that the most recent version of APPSPACK can handle linear constraints (Kolda, Lewis, and Torczon 2006; Griffin, Kolda and Lewis 2008), while a software package called HOPSPACK builds on the APPSPACK software and includes a GSS solver that can handle nonlinear constraints (Griffin and Kolda 2010a; Plantega 2009). To implement EAGLS, as in (Griffin and Kolda 2010b), a preliminary version of HOPSPACK was used. The EAGLS algorithm is designed to exploit parallelism. A goal of a parallel program is to ensure that all available processors are continuously being used. However, in practice this is often not the case. To understand this more fully in our context, consider a hypothetical black-box bound constrained optimization problem that has two real variables and an objective function with an evaluation time of at least one hour; further, we assume the user has 128 nodes with 2 processors each. There are a number of advantages that come from the use of parallelism in this context. x Most local search algorithms (even asynchronous parallel ones) have a cap on the

Figure 1: Illustration of the steps of Parallel Pattern Search (PPS) for a simple two-dimensional function. On the upper left, an initial PPS stencil around starting point ܿ is shown. In the upper right, a new stencil is created after successfully finding a new local min f . On the bottom left, PPS shrinks the stencil after failing to find a new minimum Note that in a basic GSS, after a successful iteration (one in which a new best point has been found), the step

266

x

replaceሺ‫ݔ‬ǡ ‫ݖ‬ሻpairs in the GA population pool with ሺ‫ݔ‬ොǡ ‫ݖ‬ሻ, where ‫ݔ‬ො is an improved estimate of ‫ כ ݔ‬provided by a local search. The GA still governs point survival, mutation, and merging as an outer iteration, but, during an inner iteration, individual points are improved via APPS applied to the real variables, with the integer variables held fixed. For simplicity, consider the parallel synchronous EAGLS algorithm with k local searches:

maximum number of processors they can effectively use. For our example problem, APPS will generate at most 4 trial-point per iteration. For the first hour APPS is called 252 processors will be idle. The user would need to start by hand 64 different instances of APPS centered at unique starting points, to fully exploit the computational power at hand with APPS alone. Most algorithms are synchronous by design, and parallel versions typically run in a “batch” mode. For example, a genetic algorithm requires all points in the current generation be evaluated before creating the next. Suppose a parallel GA uses a population of size 256 and submits all 256 points to be evaluated in parallel. Before the second iteration can begin, all 256 points must be evaluated; if all evaluations are complete but one, then the entire optimization processes is halted until this final evaluation is completed, even if this remaining evaluation takes hours longer to complete. Thus synchronous parallel algorithms necessarily move at the rate of the slowest evaluation.

The downsides described in the preceding bullets are actually advantageous for hybrid algorithms. Rather than attempt to redesign APPS so that it will submit more points in each iteration or invent a new asynchronous genetic algorithm that seeks to update multiple generations asynchronously, we simply tie multiple algorithms together loosely, pooling the resources in such a way that any unused resources can be shared. In the case of EAGLS, a single GA is run, and remaining idle processors are used to perform local searches. However, because local searches are often much faster than a GA at finding a local minimum, priority is given each generation to the local searches in the evaluation queue until that iterations local search evaluation budget has been expended. An immediate consequence and benefit of the EAGLS structure is that there is virtually no cap on the maximum number of processors that can be utilized for a given problem. At the same time, even with a few extra processors, significant wall-clock gains can be achieved, as the local search can be used to quickly find the global minimum once the GA is sufficiently near.

1.

Evaluate initial population in parallel

2.

While not converged Do a.

Choose a subset ࣦ of ݇ points from current population for local search

b.

Simultaneously run ݇ instances of APPS centered at points in ࣦ

c.

Replace respective points with their optimized values

d.

Perform crossover

e.

Evaluate new GA points in parallel

selection,

mutation,

Figure 2: In EAGLS the genetic algorithm optimizes over both integers and real variables, while local search instances work solely within a given integer plane (Griffin, Fowler, Gray, Hemker, and Parno). To select points for the local search, EAGLS uses a ranking approach that takes into account individual proximity to other; better points (see Figure 2). The goal of this step is to choose promising individuals representing distinct integer subdomains. The EAGLS algorithm allows the local search and the GA to run simultaneously using the same pool of evaluation processors. For the most part, the GA and each local search run asynchronously. However, after each GA generation, a new batch of local searches are created and given priority in the evaluation queue. This implies that given an adequate number of local search instances, the GA generations and local search generation will necessarily be nested, as the number of local search trial-points will always be greater than the number of available processors in the evaluation queue. This forces the GA to wait until the local search generation depletes its current evaluation budget prior to proceeding. Once the GA population has been evaluated, the local

2.4. EAGLS Algorithm EAGLS uses the GA's handling of integer and real variables for the global search, and APPS's handling of real variables in parallel for local search. Note that a MINLP could be immediately reduced to an integer programming problem if there was an analytic formula that provided ‫ כ ݔ‬where ‫  כ ݔ‬ൌ ƒ”‰ ‹ ݂ሺ‫ݔ‬ǡ ‫ݖ‬ሻ ௫

given an integer variable ‫ݖ‬. Though for a general MINLP, such a formula may not exist, local searches can be used (in parallel) to repeatedly

267

ܿଵ ሺ‫݌‬ሻ  ൌ  ‫ݔ‬ଵଶ ൅ ‫ݖ‬ଵ െ ͳǤʹͷ ൌ Ͳ ܿଶ ሺ‫݌‬ሻ  ൌ  ‫ݔ‬ଶଵǤହ ൅ ͳǤͷ‫ݖ‬ଶ െ ͵ǤͲͲ ൌ Ͳ ܿଷ ሺ‫݌‬ሻ ൌ  ‫ݔ‬ଵ ൅ ‫ݖ‬ଵ െ ͳǤ͸Ͳ ൑ Ͳ ܿସ ሺ‫݌‬ሻ ൌ ͳǤ͵͵͵‫ݔ‬ଶ ൅ ‫ݖ‬ଶ െ ͵ǤͲͲ ൑ Ͳ ܿହሺ௣ሻ ൌ  െ‫ݖ‬ଵ െ ‫ݖ‬ଶ ൅ ‫ݖ‬ଷ  ൑ ͲǤ

searches begin and operate asynchronously. To avoid re-evaluating points, all function values are stored in cache. The external parallel paradigm is nearly identical to that used in (Griffin and Kolda 2010b; Gray, Griffin et al. 2008). Whenever an improved point is found with respect to the real variables, the corresponding population member is immediately updated. See Figure 3 for a short point-flow sketch of this process.

The constraints on both the integer and real variables make this problem challenging. For constraint handling, we use the κଵ and the κଵ -smoothed penalty function where the constraint violation is incorporated with the objective function to form a corresponding merit function (Griffin and Kolda 2010b). Although the problem is small dimensionally, it is non-convex and some of the sub problems obtained by fixing the integer variables contain a unique local minimum which is challenging for standard MINLP solvers to avoid, as shown in (Kocis and Grossman 1988). Thus, this problem was ideal for testing the integer capabilities of EAGLS (Griffin, Fowler, Gray, Hemker, and Parno) and thereby was chosen here to study the asynchronous parallel local search capabilities. The known solution has a function value of 7.667 and the local minimum has a value 7.931. To add computational expense to each function evaluation and test the asynchronous nature of the algorithm, we add a random pause between one and three seconds to each function evaluation. This approach was used to test parallel optimization approaches in (Hough, Kolda, and Torzan 2001; Griffin and Kolda 2010b).

Figure 3: The EAGLS user can decide the population size and the number of local searches in an input file. The algorithms are run asynchronously in parallel with the local searches periodically inserting new improved points into the current GA population.

3.2. Algorithmic Parameters and Platform Since the solution to the test problem is known, we stop when the best point found is within 1% of the known solution. We provide the other relevant optimization parameters in Table 1. The numerical experiments were performed on a 102 processor Beowulf blade cluster (IBM e1350) with 3.0 Ghz Intel Xeon processors and Myrinet Networking.

3. NUMERICAL RESULTS 3.1. Test Problem To evaluate the parallelism of EAGLS we consider two studies. In the first, we fix all the optimization algorithm parameters and increase the number of processors used. In the second, we fix the number of processors to 16 and vary only the number of local searches while all other optimization parameters are held fixed. We use a classical mixed-integer test problem taken from (Kocis and Grossman 1988) that was proposed to study process synthesis applications with the outer approximation method. While this may seem to be simple, it is representative of the MINLPs encountered in process design and engineering. Thus, understanding how the parallelism and local search components of EAGLS affect its solution will aid in our ability to more efficiently solve similar MINLPs. The decision variables are ‫ ݌‬ൌ ሺ‫ݖ‬ଵ ǡ ‫ݖ‬ଶ ǡ ‫ݖ‬ଷ ǡ ‫ݔ‬ଵ ǡ ‫ݔ‬ଶ ሻ் with bound constraints given by

Table 1: Optimization Parameters

‫ א ݌‬ȳ ൌ ሼ‫݌‬ȁ‫ݖ‬ଵ ǡ ‫ݖ‬ଶ ǡ ‫ݖ‬ଷ ‫ א‬ሼͲǡͳሽǡ ‫ݔ‬ଵ ǡ ‫ݔ‬ଶ ‫  א‬ሾͲǡͳͲሿሽǤ We seek to minimize the objective function ݂ሺ‫݌‬ሻ where ݂ሺ‫݌‬ሻ  ൌ ʹ‫ݔ‬ଵ ൅ ͵‫ݔ‬ଶ ൅ ͳǤͷ‫ ͳ ݖ‬൅ ʹ‫ݖ‬ଶ െ ͲǤͷ‫ݖ‬ଷ

(4)

(3)

subject to the following constraints,

268

Parameter

Value

Population size

40

Number of Generations

250

Real Crossover Probability

0.9

Real Mutation Probability

0.5

Binary Crossover Probability

0.9

Binary Mutation Probability

0.0125

GSS Contraction Factor

0.5

GSS Sufficient Decrease Factor

1e-9

GSS Step Tolerance

1e-5

Maximum Generation Evaluations

840

Maximum Function Evaluations

3000

3.4. Varying Number of Local Searches To further understand how the asynchronous nature of APPSPACK impacts the search phase of EAGLS, we vary the number of local searches. For these experiments, 16 processors were used and all optimization algorithmic parameters were fixed except the number of local searches, which was varied from 4 to 8. We also consider the case of no local searches, which means EAGLS is simply a genetic algorithm with function evaluations performed in parallel. Figure 5 shows the average run times and number of function evaluations needed for convergence.

3.3. Varying Number of Processors Since the GA has stochastic optimization parameters and APPS is asynchronous, EAGLS is not a deterministic method, thus each optimization experiment was run five times and average values are reported. This approach has been used in numerous studies for APPS (Griffin and Kolda 2010b). Average run times and number of function evaluations required for convergence are shown in Figure 4 as the number of processors doubles from 2 to 64. For these experiments EAGLS used 8 local searches. Since there are only ݊௥ ൌ ʹ real variables, for each local search APPSPACK would not see increased speed up beyond ʹ݊௥ ൌ Ͷ processors for a total of 32 while the additional processors can be used to evaluate the GA population. The figure on the left shows the speed-up one would expect. The figure on the right is interesting in that the number of function evaluations increases with the number of processors. This is because as APPSPACK is run on more processors, the algorithm may move the stencil to a new location if a point is found with a lower function value but older points are not deleted from the queue if sufficient processors are allocated. So if a point from an older stencil does return a lower function value, the algorithm would move back to that location and continue. Note that because significantly more processors are being used, the computational time still shows linear speed-up despite the increased number of function evaluations.

Figure 5: Computational time and number of function evaluations required as the number local searches varies. Run times are shown in the upper picture and number of function evaluations are shown in the lower picture. The local searches have a significant impact on the optimization history using roughly one fifth of the computational effort of the GA alone. As the number of local searches increases, the number of function evaluations increases as one would expect but it is not significant. This is due in part to the fact that the algorithm is terminating based on proximity to a known solution. Future work will include exploring the behavior on larger dimensional problems which may show more dynamic results in terms of the optimal

Figure 4: Computational time and number of evaluations required as the number of processors varies. Run times are shown in the upper picture and number of function evaluations are shown in the lower picture.

269

Hydraulic Capture Community Problems. Adv. Water Resourc., 743-757. Fowler, K., & C.T., K. (2004). Solution of a Well-Field Design Problem with Implicit Filtering. Opt. Eng. Goldberg, D. (1989). Genetic algorithms in search, optimization, and machine learning. Addison Wesley. Gray, G., & Fowler, K. (2011). Traditional and Hybrid Derivative-free Optimization Approaches for Black-box Optimization. In Computational Optimization and Applications in Engineering and Industry. Springer. Gray, G., & Griffin, J. (2008). HOPSPACK: Hybrid optimization parallel search package. Livermore, CA: Sandia National Labs. Gray, G., & Kolda, T. (2006). Algorithm 856: APPSPACK 4.0: Asynchronous Parallel Pattern Search for Derivative-Free Optimization. ACM TOMS. Gray, G., Fowler, K., & Griffin, J. (2010). Hybrid Optimization Schemes for Simulation Based Problems. Procedia Comp. Sci., 1343-1351. Grefenstette, J. (1986). Optimization of Control Parameters for Genetic Algorithms. IEEE Trans. Sys. Man Cybernetics. Griffin, J., & Kolda, T. (2010). Asynchronous parallel hybrid optimization combining DIRECT and GSS. Optim. Meth. Software. Griffin, J., & Kolda, T. (2010). Nonlinearly-constrained optimization using heuristic penalty methods and asynchronous parallel generating set search. Appl. Math. Res. eXpress. Griffin, J., Fowler, K., Gray, G., Hemker, T., & Parno, M. (n.d.). Derivative-free Optimization via Evolutinary Algorithms Guiding Local Search (EAGLS) for MINLP. Pacific Journal of Optimization. Griffin, J., Kolda, T., & R., L. (2008). Asynchronous Parallel Generating Set Search For LinearlyConstrained Optimization. SIAM J. Sci. Comp. Gropp, W., & Lusk, E. (1996). User's Guide for mpich, a Portable Implementation of MPI. Mathematics and Computer Science Division, Argonne National Lab. Gropp, W., Lusk, E., Doss, N., & Skjellum, A. (1996). A high-performance, portable implementation of the MPI message passing interface standard. Parallel Comput. Holland, J. (1975). Adaption in Natural and Artificial Systems. University of Michigan Press. Holland, J. (1975). Genetic algorithms and the optimal allocation of trials. SIAM J. Comput. Hough, P., T.G, K., & Torczon, V. (n.d.). Asynchronous Parallel Pattern Search for Nonlinear Optimization. SIAM J. Sci. Comput., 2001. Karr, C., & Freeman, L. (1998). Industrial Applications of Genetic Algorithms. CRC Press. Kocis, G., & Grossman, I. (1988). Global Optimization of Nonconvex Mixed-Integer Nonlinear

number of local searches, but for this work we are staying in the context of simulation-based MINLPs which typically are not too large. We should further note that this test problem does have a feasible local minimum with a function value of roughly 7.931, and EAGLS avoided convergence to this suboptimal point in all trials. 4. CONCLUSIONS These experiments are the first step in understanding an asynchronous hybridization of a genetic algorithm with a local search based on a generating set search method for mixed-integer problems. This approach has extended the APPSPACK software to handle integer variables, improved its global search capabilities, and added parallelism and a local search to the NSGA-II software package. The tests done here are promising in showing that using local searches can help accelerate the convergence of the GA but also indicate that there is a complex interaction among algorithm parameters. The GA is well-known to be sensitive to parameter settings and the addition of an asynchronous local search with additional parameters warrants a more extensive study to better guide users. Future work will include a sensitivity study similar to that in (Matott, Bartlelt et al. 2006) to understand the interaction and main effects of the optimization settings. ACKNOWLEDGMENTS This work was made possible by support from the American Institute of Mathematics. REFERENCES Alba, E. (2005). Parallel Metaheuristics. John Wiley & Sons, Inc. Blum, C., Blesa Aquilera, M. J., Roli, A., & M., S. (2008). Hybrid Metaheuristics. Springer. Conn, A., Scheinberg, K., & Vincente, L. N. (2009). Introduction to Derivative Free Optimization. SIAM. Deb, K. (2000). An efficient constraint handling method for genetic algorithms. Computer Methods in Applied Mechanics and Engineering. Deb, K., & Goel, T. (2001). Controlled Elitist Nondominated sorting genetic algorithms for better convergence. Proceedings of the First International Conference on Evolutionary Multi-Criterion Optimization {EMO} 2001. Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002). A Fast and Elitist Multi-Objective Genetic Algorithm: {NSGA-II}. {IEEE} Transactions on Evolutionary Computation. Dejong, K., & Spears, W. (1990). An Analysis of the Interacting Roles of Population Size and Crossover in Genetic Algorithms. First Workshop Parallel Problem Solving from Nature. Springer-Verlag, Berlin. Dennis, J. E., & Torczon, V. (1991). Direct search methods on parallel machines. SIAM J. Optim. Fowler, K. e. (2008). A Comparison of Derivative-free Optimization Methods for Water Supply and

270

Programming (MINLP) Problems in Process Synthesis. Ind. Eng. Chem. Res. Kolda, T. (2004). Revisiting Asynchronous Parallel Pattern Search. Livermore, CA: Sandia National Labs. Kolda, T., Lewis, R. M., & Torczon, V. (2006). Stationarity results for generating set search for linearly constrained optimization. SIAM J. Optim. Lewis, R., Shepherd, A., & Torczon, V. (2005). Implementing generating set search methods for linearly constrained minimization. Williamsburg, VA: Department of Computer Science, College of William & Mary. Lobo, F., Lima, C., & Michalewicz, Z. (Eds.). (2007). Parameter settings in evolutionary algorithms. Springer. Matott, L., Bartlelt-Hunt, S., & Rabideau, A. F. (2006). Application of Heuristic Techniques and Algorithm Tuning to a multilayered sorptive barrier system. Environmental Science \& Technology. Mayer, A., Kelley, C., & Miller, C. (2002). Optimal design for problems involving flow and transport phenonmena in saturated subsurface systems. Advances in Water Resources. Plantega, T. (2009). HOPSPACK 2.0 User Manual (v 2.0.1). Livermore, CA: Sandia National Labs. Raidl, G. R. (2006). A unified view on hybrid metaheuristics. {HM06:} Third International Workshop on Hybrid Metaheuristics. Reed, P., Minsker, B., & Goldberg, D. (2000). Designing a competent simple genetic algorithm for search and optimization. Water Resources Research. Talbi, E. (2004). A taxonomy of hybrid metaheurtistics. J. Heuristics 8, 541-564. Torczon, V. (1992). PDS:Direct Search Methods for Unconstrained Optimization on Either Sequential or Parallel Machines. Houston, TX: Rice Univ. Zitzler, E., Deb, K., & Thiele, L. (2000). Comparison of Multiobjective Evolutionary Algorithms: Empirical Results. Evolutionary Computation Journal.

271

RECONSTRUCTION OF CLINICAL WORKFLOWS BASED ON THE IHE INTEGRATION PROFILE “CROSS-ENTERPRISE DOCUMENT WORKFLOW” Melanie Strasser (a), Franz Pfeifer (a), Emmanuel Helm (a), Andreas Schuler (a), Josef Altmann (b) (a)

Upper Austria University of Applied Sciences, Research & Development, Softwarepark 11, 4232 Hagenberg, Austria (b) Upper Austria University of Applied Sciences, School of Informatics, Communications, and Media, Softwarepark 11, 4232 Hagenberg, Austria (a)

{melanie.strasser, franz.pfeifer, emmanuel.helm, andreas.schuler}@fh-hagenberg.at, (b) [email protected]

Enterprise Document Sharing (XDS), Patient Demographics Query (PDQ), Patient Administration Management (PAM) and Patient Identifier Crossreferencing (PIX). These integration profiles are well suited to define administrative processes but are not usable in the context of clinical processes. Moreover, the reconstruction of patient pathways was based on PAM audit messages stored in the Audit Record Repository (ARR). Although the entries stored in the ARR are suited to reconstruct a patient’s way through a healthcare facility, a complete reconstruction of a clinical workflow failed. In this work we focus on the reconstruction of clinical workflows based on the Cross-Enterprise Document Workflow (XDW), an IHE integration profile which handles the workflow of documents in a clinical context. The presented approach is part of the research project IHExplorer (IHExplorer 2011), whose main objective is to support hospital operators and clinical process managers with a set of tools to monitor, analyze and visualize clinical transactions and workflows.

ABSTRACT Process-oriented workflow management in healthcare is a prerequisite to deliver high quality of care and to decrease treatment costs. Clinical workflow management systems (WfMS) cover the definition, execution and reconstruction of healthcare processes, which aims to identify main causes of high medical costs. Nevertheless, almost all existing clinical WfMS share one common drawback: they build upon proprietary hospital information systems (HIS). This paper presents a HIS independent workflow reconstruction approach based on the IHE integration profile "Cross-Enterprise Document Workflow" (XDW). XDW addresses the management of clinical, cross-enterprise workflows and makes use of a specific document, the Workflow Document. This document tracks every step of a workflow (e.g. ePrescription) including control information as well as input and output documents. Therefore it is ideally suited to reconstruct a fine-grained clinical workflow including associated documents and metadata. Furthermore this approach enables a nominal-actual comparison between a clinical workflow and a process definition.

2. METHODS This section describes the essential methods and standards used by the XDW-based process reconstruction approach.

Keywords: Integrating the Healthcare Enterprise, CrossEnterprise Document Workflow, Clinical Workflow Management System, Reconstruction of Clinical Workflows

2.1. Integrating the Healthcare Enterprise Integrating the Healthcare Enterprise (IHE) is an international initiative by healthcare professionals and industry to improve the integration and interoperability of medical information systems with standardized descriptions of medical use cases and the systematic use of well established communication standards like Health Level 7 (HL7) and Digital Imaging and Communications in Medicine (DICOM). IHE issues technical guidelines called integration profiles that describe clinical use cases with actors, which represent software systems or software components, and standard-based transactions,

1. INTRODUCTION AND MOTIVATION In our previous work (Strasser et al. 2011; Altmann and Mayr 2011) we focused on the definition and execution of administrative and clinical processes and the reconstruction of patient pathways based on Integrating the Healthcare Enterprise (IHE) (IHE International Inc. 2011) and Business Process Model and Notation 2.0 (BPMN) (Object Management Group Inc. 2011). First results showed that this approach has some limitations due to insufficient prospects on the part of IHE. First of all, the definition and execution of processes was based on selected IHE integration profiles, such as Cross-

272

representing the communication between IHE actors. Integration profiles provide instructions for software manufacturers to develop interoperable software systems (IHE International Inc. 2011).

3.

RECONSTRUCTION OF CLINICAL WORKFLOWS BASED ON XDW This chapter presents a novel clinical workflow reconstruction approach based on the IHE integration profile XDW. The functionality can be used in addition to the PAM-based reconstruction approach (Strasser et al. 2011) or standalone. Finally, the reconstructed clinical workflow can be compared to an existing BPMN 2.0 process definition.

2.2. Cross-Enterprise Document Workflow The IHE integration profile Cross-Enterprise Document Workflow (XDW) focuses on the management of crossenterprise clinical workflows and makes use of a specific document, the Workflow Document. This document administrates all documents related to one clinical workflow and handles the changing of document states (Zalunardo and Cocchiglia 2011). The Workflow Document is a structured document, characterized by tasks, representing a single step in the workflow. Every task results in the creation of a new document or represents a document state change. Each task has the same structure, based on three elements (see Figure 1): control information, input and output data. The control information element contains metadata needed to describe the specific step (e.g., author, date and time, organization). The input could be data or references to documents needed to perform the current step. The output of a task is a reference to one or more documents created during this step. The structure of the Workflow Document is kept general and extensible to take account of further use cases. The current revision of the XDW integration profile describes a large number of different use cases to cover as much scenarios as possible. The use cases are simplifications of real life scenarios, e.g. ePrescription and eReferral. The XDW integration profile is currently submitted for public comment and is a supplement to the IHE IT Infrastructure Technical Framework 7.0 (IHE International Inc. 2010).

3.1. Process definition and execution The initial point of the presented reconstruction approach is a process definition based on BPMN 2.0, describing a clinical pathway which in turn is executed on a WfMS. Each action performed during process execution creates a new entry in the XDW Workflow Document according to the structure presented in section 2.2. The execution of a process leads to a comprehensive record of actions performed. As the main field of application of the XDW integration profile are document-based workflows such as eReferral or ePrescription, the structure of the XDW Workflow Document contains references to other documents created during process execution as well as information about the authors and their organizations. A typical example of an action which creates a new entry in the Workflow Document is the creation of an electronic prescription (see Figure 2). A prescription depends on clinical information which is often provided by means of one or more documents. Therefore, the input section of the appropriate task in the Workflow Document is used to store references to the according input documents. Moreover, the result of the prescription placement task is a new document. This information is stored in the output section of the according task. For better comprehension and traceability of the approach this paper exclusively focuses on the IHE pharmacy process ePrescription which is described and illustrated with a sequence diagram in the XDW integration profile (Zalunardo and Cocchiglia 2011). Figure 2 shows the ePrescription use case as BPMN 2.0 process definition. 3.2. Workflow reconstruction Workflow Documents are updated every time a new task is executed, so all documents created during a patient’s treatment are referenced in the input- and output section of a task. The tasks in the Workflow Document can be sorted chronologically by using date and time of the control information. Due to the fact that XDW Workflow Documents are well-formed and valid XML documents, it is possible to use a standard mechanism to display and transform the documents. Extensible Stylesheet Language Transformation (XSLT) is a declarative, XML-based language used for the transformation of XML documents (W3C 2007).

Figure 1: The structure of a task in the XDW Workflow Document

273

Figure 2: The ePrescription use case as BPMN 2.0 process definition In this context XSLT is used to transform the XDW Workflow Document into a BPMN 2.0 process (see Figure 3). The transformation process relies on the correct implementation of the Workflow Document. Each task in the Workflow Document (e.g. Create Advice) is mapped to multiple BPMN 2.0 activities; threre are one or more activities to download clinical documents (e.g. Download Prescription and Workflow Document), one ore more activities to create or update clinical documents (e.g. Create Advice) and finally the upload of the modified documents (e.g. Upload Advice and Workflow Document). The activities are consecutively numbered to guarantee unique identifiers. For better readability it is important to define meaningful display names in the tasks code system, because these names are used to title the activities in the reconstructed process. In terms of codes and display names, there is still a huge amount of work to be done by the XDW technical committee, as there is currently no common code system available. Depending on the author of a task we distinguish between user tasks and service tasks. On the one hand an author can be human like a nurse or a doctor, so the task is obviously an user task, but on the other hand tasks may also be executed by a program e.g. a HIS, which identifies a task as a service task. To access the underlying IHE infrastructure with service tasks,

Figure 3: XSLT transformation to a BPMN 2.0 process existing actors of the Open Source framework Open Health Tools (Open Health Tools Inc. 2011) are used. To meet the requirements of the BPMN 2.0 standard definition at least two events must be implemented. The start and end events enclose the entire sequence flow. All activities between them are ordered chronologically in a straight line (see Figure 4).

274

Since a reconstructed BPMN 2.0 document represents only one patient pathway, it is linear and without any gateways. The reconstructed patient pathway can be visualized in any BPMN 2.0 editor, because it is available in a standard format. Furthermore functionality for actual-theoretical comparison of BPMN 2.0 documents is provided.

3.3. Process comparison In order to compare an document-based, clinical workflow with an existing, predefined BPMN 2.0 process definition one needs to have a tool to match two BPMN 2.0 documents. Currently there is no such tool available, so we decided to implement one on our own. The tool enables a nominal-actual comparison, by examining all the tasks listed in an existing process definition and comparing them with the display names given in the Workflow Document. Subsequently the resulting information can be used to emphasize the actual workflow in the process definition. Figure 5 shows the ePrescription process definition with the highlighted reconstructed workflow. 4. RESULTS The active cooperation with L. Zalunardo, the main author of the XDW integration profile, resulted in the creation of a conceptual design of the HL7 v3 CDA Workflow Document. The Workflow Document draft was introduced to the latest version of the XDW integration profile. The Workflow Document is ideally suited for document-based workflows, tracking document states, like document creation and document updates. All workflow steps are summarized in one standard-based, structured document. Moreover, the current status of a clinical workflow can be determined with the Workflow Document.

Figure 4: Reconstructed ePrescription process

Figure 5: A reconstructed BPMN 2.0 process definition based on a XDW Workflow Document

275

A further result is the detailed workflow reconstruction based on the Workflow Document (see section 3.2) which can be used as an extension to the previous process reconstruction work. As every task in the Workflow Document contains control information as well as input and output documents, a fine-grained, document-centered process reconstruction is available. As reconstructed workflows are available in the standard format BPMN 2.0, new possibilities for further processing open up. This approach, for example enables a nominal-actual comparison based on two BPMN 2.0 documents, e.g. a clinical workflow and a clinical process definition (see section 3.3). The result of the calculation, another BPMN 2.0 document, highlights the executed workflow in the process definition. This functionality enables delta analysis, as there are often discrepancies between the defined and the executed process.

ACKNOWLEDGMENTS The project IHExplorer is funded by the Austrian Research Promotion Agency, Tiani-Spirit GmbH, OÖGKK, X-Tention Informationstechnologie GmbH and the OÖ Gesundheits und Spitals AG. REFERENCES Altmann, J. and Mayr, H., 2011. e-Health: Die IT-Basis für eine Integrierte Versorgung. Linz: Wagner Verlag; Auflage 1. IHExplorer, 2011. Research Project of the Upper Austria University of Applied Sciences. Available from: http://ihexplorer.fh-hagenberg.at [June 2011] IHE International Inc., 2011. Integrating the Healthcare Enterprise. Available from: http://www.ihe.net [June 2011] IHE International Inc., 2010. IHE IT Infrastructure (ITI) Technical Framework, Volume 1 (ITI TF-1) Integration Profiles. Available from: http://www.ihe.net/Technical_Framework/upload/I HE_ITI_TF_Rev7-0_Vol1_FT_2010-08-10.pdf [June 2011] W3C, 2007. XSL Transformation (XSLT) Version 2.0. Available from: http://www.w3.org/TR/2007/REC-xslt2020070123/ [June 2011] Object Management Group Inc., Documents Associated With Business Process Model and Notation (BPMN) Version 2.0. Available from: http://www.omg.org/spec/BPMN/2.0 [June 2011] Open Health Tools Inc., 2011, Improving the world’s health and well-being by unleashing health IT innovation. Available from: http://openhealthtools.org/ [June 2011] Strasser, M., Pfeifer, F., Helm, E., Schuler, A., Altmann, J., 2011. Defining and reconstructing clinical processes based on IHE and BPMN 2.0. Accepted as full paper at the XXIII International Conference of the European Federation for Medical Informatics. August 28-31, Oslo. Zalunardo, L. and Cocchiglia, A., 2011. IHE IT Infrastructure (ITI) Technical Framework Supplement, Cross-Enterprise Document Workflow (XDW). Available from: ftp://ftp.ihe.net/IT_Infrastructure/iheitiyr9-20112012/Technical_Cmte/Profile_Work/XDW/IHE_ XDW_v16_23_03_11_draft.doc [June 2011]

5. CONCLUSION The presented approach shows the reconstruction of clinical workflows based on the IHE integration profile XDW. In the following, we want to point out the advantages of this new approach in comparison to the previous PAM-based reconstruction approach. Both approaches are hospital information system independent, because the reconstructions build upon IHE integration profiles. Certainly both approaches can be combined to receive further details about an executed workflow. The PAM-based reconstruction approach shows some limitations, mainly because IHE doesn’t provide enough integration profiles to reconstruct comprehensive, clinical workflows. The existing integration profiles are restricted to the definition of administrative healthcare processes. Furthermore the PAM-based appraoch cann‘t reconstruct holistic workflows without using the event logs of the WfMS. The assignment of clinical documents to a certain workflow is a time-consuming calculation and sometimes infeasible becasue of multiple parallel patient workflows. In contrast to the PAM-based reconstruction approach, XDW enables a complete reconstruction of a clinical workflow with an unique assignment of clinical documents to a specific workflow step. Furthermore the presented approach allows a fine-grained reconstruction, because the Workflow Document describes every workflow step in a high level of detail with control information and associated input and output documents. Moreover the reconstructed workflow is available in the standardized format BPMN 2.0. Next to the reconstruction, functionality was developed to make a comparison between an executed workflow and a process definition. To achieve further information about a certain workflow, the PAM events in the ARR and the tasks in XDW Workflow Document might be combined.

AUTHORS BIOGRAPHY Melanie Strasser finished her studies in „Information Engineering and Management“ at the Upper Austria University of Applied Sciences in 2009. Since her master thesis in 2008 she is a scientific researcher at the Research Center Hagenberg. Her research focuses on eHealth and IHE. Franz Pfeifer is a research associate at the Upper Austria University of Applied Sciences, Campus

276

Hagenberg. His research interests are informatics and digital image processing.

medical

Emmanuel Helm is a researcher at the Research Center Hagenberg and a student at the Upper Austria University of Applied Sciences. His research work concentrates on IHE focusing on the IHE compliant transmission of discrete and continuous data. Andreas Schuler finished his sutdies in “Software Engineering” at the Upper Austria University of Applied Sciences in 2011. Since September 2010 he is part of the research project IHExplorer and his work focuses on healthcare process management. Josef Altmann is head of the department „Communication and Knowledge Media“ of the Upper Austria University of Applied Sciences. Moreover, he is head of the research project IHExplorer. His research interests are in the area of component and service oriented software development as well as in the area of data and information integration.

277

A METHODOLOGY FOR DEVELOPING DES MODELS: EVENT GRAPHS AND SHARPSIM Arda Ceylan (a), Murat M.Gunal (b) (a)

Institude of Naval Science and Engineering Turkish Naval Academy, Tuzla, Istanbul, Turkey (b) Department of Industrial Engineering Turkish Naval Academy, Tuzla, Istanbul, Turkey (a)

[email protected]

(b)

[email protected]

1993). Simulation software adopts one of these worldviews. DES worldviews in the literature can be categorized into:

ABSTRACT In this paper, a methodology for fast development of Discrete Event Simulation (DES) models is presented. The methodology simply works in two stages. In the first stage the modeler builds a Conceptual Model (CM) of the system to be modeled. A CM is represented as an Event Graph (EG). EGs are used to document the events and their causes in a system. In the second stage the CM is translated to an Event Based DES model. To fulfill this task we developed a DES library, SharpSim, using C# (CSharp) programming language. This paper gives an introduction to our methodology. We provide an insight into SharpSim and EGs, and illustrate a modeling example.

• • • •

Process Interaction Activity Scanning Three Phase Event Scheduling

Process Interaction focuses on processes which can be described as “set of events” (Roader, 2004). In this approach, entity flows play the main role where flows include all states of objects. The process is described as “a time-ordered sequence of events, activities and delays that describe the flow of a dynamic entity through a system” (Carson, 1993). Process Interaction is popular and widely used since it is easier to conceive and implement, but “deadlock problem” (Pidd, 1998) stands as the weak point. This approach is used commonly by COTS simulation software, such as Automod. Another common approach, Flow Transaction, is a derivative of Process Interaction. Arena, ProModel and Witness are some of popular software using this approach (Abu-Taieh and Sheikh, 2008). In Activity Scanning, all activities are scanned in each time step and initiated up to their conditions. It is also called as Two Phase. Three Phase approach is a variant of Activity Scanning. It is more tedious to model, but faster since only conditional activities are scanned at each step. Event Scheduling requires the identification of events and their impacts on system’s state variables. This approach is most efficient but can be complicated to conceptually represent when the model size is big. In this paper, we particularly focus on Event Scheduling world-view. As a first step of our interest we review the methods for conceptual modeling, such as Event Graphs (EG). Secondly, we present a new DES library developed in C#: SharpSim. Additionally we give a general idea about some basics of DES and pertinent general purpose DES software in use, and then position the SharpSim in this picture. Finally, we

Keywords: Discrete Event Simulation Library, Event Scheduling, Event Graph, Simulation Software. 1. INTRODUCTION Due to its popularity in simulation world, plenty of Discrete Event Simulation (DES) software has been developed and this trend is likely to continue in the future. There are two main types of simulation software; those which aim at non-programmers (e.g. a graphical user interface which provides drag and drop facilities to build a model by simple mouse clicks) and others which require programming skills (e.g. extending a given source code library to write a full program). First type of software is Commercials-Off-The-Shelf (COTS) such as Arena, Simul8, and Flexsim and they reach a wider user community than the other type does. It is in fact for this reason why the first type dominates the market. The obvious difference of the two types is the user friendliness; one requires the knowledge of how the software is used, and the other requires special expertise, e.g. programming. It is noteworthy that it is dangerous to strictly separate the two types since most COTS software today provides limited programming features. In either type of simulation software, a DES is approached by a variety of worldviews. A worldview is described as a “modeling framework that a modeler uses to represent a system and its behavior” (Carson,

278

The translation of the EG in Figure 1 is as follows; “If condition (i) is true at the instant event A occurs, then event B will immediately be scheduled to occur t time units in the future with variables k assigned the values j”.

provide an EG of M/M/n queuing system and a short tutorial on how a SharpSim model can be built. 2.

A BRIEF REVIEW OF CONCEPTUAL MODELING METHODS AND EVENT GRAPHS Conceptual modeling in DES is an active research area and there is still no consensus among simulation modelers on its representation, although Onggo (2009) is an attempt in which unified conceptual modeling is discussed. There are a variety of methods to conceptualize the problems in hand in terms of logical flow of objects and events in the system. We review three methods here. The first method is the most commonly used; Process Flow Diagrams (PFD). PFDs focus on the flows of entities in a system and are used by most COTS simulation software. A PFD is created by simply placing drag-and-drop objects to represent processes and links between these processes to represent interactions between processes. The modeler, in a way, treats him or herself as an entity and follows the processes which transform an entity. The second method is Activity Cycle Diagrams (ACD) for conceptualizing the logical flows of objects in the system. In an ACD, life cycle of entities in the system is shown. In their life time, entities changes state and interact with each other. Entity states alternate from active to dead states. Simulation time moves forward and entities of the system spend time in these states. Active states represent activities which different types of entities can cooperate. Once an entity enters an active state, its duration can be determined, generally by taking a sample from a probability distribution. However some conditions must be satisfied for an entity to be in an active state, for example, if there is a server available and there is a client waiting in a queue, a customer entity enters to a service active state. Dead state is the opposite of an active state that is when an entity is idle or waiting for something to happen. This generally means a waiting area. Unlike an active state, duration of a dead state cannot be determined in advance since the time an entity spends in dead state is bound to preceding and succeeding activities. Finally, Event Graphs (EG) are used to conceptualize a system by focusing on its events. EGs work well with Event Scheduling approach since “Event Graphs are a way of representing the Future Event List logic for a discrete-event model” (Buss, 2001). There are two main components of EG; nodes to represent events and edges to represent transitions between events. Figure 1 shows the basic structure of EG (Roader, 2004).

3. SHARPSIM OVERVIEW In the second stage of our methodology an Event Graph is translated into computer code to build a simulation model. SharpSim is developed for this purpose. SharpSim is an open-source Discrete Event Simulation (DES) code library developed in C# (The code can be It accessed at http://sharpsim.codeplex.com). implements Event Scheduling world-view which involves three main classes; Simulation, Event, Edge and 3 secondary abstract classes; Entity, Resource and Stats. The objects instantiated from these classes are used to implement the EG drawn, as described in the first stage of our methodology. SharpSim is appropriate for multi threading. This is particularly helpful for animation, for example a simulation model running as a thread can communicate with animation classes, e.g. updating screen objects periodically. In this section we briefly explain how SharpSim works. 3.1. Simulation Class Simulation class is the core of SharpSim. It includes the main Event Scheduling algorithm and the thread that executes the model. Number of replications and seed number for random number generation are the parameters of this class. There are four properties of simulation class and their descriptions are as follows; • Future Event List (FEL): This collection involves the set of events that will be executed in the future. An instance of the event that is to be scheduled is inserted into FEL. FEL is sorted by the due time of the events, e.g. earliest event is on top of the list. Note that the event scheduling simulation algorithm scans FEL repeatedly until no events exist in FEL. After the execution of an event, it is removed from the list. • Clock: The clock variable keeps the simulation time. It is handled in Run method and proceeds to the execution time of the next event. • Events: This collection involves the set of events instantiated at the beginning of simulation and provides easy manipulation of events. Note that the events of a model are instantiated in the model class that is coded outside of SharpSim library. • Edges: This collection involves the set of edges instantiated at the beginning of the simulation and provides easy manipulation of edges as in the Events list. Simulation class includes two main and two supplementary methods. Main methods are described below. The two supplementary methods, Create Events and Create Edges, are useful for reading event and edge details direct from an Excel input file. Events and Edges are instantiated and added to Events and Edges collections.

Figure 1: A Basic Event Graph

279

• Run: The Event Scheduling algorithm is handled in this method. It involves a loop for each replication and another embedded loop for each event in the future event list. The first loop iterates for a number of replication times while the second embedded loop iterates till termination event is executed. In the second loop, first the clock is set to next event’s execution time, then event is executed and at the end of the execution it is removed from future event list. • Start Simulation Thread: This method is used to start simulation thread which is created when a simulation object is instantiated.

There is one method in the Event class, Event Executed, which is a delegate method associated with the next event. This is the point where C# event handling mechanism meets with the simulation’s events. When the due time of an event comes this method is called to schedule the next event linked to the current event being executed. The event schedule occurs if the condition on the edge is true. It clones a new independent event from the following event, provides parameter passing between edge and cloned event, and sets its execution time and finally insert cloned event into the FEL.

3.2. Event Class Event is an activity which causes a state change. The set of events together with edges forming a system is created at the start of the simulation and according with interrelations among events and edges new events are cloned and added to future event list during simulation. The constructor of this class has four arguments; event id, event name, priority, and event due time. If an instance of an event class is created with an event due time, the event is inserted to FEL directly. When more than one event has the same execution time, a second parameter is needed to decide which event will be executed first. Priority provides this secondary regulation. It is crucial to assign priorities on events particularly in complex systems. Properties of this class are explained below; • Execution Time: Each event has an execution time. The execution time of an event is mostly set during the simulation. • Parameter: This property is used to implement parameter passing on edges in EGs. When an event is executed, a parameter, either a single value such as an integer or an object such as a customer object, can be set into the next event. With this mechanism, for example, individual entities can be transferred from event to event. • Queue: This property is used to keep the entities that are waiting to be scheduled into the FEL.

3.3. Edge Class Edge is a link between two events. It defines relations between events and accordingly flow of the system. Scheduling of events is decided up to edge conditions. Furthermore, execution times of newly cloned events are set according with edge’s next event time value. The constructor of this class has three parameters; name of the edge, source event, and target event. Target events subscribe to source events. There are three properties of the edge class; next event time, attribute, and condition. Next event times can be deterministic or stochastic. Attribute is a variable which is set when parameters are passed between events. Condition is the condition of scheduling an event. The modeler can create entities and resources by inheriting from the entity and resource classes. Additionally Stats class provides an easy output manipulation for the simulation. 4. M/M/N SERVICE SYSTEM SIMULATION In this section, we aim to describe how a model of an M/M/n service system can be built using our methodology. As stated earlier, the first stage requires drawing an event graph of the system to be modeled. M/M/n queuing system’s event graph is drawn in Figure 2. There are four events and six edges in this graph.

Figure 2: Event Graph of M/M/n

280

The Event Graph (EG) in Figure 2 has four events which represent start of simulation, arrival of customers, start of service, and end of service events in a queuing system. The variables are ID (arriving customers’ ID number), S (number of available servers), and C (Customer Entity). The explanation of this EG is as follows; When the Run event occurs, set the ID to 1 (first customer’s ID), the S to n (there are n servers) and create C (an instance of Customer object). Setting the attribute of Edge [1-2] to C is required to pass the C object to the next event. When the Arrival event occurs you first need to add the receiving Customer object to the next event’s queue. Later you pull a Customer entity from the Start Event’s queue and set this to the edge’s attribute. Likewise, you need to create a new Customer instance and set it to the self loop attribute. Finally, you need to set the condition on edge between Arrival and Start events based on S (number of available servers). When the Start event occurs, first customer in the Start event’s queue is removed, since it is time for that customer to be served. Number of available servers decremented by one and the receiving customer entity is set as the parameter on the edge. This is to transfer the customer to the leave (end of service) event. Final event is Leave event. Executing a leave event means that a customer finished the service and therefore number of available servers (S) must be incremented by one and a new Start event must be scheduled. Scheduling a Start event is possible if number of customers waiting in the queue (Q) is non-zero. Q is the queue count of Start event (Event[3]).

which triggers when to stop the simulation, and has a due time. Termination event’s due time is the replication length of the simulation. After creating the events, you need to add State Change Listeners. State change listeners are related to C# event handling mechanism and help connect SharpSim events with C# form events, for example when the Run event occurs in SharpSim, Run method of the form is executed. An instance of Edge is instantiated for every edge in Figure 2. The Edge class has three parameters; name, source event, and target event. The modeller can set the time distribution and the distribution parameters on an edge by setting its “.dist” and “.mean” properties. After the definitions, a stats collection line can be written, such as the delay time between arrival and leave events. The delay between these two events means the total time of customers in the system. And finally, the run method of the simulation instance is called which causes the simulation to start. 4.2. State Change Handlers When a SharpSim event occurs, its corresponding method in the form is also executed. These methods are coded in the model file and inside these methods there are state change related codes, such as incrementing state variables and creating new entities. Inside a state change handler, it is essential to write code inside “evt.EventExecuted += delegate(object obj1, EventInfoArgs e){ ...}” block. For example for the Run event in Figure 2, write a “public void Run” method and inside the method write the delegate line and then set the ID variable to 1 which means that the very first arriving customer’s ID will be 1, set the S variable to 2 which means that we have initially 2 servers, create a new customer instance, and set the edge between Event 1 and 2 parameter value to this new customer.

4.1. Building a SharpSim model To build a simulation model in SharpSim, a C# project must be created. You need to create a Windows Forms Application project in a C# compiler such as Microsoft C# Express Edition 2008. The project must include the SharpSim library. SharpSim Library is a DLL file although full source code is provided and can be added to the project. On the default form in your project, generally named as Form1, a variable of type Simulation, five variables of type Event, and five variable of type Edge are defined. You need to add a button and a richtextbox components on to the form to create a basic user interface. The model will run when the button is clicked and an output will appear in the text area. On the button’s click event, firstly you need to instantiate your simulation model by calling Simulation class’s constructor. The constructor has three parameters; track list to show a default output screen, replication number, and seed number. Secondly, events are instantiated. An event has an ID, name, and priority parameters. Note that these simulation events have no due time given since all four events instantiated, and shown in Figure 2, are dynamic event. This means that their due time will be known during the simulation. On the other hand, the fifth event is the Termination event,

For every event in the model, there must be a State Change handler method in the code. 4.3. Running the model After buiding a SharpSim model as described above, the project is built and run. Clicking the button on the default form will start the simulation. This means that an instance of Simulation class, instances of events and edges will be created. Since the starting event is Run and scheduled to time 0, the model will start with this event. Execution of Run event will then cause scheduling an arrival event and in turn new arrivals will be created and so on. In the text area, text outputs will appear as simulation runs. Note that any message, simulation related or not, can be written to the text area on the form inside the state change handlers.

281

5.

CONCLUSION

In this paper we presented a methodology for building DES models. The methodology incorporates Event Graphs, as a conceptual modeling tool, and SharpSim, a new Discrete Event Simulation (DES) library. The DES library, SharpSim, is created using C# language and allows modelers to build DES models by programming in C#. The SharpSim library is an implementation of event scheduling simulation approach. It aims at translating event graphs into simulation models easily. SharpSim is open source and can be downloaded at http://sharpsim.codeplex.com. The website also includes tutorials on modeling examples. REFERENCES Abu-Taieh, E.M.O., El Sheikh, A.A.R. (2008). Methodologies and Approaches in Discrete Event Simulation Abu-Taieh, E.M.O., El Sheikh, A.A.R. Commercial Simulation Packages: A Comparative Study Buss, A. (2001). Technical Notes, Basic Event Graph Modeling. Carson, J.S. 1993. Modeling and Simulation Worldviews. Pidd, M. (1998). Computer Simulation in Management Science. Chicester, UK: John Wiley & Sons. Roeder, T.M.K. (2004). An Information Taxonomy for Discrete Event Simulations Rossetti, M. D. (2008) “JSL: An Open-Source ObjectOriented Framework for Discrete-Event Simulation in Java”, International Journal of Simulation and Process Modeling, vol. 4., no. 1, pp69-87, DOI: 10.1504/ IJSPM. 2008. 020614 Evans, W.A., 1994. Approaches to intelligent information retrieval. Information Processing and Management, 7 (2), 147–168. Schruben, L. (1983). Simulation Modeling with Event Graphs. Communications of the ACM, Volume 23, Number 11. Tocher, K. D. 1963. The art of simulation. London. English Universities Press. Onggo, BSS. 2009. Towards a unified conceptual model representation: a case study in healthcare. Journal of Simulation, 2009, 3, 40-49. AUTHORS BIOGRAPHY Arda Ceylan has received his MSc Naval Operations Research degree at the Institute of Naval Science and Engineering in Turkish Naval Academy. Murat Gunal is an assistant professor at the department of Industrial Engineering in Turkish Naval Academy.

282

REVISITATION OF THE SIMULATION METHODOLOGIES AND APPLICATIONS IN MANUFACTURING . Radha Ramanana and Ihsan Sabuncuoglub a. Assistant Professor, Mechanical Engineering Department, National Institute of Technology Calicut, Calicut – 673 601. Kerala, India b. Professor, Department of Industrial Engineering, Bilkent University, Ankara, Turkey (a) [email protected],\ (b) [email protected]

Exact solutions are available for most of the manufacturing systems. In spite of it, simulation remains as a widely used tool in finding a solution to a problem. This paper focuses on the application of simulation technology to manufacturing system design and manufacturing system operations. System design generally involves making long term decisions such as facility layout and system capacity/configuration. As such, models are typically created and used for single design exercise, and model run time is not a significant factor during the simulation process (Smith, 2003). On the other hand manufacturing system operations focus on day-to-day activities within the company and are typically made by lower-level managers. Decisions made at this level help to ensure that daily activities proceed smoothly and therefore help to move the

ABSTRACT Manufacturing is one of the largest application areas of simulation. For the purpose of understanding where, how and why the simulation is used in the manufacturing, this survey classifies the manufacturing system into two broad areas viz. manufacturing system design and manufacturing system operations. The two broad areas are further subdivided for this study. The survey discusses the evolution of the subdivisions before detailing the need of simulation in each of the sub divisions of the manufacturing systems. Finally, a discussion is made in order to understand where the research is heading for and identifying the future directions. Keywords: simulation, manufacturing system design, manufacturing system operations

2. 1.

MANUFACTURING SYSTEM DESIGN

In general, manufacturing system design problem (MSDP) encompasses the problem of facility location, plant layout, materials handling system design, assembly line balancing, and other ancillary functions necessary for the production of products. We discuss below the above sub divisions in detail.

INTRODUCTION

Simulation involves the development of descriptive computer models of a system and exercising those models to predict the operational performance of the underlying system being modeled. Simulation has been one of the most widely used tools for manufacturing (Banks et al. 2005).

2.1 Location Problems The generic term of facility is used to denote a large variety of entities such as warehouses, plants, antennas, hospitals and other industrial or public structures. The problem is to choose a set of points where these facilities are located so that the sum of location costs and transportation costs are minimized and satisfy the needs of all or part of the customers. The complexity stems from a multitude of qualitative and quantitative factors influencing location decisions as well as the intrinsic difficulty of making trade-offs among those factors. In general, the location problems are formulated as un-capacitated facilities location problem or simple facility location problem and capacitated facility location problem.

The basic components of manufacturing include product design, manufacturing/production, planning and control. The product design functions include, conceptualization, function identification, modeling and CAD, material selection, design for manufacturing and dimension and tolerance setting. The manufacturing operation includes processing, assembly, material handling, inspection and test. The planning function includes material requirement planning, capacity planning, process planning. The control function includes production scheduling, inventory control and tool management. For the sake of convenience, in this paper we consider the production, planning and control components all together as manufacturing operations.

283

The optimization problems defined above are mainly handled by deterministic and static approaches and these studies resulted in a number of valuable contributions to the area.There are a few studies such as, Hidaka and Okano (1997), Kurt and Scott (2007) that utilize simulation to investigate these trade offs. In general, the researchers employ the simulation tool to understand what if scenario. Simulation is used here either because data are not available or because of interactions that exist among many variables, such as customers, warehouse location, delivery time, transportation cost, fixed costs etc. involved in the decision making process.

The stochastic nature of some of the input factors such as demand, processing time, material flow, production schedules, travelling time, etc does not only increase the complexity of the problem but also necessitates a need for simulation or other stochastic optimization tools. Thus the material handling system design offers wide scope for simulation to test the different variables playing crucial role in the design of material handling systems and their interaction effect. The input factors (or variables) required for simulation of the material handling systems may be, the type of material flow, the type of material handling equipment, level of automation, machine schedules, shift patterns, the travelling distance, demand rate, desired throughput rate. Thus the material handling design offers wide scope for simulation to test the different variables playing an important role in the design of material handling system and also its interaction effects.

2.2 Facility / plant Layout Facility/plant design applications may involve modeling many different aspects of the production facility, including equipment selection/layout, control strategies (Push pull logic), material handling design, buffer sizing, etc. In general, the overall objective of facilities design is to get the inputs (material, supplies, etc.) into, through and out of each facility in the shortest time practicable, at acceptable cost. The material flow pattern becomes the basis for an effective arrangement of physical facilities. The facility layout problem is either formulated as a static layout problem or dynamic layout problems with optimizing the transportation or material handling cost as the primary objective. With this objective in mind different mathematical models have been proposed in the literature (Amine et al. 2007 and Balakrishnan and Cheng,1998). Simulation has found a large number of applications in the facility layout problems. Specifically, it is used for better understanding and visualizing the complexity of the problems as well as evaluating the system performance for alternative layouts. The complexity increases with increasing number of planning periods, stochastic flow patterns, stochastic demand patterns, unequal size of facility, different product mixes, etc. Some of the simulation studies that are found in the literature are Greasely (2008) and Harrell and Gladwin (2007). In these studies simulation is predominantly used as an interactive modeling and analysis tool to measure the performance of the system in terms of the work-inprocess, bottlenecks, routing complexity, the machine setup, machine down time, capacity etc.

2.4 Assembly Line Balancing Assembly lines consist of successive workstations at which products are processed. Workstations are defined as places where some tasks (operations) on products are performed. Products stay at each workstation for the cycle time, which corresponds to the time interval between successively completed units. There are a large number of methods proposed to solve these problems in practice. Bhattacharjee and Sahu (1990) discuss the complexity of the assembly line balancing problem. Some of the factors, such as, work content, cycle time, standard deviation of elemental times, TF-ratio, etc., which are responsible for the complexity of the line balancing problem, are identified and their effect on the complexity of the problem is discussed. Since the problem is NP hard, a number of heuristics are proposed to solve this problem (Sabuncuoglu, Erel and Tanyer (2000)). Boysen, Fliedner and Scholl (2008) provide a classification of ALB problems. While the ALB problems are generally formulated by static and deterministic models, the stochastic nature of demand, the transport times, processing times, set up times etc. necessitates the tools such as simulation. Su and Lu (2007), Mendes et al. (2005), Bukchin et al. (2002), Hsieh (2002) propose simulation to obtain optimum results for ALB problems. The researchers predominantly use simulation packages to evaluate the performance of the system and identify bottlenecks.

2.3 Material handling System Design The material handling system includes two highly inter-related sub-problems: (i) design of the material flow network that provides the resource interconnections; (ii) sizing of the transporter fleet, and allocation of the inter-group moves to these transporters. Sub-problem (ii) vehicle routing problem.

3. MANUFACTURING SYSTEM OPERATION Operational decisions focus on day-to-day activities within the company and are typically made by lower-level managers. Decisions made at this level help to ensure that daily activities proceed smoothly and therefore help to move the company toward reaching

284

the strategic goals. Examples of operational decisions include scheduling, handling employee conflicts, and purchasing raw materials needed for production. System operation involves making decisions on a much shorter time schedule. As such, the model is generally used much more frequently, and simulation run time is a more significant factor in the software / package selection and model design process. The classifications made here for the purpose of study are operations scheduling, lot-sizing and operating policies such as push/pull systems.

sizing problems. There are a number of review papers that study the lot-sizing under different classifications. The roles of simulation in lot-sizing are: to develop the inputs for the heuristics, understand the bottlenecks, understand the different operating conditions, impact of scheduling, understanding the capacity constraints etc. The complexity is sought to be modeled and a robust approach is made to minimize the impact of uncertainties using simulation. Researchers in the future will concentrate more on doing robust design of the demand and integrating the lot sizing with scheduling as the work in this area is also limited, but the need is highlighted by many researchers.

3.1 Scheduling Scheduling is the allocation of resources to tasks in order to ensure the completion these tasks in a reasonable amount of time. The objective of scheduling is to determine the job schedules that minimize (or maximize) a measure (or multiple measures) of performance. Literature has shown that only a few instances of the scheduling problems as polynomially solvable. The majority of the problems are NP hard. Some of the recent literatures (Mejtsky (2007), Metan and Sabuncuoglu (2010)) are reviewed to understand the need of simulation in scheduling. The roles of simulation in these applications are: to test the proposed heuristics in different scenarios or operating conditions, estimate the performance of schedules, identify bottlenecks or critical resources in the schedules, and generate input data for other heuristic or meta-heuristic algorithms to arrive at an optimized objective function values. As stated by Sabuncuoglu and Goren (2009), the future applications of simulation in scheduling still lie in the area of estimation and testing alternative solutions or schedules generated by scheduling algorithms in stochastic and dynamic environments. Simulation will also be used to determine appropriate scheduling or dispatching policies for manufacturing systems. In the recent years, robust optimization and scheduling have become very popular. Simulation has a potential to be used as a surrogate measure in these applications.

3.3 Control Logic In a typical manufacturing system a job moves from workstation to workstation. The control logic for managing this movement through the system can be based on push logic, pull logic or some combination. Special modeling features are required to accommodate each class, additional flexible constructs are required to represent the specific details and exceptions of the lower level control logic. There are no known available mathematical formulations for the control logic. Simulation seems to be the best way forward to evaluate the performance of the system. The input variables for control logic required may be the set up time, the number of transporters, demand rate etc and depend on the model construct. Enns and Suwanruji (2006) have summarized one group of recent simulation studies comparing replenishment strategies. Time-phased planning, implemented using DRP and MRP logic, continuousreview reorder point and single-card Kanban systems. There are at least two types of performance measure of interest, one related to the inventory level and the other to delivery performance. A tradeoff between these two types of measure exists. Therefore the problem is one of obtaining the desired performance across multiple performance measures (such as inventory level and delivery performance) through the selection of multiple interacting decision variables (such as lot size, reorder point)

3.2 Lot-sizing The lot size is the amount produced for each machine set up or the aggregate order size. Two very important dimensions of performance relate to inventory levels and customer delivery performance. The objective is to minimize total costs for the planning horizon while satisfying all demands, without backlogging. The literature is replete with a lot of mathematical models right from linear programming, integer programming, branch and bound procedures, dynamic programming, exact formulations like Wagner and Whitin algorithm.

Some of the simulation studies that are found are Enns (2007), Jula and Zschocke (2005), Krishnamurthy and Claudio (2005), Treadwell and Herrmann (2005). Simulation is used here to understand the system performances with respect to capacity, storage space, number of transporters and simultaneously collect data for decision making such as prioritizing and routing depending on the replenishment rate based on delivery performance.

Karimi, Fatemi, and Wilson (2003) describe the eight characteristics that affect the complexity of the lot

Loading of a facility requires complete tracking of all the resources and facilities, tracking of

285

the schedule of the events to occur, the operator allocation and subsequent delivery of the material. Owing to its complexity of many interacting factor no results can be claimed as optimal. Optimizing a control logic phase of the manufacturing operations has immense scope of future research

Balakrishnan J. and C.H. Cheng, “Dynamic layout algorithms: A state-of-the-art survey”, Omega 26 (4) (1998), pp. 507–521. Banks, J., J. S. Carson, B. L. Nelson, and D. M. Nicol, 2005. Discrete-event Simulation, Prentice-Hall, Upper Saddle River, NJ, New Jersey: PrenticeHall, Inc.

4. Discussions and Future Directions

Bhattacharjee T. K. and Sahu S., 1990. Complexity of single model assembly line balancing problems Engineering Costs and Production Economics 18(3), pp 203-214.

In this paper, the simulation studies in the manufacturing area are analyzed. A fairly comprehensive review is presented for the design and operational problems. The recent developments and applications of simulation are also discussed by identifying the future research directions.

Boysen N., Fliedner M. and Scholl A., 2008. Assembly line balancing: Which model to use when? International Journal of Production Economics, Volume 111(2), pp 509 – 528.

This survey indicates that manufacturing is one of the prime application areas of simulation. At the same time, simulation is one of the indispensible tools for manufacturing. Design problems are usually viewed as tactical or strategic decision problems that contain lots of randomness. Hence, stochastic simulation with appropriate output data analysis is generally required to estimate the long term or steady state performance of the systems. The general purpose simulation languages available in the market place today are quite sufficient to answer the design questions. In these applications, simulation is mostly used in the off-line mode as a stand-alone decision tool to enforce the decisions made by analytical or other models. Since the time is not the main constraint in this decision making environment, computationally demanding simulation optimization procedures can be used to make better decisions. Because the implication of false or incorrect conclusions from a simulation study can be disaster for a firm which has to make long-term design decisions.

Bukchin J., Ezey M. Dar-El and Jacob Rubinovitz, (2002) “Mixed model assembly line design in a make-to-order environment, Computers & Industrial Engineering, volume 41(4), pp 405-421 Enns S.T and Suwanruji Pattita, 2006. Observations on material flow in supply chains. Proceeedings of Winter simulation conference, Monterey, California December 03 - 06, 2006, Pages: 1446 – 1451. Enns Silvanus T. 2007. “PULL” replenishment performance as a function of demand rates and setup times under optimal settings” Winter simulation conference, Washington D.C. December 09 - 12, pp. 1624-1632 Harrell, Charles and Gladwin, Bruce, 2007. Productivity improvement in appliance manufacturing, Proceedings of Winter simulation conference, Washington D.C. December 09 - 12, 2007 pp 1610-1614.

In contrast, operational issues span relatively short time horizons. Hence, deterministic simulation (or stochastic simulation with a few random variables) is normally sufficient. Output data analysis and other statistical issues are not the main concern in these applications.

Hidaka, Kazuyoshi and Okano Hiroyuki,1997. Simulation-based approach to the warehouse location problem for a large-scale real instance, Proceedings of the 1997 Winter Simulation Conference, Atlanta, USA.

Since for the operational problems, simulation is used as an on-line tool, its integration to the existing decision support system is an important issue. Depending on the type of the application, web-based and/or distributed simulations may also be employed to improve the effectiveness of simulation studies. Virtual reality is also a challenge for the real-time applications of simulation in future studies.

Greasley A., 2008. Using simulation for facility design: A case study. Simulation Modelling Practice and Theory, 16 (6), pp 670-677. Karimi, B. Fatemi Ghomi S.M.T. and Wilson J.M. 2003, The capacitated lot sizing problem: a review of models and algorithms, Omega 31, pp. 365– 378. Krishnamurthy Ananth and Claudio David, 2005. Pull systems with advance demand information, Proceedings of the 37th conference on Winter

REFERENCES

286

simulation 2005, Orlando, Florida - 07, 2005, pp. 1733 – 1742.

December 04

Conference, pp. 183-188. July 24-28, Philadelphia (Pennsylvania, USA).

Kurt De Maagd and Scott Moore, 2007. Using Classifiers to Solve Warehouse Location Problems”, Proceedings of the 40th Hawaii International Conference on System Sciences.

Mercer, P.A. and Smith, G., 1993. Private view data in the UK. 2nd ed. London: Longman. Bantz, C.R., 1995. Social dimensions of software development. In: J. A. Anderson, ed. Annual review of software management and development. Newbury Park, CA: Sage, 502–510. Holland, M., 2004. Guide to citing Internet sources. Poole, Bournemouth University. Available from: http://www.bournemouth.ac.uk [accessed 15 July 2005] Das, A., 1992. Picking up the bills, Independent, 4 June, p. 28a. Agutter, A.J., 1995. The linguistic significance of current British slang. Thesis (PhD). Edinburgh University.

Mendes A.R., Ramos A.L., Simaria A.S. and Vilarinho P.M., 2005. Combining heuristic procedures and simulation models for balancing a PC camera assembly line, Computers & Industrial Engineering 49, pp. 413–431. Metan, G., Sabuncuoglu, I. and Pierreval, H., 2010, Real time selection of scheduling rules and knowledge extraction via dynamically controlled data mining, International Journal of Production Research (forthcoming). Mejtsky George Jiri, 2007, A metaheuristic algorithm for simultaneous simulation optimization and applications to traveling salesman and job shop scheduling with due dates. Proceedings of Winter simulation conference, Washington D.C. December 09 – 12.

AUTHORS BIOGRAPHY T. Radha Ramanan after his graduation in Mechanical Engineering pursued his post graduation in Industrial Engineering and PhD at NIT Trichy. His doctoral dissertation was in the area of Flow shop scheduling applying Artificial Neural Networks. He worked for a few years in industry before joining the academics. His areas of interest are operations management, supply chain management, production planning and control, and technology management. His research interests include sequencing and scheduling, lotsizing, simulation modelling etc. Four of his articles are published in referred international journals such as International Journal of Production Research, Journal of Intelligent Manufacturing, International Journal of Advanced manufacturing Technology. He is a life member of ISTE (Indian Society of Technical Education)

Sabuncuoglu, I., Goren S., 2009. Hedging production schedules against uncertainty in manufacturing environment with a review of robustness and stability research. International Journal of Computer Integrated Manufacturing Vol. 22 (2), pages 138 – 157. Sabuncuoglu I., Erel E. and Tanyer M., 2000. Assembly line balancing using genetic algorithms. Journal of Intelligent manufacturing, 11 (3), pp 295-310 Smith, Jeffrey S, 2003. Survey on the use of simulation for manufacturing system design and operation, Journal of Manufacturing systems, 22(2), pp. 157171.

Ihsan Sabuncuoglu is Professor and Chair of Industrial Engineering at Bilkent University. He received the B.S. and M.S. degrees in Industrial Engineering from Middle East Technical University and the Ph.D. degree in Industrial Engineering from the Wichita State University. Dr. Sabuncuoglu teaches and conducts research in the areas of simulation, scheduling, and manufacturing systems. He has published papers in IIE Transactions, Decision Sciences, Simulation, International Journal of Production Research, International Journal of Flexible Manufacturing Systems, International Journal of Computer Integrated Manufacturing, Military Operations Research, Computers and Operations Research, European, Journal of Operational Research, International Journal of

Su, Ping and Lu Ye 2007, Combining Genetic Algorithm and Simulation for the Mixed-model Assembly Line Balancing Problem, Third International Conference on Natural Computation (ICNC 2007). Treadwel A. l Mark and Herrmann Jeffrey W. 2005. A kanban module for simulating pull production in arena. Proceedings of the 37th conference on Winter simulation 2005, Orlando, Florida December 04 - 07, pp. 1413 – 1417. Bruzzone, A.G., Longo, F., 2005. Modeling & Simulation applied to Security Systems. Proceedings of Summer Computer Simulation

287

Production Economics, Production Planning, Control, Journal of Operational Research Society, Journal of Intelligent Manufacturing, Computers and Industrial Engineering. He is on the Editorial Board of International Journal of Operations and Quantitative Management, International Journal of Systems Sciences, Journal of Operations Management, and International Journal of Computer Integrated Manufacturing. He is an associate member of Institute of Industrial Engineering and Institute for Operations Research and the Management Science.

288

Insights into the Practice of Expert Simulation Modellers Rizwan Ahmed (a), Mahmood H. Shah (b) (a)

(b)

Lahore Business School University of Lahore Lancashire Business School, University of Central Lancashire (a)

[email protected], (b) [email protected]

Modellers in business and industry develop their models under a variety of constraints and contexts. The contextual factors may have an effect on the way modellers go about developing their models (Robinson 2002, Salt 2006). The problem domain, the scope of the problem, simulation language/technique/package used, the size and complexity of the problem simulated are some of the contextual factors which may affect a modeller’s approach to model development. Therefore it is important to reflect on how simulation context and practice relate with each other. Quite a few surveys have been reported in BPMS literature aiming to explore characteristics of modellers (Murphy and Perera 2001, Hollocks 2002), and practice (Melao and Pidd 2003, Cochran 1995), nevertheless, there is rare accounts of in-depth studies of modelling & simulation practice. These quantitative studies have provided useful indicators to understand characteristics of modellers and their backgrounds, nevertheless, these studies may not provide an in depth view of practice. One of the prominent in depth study of simulation modelling practice has been conducted by Willemain (1994, 1995), that explores the way expert modellers develop their models. Willemain (1994) studies the practices of expert modellers and suggests that practical guidelines for model formulation should be developed for novices in order to become experts. Foss et al (1998) reports a field study of industrial modelling process. Foss et al. (1998); interviewed 10 expert modellers and explored their process of simulation model development and proposed guidelines for improving simulation practice. This study empirically investigates as to how expert modellers develop their simulation models and how their context may affect their simulation practice. We believe that investigating the practices of expert modeller will enable further understanding of simulation practice and underpin the simulation methodology research. The paper has been organised in 6 sections. Section 2 gives an overview of the research methodology, Section 3 summarises study participants and their contexts. Section 4 discusses participants’ simulation practice and processes, Section 5 provides a discussion on the results and Section 6 concludes the paper.

ABSTRACT In this study we report the result of an empirical study investigating simulation modelling practices and processes of expert modellers in business and industry. The results suggest that most of the participants do not have a clearly defined or a formal process for developing their models, rather a set of key steps or stages depending on certain contextual factors and personal style. A number of contextual factors such as the problem domain, the scope of the problem, the size and complexity of the model, may affect the way a modeller goes about developing his/her simulation models. Generally a three phased approach is identifiable which can be named as problem definition, model development, and model usage. Model documentation largely depends on model life, client requirement, and type of model being developed. Maintenance and reuse of model is generally not practiced, given most of the models developed are of short to medium term use; however, experience and knowledge is something that is reused. Keywords: business process modelling, modelling practice, simulation context, modelling process

simulation simulation

1. INTRODUCTION We present the results from an interview study that investigates the practices of business process simulation modellers in order discover they underlying process of model development. Twenty expert simulation modellers selected from industry and academia described their simulation contexts and practices. Business process modelling & simulation (BPMS) generally lacks a rich body of literature reflecting on the modelling and simulation practices of modellers in real world. Successful application of modelling and simulation may depend very much on the personal practices of a simulation modeller (Willemain 1994). A huge number of case studies and personal anecdotes of successful application of simulation in different areas of business and industry can be found in simulation and modelling literature, however, little can be found in these studies as to how these modellers go about developing their models and simulation.

289

Table 1: Participants and modelling contexts Summary of Education and Professional Roles

Education summary

PhD Masters Bachelor Avg. Experience

14 3 3 8.5 years

Professional summary

role

Consultant (C) Researcher (R) C/R

9 5 6

Experience Summary of Model life, size, complexity, and Modelling Techniques Model Life Modelling Technique

Short-term 8 (40%)

Long-term 2 (10%)

Long/Short-term10 (50%)

DE: 8 (40%)

SD: 3 (15%)

Both DE and SD: 9 (45%)

Size Medium: 14 (70%) Small: 3 (15%) Large: 3 (15%) Complexity Low: 3 (15%) Medium: 12 (60%) High: 5 (25%) Summary of Types of Models Aims of models : Insights, cost and schedule, forecasting, Resource planning, allocation and evaluation Process improvement, Quality assurance, Understanding, Process performance monitoring and measurement, Process design Application area: Process change, improvement, and optimisation, Planning, Technology adoption, Project management, Education and training, Project control and operational management Problem domain: Safety control systems, Oil and gas pipelines, Mining, Supply chain and logistics, Airport processes, Call centres, Manufacturing, Financial services, Defence (weapons, vehicles), Telecom, Retail, Road and traffic, Health care, Software development processes, Scientific (physical, bioinformatics) Key: C=Consultant, R=Researcher, DE = Discrete event, SD = System dynamics, HB = Hybrid models, SB = State based

relevance on an initial draft and questions were improved on the basis of feedback by participants. Piloting the interview sessions with four other participants to evaluate the research instrument helped assessing the appropriateness of the structure and flow of the interview questions. It also helped testing and improving interviewing approach and provided valuable practice for the main set of interviews. The use of audio recording equipment was also evaluated. Moreover, it helped determine the time necessary for interviews.

2. METHODOLOGY This study follows a preliminary survey of 17 expert modellers (Ahmed et. al. 2008) which was an adaptation of Willemain’s survey. Insights from this survey instigated our interest in exploring the context and practices of expert modellers in depth. The results from survey allowed construction of a framework of ideas, relevant to the context and practices of simulation modellers, explored in this study. We wanted to study the context and practices of expert modellers in-depth and generally in a structured manner, therefore, we used semi-structured interviewing technique. Answers to the following research were explored with the participants:

3. THE PARTICIPANTS & THEIR CONTEXTS The participants in this study consist of both simulation practitioners and researchers. There are 20 participants in total coming from USA, UK, Germany, Spain and South Africa. Table 1 provides a summary of participants’ contexts. A thorough discussion on participants’ contexts has been provided in an earlier paper (Ahmed & Robinson 2007), however, here we will provide a summary of their contexts. The participants consisted of three groups; researchers (R), consultants (C), and researchers cum consultants (C/R); inclusion of both groups gives an insight both into the industry and academia. Table 1 shows that there are 14 participants with a PhD, 3 participants with Master degrees, and 3 participants hold Bachelor degree. This suggests that the participants in this study are highly educated and most of them had some modelling education as part of their professional or research degrees. The average experience of the participants in simulation is 8.5 years. This suggests a high level of simulation experience amongst the participants. The types of model developed by the participants have been classified with regard to their aims, application area, problem domain, size, complexity, and term of use.

RQ1: What are the modelling contexts of business process simulation modellers? RQ2: What are the modelling practices of business process simulation modellers? A pool of interview questions was prepared, consisting of some main open ended questions and several auxiliary questions which were to be asked depending on the flow of interview. A questionnaire consisting of open ended questions was sent to the participants a week prior to conducting the interviews. We also prepared an interview script document, which was used during the interview to ensure a generally uniform way of conducting interviews with all the participants. We also conducted an intensive pilot study to evaluate the interviewing instrument. This pilot study was conducted in two phases; first, pre-testing the interview questions validity and second, piloting the interview sessions. In the pre-testing, four participants evaluated each question for its understandablity and

290

simulation model, model design, construction, verification and validation, and experimentation. Table 2 shows that some of the participants tend to use software engineering terms such as requirements, requirements analysis, basic and detailed design, and testing. S2 describes a spiral approach to simulation model development and S8 describes an evolutionary and iterative approach. S4, S5 and S10 describe a process similar to the waterfall model of software development, with steps such as requirements gathering, analysis, design, implementation, and testing (validation and verification). S7 said that he/she has a completely ad-hoc approach to simulation model development with no specific process steps. S3, S4, S5 and S9 described their process in much more detail than the others. S3 and S4 develop highly complex models and S5 develop large models; perhaps this could explain the detailed natured of their process. Also S3 and S4 have experience of working both with discrete event and continuous simulation. S11, S13 and S14 described their process in a highly detailed manner. S15 and S19 described their process at a very low detail.

Most the participants develop process simulation models to study, plan, control, and manage the issues of cost, quality, and resources as shown in Table1. Table 1 shows that they mainly develop simulation models that fall in the application areas of process improvement, process understanding, project planning and management, technology adoption, and project/process control and operational management. Moreover, the participant have developed simulation models in the problem domains of airport processes, passenger flow, cargo, logistics, supply chain management, mining, oil and gas pipelines, call centres, manufacturing, telecom, financial sector, banks, healthcare policy planning, defence, and software development processes. Table also shows that most models developed by the participants are for short-term use, however, on rare occasions they have also developed models for longer term use. The model’s life of use may have an effect on the practices of simulation modellers (Ahmed & Robinson 2007), which will be described in the upcoming sections. Most of the participants have experience of working both with discrete event and continuous techniques. Only 3 participants have experience of using continuous simulation exclusively while 8 participants have worked exclusively with discrete event simulation. The participants use different tools for developing simulation models; Witness and Extend for discrete event and Vensim for system dynamics are the most popular tools amongst these participants. Participants claim that choice of simulation tool may have a positive or negative effect on the simulation practice of a modeller (Ahmed & Robinson 2007). They mostly develop simulation models of small and medium size. Also most of the participants develop simulation models of low or medium complexity. Most of the participants also believe that simulation model size and complexity are related, i.e. the bigger the simulation model, the higher the complexity will be, however, some participants also noted that a small model may also be very complex depending on the nature on a problem (Ahmed & Robinson 2007).

In Table 2 we summarise findings about the simulation modelling process practice of the participants. Apparently the simulation modelling process of the participants can be categorised into three phases as Problem Definition, Model Development, and Model Usage and Experimentation. Following we describe findings related to each phase and subsequently some other related themes. 4.1. Modelling Process Phase I: Problem Definition 1. Only three participants mentioned simulation user identification as a step in their process. The user can be the client or some other person in the organisation who needs results from the simulation study. They claim that establishing who the user of the simulation is very important to increasing confidence in the study results. This is because without close interaction with the user, a simulation study may not be of any value to its users. Moreover it is also important to identify the domain or subject matter experts with whom the simulation modeller may need to liaise during the model development. 2. Most of the participants indicate that the identification of simulation goals/objectives and simulation questions is one of their earliest steps in a simulation study. 3. Some of the participants used the term “requirements gathering” while talking about simulation goals and questions. This is perhaps because of their software engineering background. 4. Some participants (S7, S8, S12, S15, S20) do not spend much time on analysis and design, rather they identify simulation goals, gain a basic understanding of the problem and develop a simple and small simulation model straightaway, adding details as they go; a rapid approach.

4.

SIMULATION MODEL DEVELOPMENT PROCESS In this section we present an analysis of the simulation model development process of the participants. There are 35 themes identified from the interview transcripts which are relevant to simulation modelling processes. Each participant described his/her simulation modelling process at varying levels of detail. Each participant’s simulation modelling process has been summarised in a process matrix in Table 2. Most of the participants described their process in a linear fashion, emphasising that there is always a fair amount of iteration in their process. The main process activities described by the participants are problem communication with the client, defining simulation objectives and questions, problem understanding and analysis, definition of inputs and outputs from the

291

292

Design Basic design Detailed design Construction/implementation Model verification Model validation Calibration Testing

Design Experiments Conduct experiment Experiment results analysis Results presentation Maintenance

31 32 33 34 35

Initial contact with client Problem communication Quick sessions with customer Simulation user/domain expert identification Setting goals Questions Requirements gathering Req. Validation Identify and define model inputs Identify and define model outputs System/problem understanding and Scope Requirements/process Analysis Data/ analysis Conceptual modelling Conceptual model validation Influence diagram Scenarios Technical feasibility check Build prototype V&V of prototype Planning Tool selection

23 24 25 26 27 28 29 30

11 12 13 14 15 16 17 18 19 20 21 22

4 5 6 7 8 9 10

1 2 3

X

X

X

X X X X

X X

X X X

X

X

X

X

X X X

X

X

X

X X

X X X

X X X X X

X X

X

X X

X X

X

X

S5 S6 S7 S8 S9 S10 Phase – I: Problem Definition X X

X X

X

X

X X

X X X

X X

S11

X X Phase-II: Model Development X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X Phase III: Model Usage and Experimentation X X X X X X

X

X X

X X

X

X X X

S4

X

X

S3

X X

X

S2

X X X X X X

X X

S1

Table 2: Simulation Modelling Process matrix of the participants

X X

X X X

X

X X

X X X

X

S12

X X X X

X X X

X X X X X X

X X

X

X

S13

X X X X

X X X

X X

X

X X

X X

X X

S14

X X X

X

X

S15

X X X

X X X

X

X

X X

X

S16

X X

X

X X X

X

X

X X

X

X

S18

X X

X X X

X

X

X

X X

S17

X X X

X

S19

X

X

X X X

X

X X

X X

S20

three of these participants claim to be developing big and highly complex models. The results from our preliminary survey (Ahmed et. al. 2008) indicate that simulation model design is considered to be an issue, however, only a few participants in this study indicate that they do model design any formally. One possible explanation, as mentioned by S2, that the nature of simulation modelling does not require to devise a design prior to constructing the model; because most of the time in the early stages of modelling, neither client nor modeller understand the problem for which the model is to be designed; therefore it is difficult to design a model for which requirements are not clear. Another possible explanation could be that most of the simulation projects developed by these participants are small or medium which take a few days, weeks or months to develop; for such small projects as S8 says, it is not feasible to spend too much time on formally designing the simulation model. 14. All participants talked about building or constructing the simulation model using some simulation tool or programming language. Verification of the model is performed as the model is constructed. Most of the participants say that the whole simulation should not be constructed in one go, rather the validation of the model with the customer should be performed as parts of model are completed. During verification or validation, the modeller may discover some bug or problem with the model and may have to go back to develop further understanding of the problem. Almost all the participants emphasise that a modeller must provide sufficient comments in code or comment boxes while developing the model. This is crucial to understanding the model in case the modeller or some other person has to change the model at some later time. 15. Most participants consider validation and verification as equivalent to evaluation. Evaluation is driven more by customer satisfaction than any other factor. Moreover, some participants refer to model validation and verification in numerous ways such as testing, calibration and validation and verification.

5. Most of the participants emphasised developing a firm understanding of the problem and capturing the scope of the problem. They talked about identifying the factors contributing to a system/process, understanding relationships between different factors/variables, and confirming those relationships with the client/user. 6. Some of participants emphasised that diagramming methods should be used to illustrate relationships between various factors. This would not only enhance problem understanding but also helps validating the problem understanding with the client. 7. Most of the participants say that identification and definition of inputs and outputs of a simulation model is very important and should be started in the earliest stages of a simulation study. 8. Two participants mentioned conceptual modelling as part of their simulation process. Conceptual modelling in the general simulation literature is said to consist of detailed analysis of the problem and designing the simulation. Analysis would be a detailed account of all the activities performed for problem understanding, identification of variables and the relationship between them. Robinson (2004) defines a conceptual model as, “a nonsoftware specific description of the simulation model that is to be developed, describing the objectives, inputs, outputs, content, assumptions and simplifications of the model”. 9. Four participants mentioned checking technical feasibility; i.e. whether simulation is an optimum tool for answering the problem. Moreover, simulation may not be needed to solve certain simple problems; in such cases simulation would prove to be rather an expensive solution. 10. S8, S9, S11, S14, S16, S18, and S20 emphasised on prototyping or building an initial simple abstraction of the whole problem explicitly talked about prototyping. These participants think that building a prototype and then getting feedback from the client helps validate problem understanding and also in checking the feasibility of simulation tool. 11. Only one participant, R6, mentioned planning as a step in the simulation modelling process. S16 generally developed very big and highly complex defence simulation models with a team of people; perhaps this is the reason that he/she mentioned planning as an important step. 12. Simulation tools can positively or negatively impact the efficiency and performance of simulation modellers, according to S6, S7, and S17. None of the other participants mentioned tool selection as a part of their process.

4.3. Modelling Process: Phase III: Model Use and Experimentation 16. Most of the participants explicitly mention experimentation as part of modelling process. They describe that designing the experiments, analysing the results and presenting the results to the client are important tasks for conducting experiments with the simulation models.

4.2. Modelling Process Phase II: Model Development 13. Only a few of the participants mention simulation model design as part of their process. Only six participants talk about design as a process step;

4.4. Client contact and rapid development 17. Most of the participants emphasise heavy client contact. It is important to note that those who have emphasised heavy client contact are consultants or

293

different educational backgrounds.

researchers cum consultants. This is perhaps because in a research environment there is usually no client; therefore, the researchers do not mention heavy client contact as an important part of their process. 18. Most of the consultants indicate that in the commercial world it is very important to deliver a solution to the client very rapidly; because processes have to be adapted according to changing business need. If a simulation study takes months or years to deliver the results, it may not be of use to the client because during that time the business would have changed even further. Moreover, when the client is spending money on a simulation study, he/she wants to see the results instantly. Therefore, a simulation modeller must involve the client heavily and adapt his/her modelling process according to the client needs in order to deliver the results and recommendation quickly.

and

professional

4.6. Documentation practices 21. Most of the participants think that the best documentation for a simulation model is to put comments in the code or the comment boxes provided by the simulation tool rather than producing formal documents. 22. As shown in Chapter 6 (Table 6.10), most of the participants say that simulation goals and objectives should be clearly stated in the documentation (in comments or in formal documents) and be agreed upon with the client. However, a few of the participants also think that the scope of the model should also be defined in the documentation. 23. Some of the participants recommend that model inputs and outputs should also be defined so that the model can be well understood in future if needed. 24. Some of the participants think that the relationships between data items (inputs and outputs) should also be documented along with an influence/process diagram or using some other diagram methods. An overview of model structure or model working is also necessary to understand the model. 25. Most of the participants say that they produce reports or presentations of the simulation results which are presented to the client. These reports or presentation include the report of experiments, the scenarios and assumptions under which experiments have been run, analysis of results and recommendations from the analysis.

4.5. Individual Nature of Simulation Practice 19. All the participants say that they typically develop simulation models alone. However, they have to interact with the client, model users or the domain experts to understand the problem and collect data. Most of the participants say that sometimes they have worked and collaborated with other modellers; however, it seldom happens that they work on the same model concurrently. Only S16 says that he has worked and managed simulation model development where multiple people worked on the same model. However, in that case the project was an enormous defence simulation on which around 200 people worked. In other cases, as for instance S2 and S10 say, they worked with other modellers in a managerial role. S5, S9, S10, S11, S12, S13, say that they have worked on simulation projects in teams; however, in such situations roles such as simulation modeller, data collector, and process-mapper/system-engineer were well defined. 20. The participants give different reasons as to why simulation modellers tend work alone on a simulation study. One reason is that the nature of the simulation problems and the nature of modelling itself that do not require many people to work on the same project. Having more than one person introduces a time overhead because all the people involved have to have a similar level of understanding; S12 says this makes a project inefficient. S13 and S14 believe that having more than one person developing the same model introduces the problem of version/modification control and integration. In the view of S12, S15, and S16, the biggest problem in teamwork is the communication between different team members. S16 states that communication becomes even more problematic if the team members come from

4.7. Others 26. Model reuse for a similar problem is not important for most participants. This is because they think that a model developed at one point in the past may be not depict the real world as it is now; as R3 says “the business changes so much that the objects become out of date; I wonder if they are updatable”. However, some of the participants mention that the experience and learning gained from simulation projects is reused in subsequent projects. This finding is similar to what is found in literature that reuse in simulation is difficult therefore not much practiced (Robinson et al. 2004). Two of the participants, S4 and S6, mention that they reuse parts of their existing models. However, some participants said that it is the experience that is reused in subsequent simulation projects. 27. Majority of the participants do not emphasize simulation model maintenance. Only S9 explicitly mentions maintenance as part of the process; no one else discuss maintenance as part of their process. This is perhaps because majority of the models developed by the participants are of short-term use.

294

from scratch (Taylor et al. 2004, Robinson et al 2004). On the other hand, the importance of maintenance, reuse, and documentation has been highlighted by Gass(1987) for large scale models. Most of our participants also suggest that their models are rarely reused, however, the participants from military simulation background say that they emphasise model reuse. This finding supports the view of Salt (2006) where he suggests that defence modellers are obsessed with reuse while civilians do not bother reusing their models. One of the main reasons suggested by the participants is that reusing a simulation model is difficult; because most often simulation model represent a reality in business process at a given time but as the time passes the reality changes therefore an old model of that reality is of little use after the reality changes. However, the knowledge and experience gained from an old model can be reused in a new project. A similar view is held by the authors in (Taylor et al. 2004), Robinson et al 2004). Gass (1987) suggests that the evaluation of models encompasses both validation and verification activities along with an assessments of the models’ quality, usability, and utility. However, the results of our study suggest that this form of evaluation does not have a formal position in the simulation modelling practice of our participants. In general, simulation modelling literature seems to emphasize validation and verification activities; therefore, most possibly modellers consider this to be equivalent to evaluation. However, the extent of evaluation largely depends on contextual factors such as requirements. Another reason for not conducting a holistic evaluation by our participants could be that most models are used by themselves and results are provided to the client. Therefore, evaluating models in the aspects other than validation and verification is not of importance. However, if a model is to be handed over to the client; perhaps, evaluating usability and documentation is given some conscious consideration. The results from this study also suggest that majority of the participants in this study don’t seem to be using a highly defined formal process framework for their simulation modelling practice. However, most of them seem to have some specific steps, perhaps, unconsciously infused in their simulation modelling practice. Simulation modelling in commercial context involves people, technology and tools. A well-defined process is believed to provide a framework where tools, technology and people collaborate, to enhance productivity and quality (Humphrey and Kellner 1989). Humphrey (1997) states that a good process brings discipline in human activities and improve the quality of software. It is the process that can effectively help engineers to produce high quality products, with reduced time, and control over cost (Cugola & Ghezzi 1998). This suggests that a good simulation modeling process may also improve quality and increase the productivity. However, it is rare to find such studies in simulation modelling literature where relation has been

5. DISCUSSION These results provide a general picture of a model development practice of the participants and the type models they develop under a variety of contexts. The results indicate that most of the participants develop their models alone supporting the literature finding of Robinson (2002); however, for relatively larger projects a number of people may be working in different aspects such as problem understanding, data collection, model construction and validation and verification. Most participants do not produce design prior to constructing their models. A possible explanation as mentioned by one of the participant is that the nature of simulation modelling does not require to devise a design prior to constructing the model; because most of the time in the early stages of modelling, neither client nor modeller understand the problem for which the model is to be designed; therefore it is difficult to design a model for which requirements are not clear. Another possible explanation could be that most of the simulation projects developed by these participants are small or medium which take a few days or weeks; for such small projects as S8 says, it is not feasible to spend too much time on formally designing the simulation model. However, for large models designing prior to model development and adapting the design during development is a must Maintainability of models is not an issue for majaority of our respondents; however maintainability will inevitably become an issue if these models are to be capable of being evolved so that they remain useful in the long term. Our literature review suggests that the maintainability of models has not been given much attention in the general simulation literature; similarly in this study only a few participants indicate that they are concerned about maintainability. Maintenance and documentation are low priority issues. Another potential reason could be that perhaps the simulation models developed are too small (though they say they mostly build medium sized models as we have no agreed measure of size); or large but conceptually too simple to be documented and maintained. Another reason could be that most simulation models may not be used in the long term, therefore documentation and maintenance is not a problem. The participants believe that extent of documentation and maintainability varies in each individual case depending on the contextual factors such as client requirements, budget, time, expertise, and simulation model size and complexity. Issues of simulation model documentation and maintenance are also seldom discussed in the general simulation literature. Foss et al. (1998) say that most simulation models are poorly documented and are therefore rarely reused. The models evolve and are redefined over the period of time, and the managers who use the models may change their minds about priorities. Foss et al. (1998) further state that poor documentation makes it very hard to maintain the models. However, it is generally believed that reusing simulation models is difficult and less cost effective than building a new one

295

drawn between simulation model quality, modeller’s productivity and use of a disciplined process. On the other hand, Shannon (1975) says that simulation modelling is both art and science; producing art needs creativity (Kneller 1965), therefore simulation modelling needs creativity. Many simulation modellers believe that simulation is a creative accomplishment and if it is constrained by a process, creativity may suffer (Powell 1995). Paul et al. (2003) says, “One can instantly see that fixed structure to develop simulation models will not be able to cope with all the situations at all times”. This suggests that consider the context of a simulation modellers is important for applicability of a simulation modelling process. Simonton (2002) suggests that creativity can be considered a constrained stochastic process; that is creativity is not completely random or stochastic, rather loosely bound in the rules of the domain for which creativity is needed. Johnson-Laird (1988) says that there can be many criteria of creative processes on which a creator may rely; some of those criteria will be common to many practitioners while others may depend on individual aptitude and style. This suggests that creative process does not consist of only stochastic random activities but there is some structure in the creative process. Ferguson et. al. (1997) suggest why discipline is needed alongside creativity:

efficiently. Therefore if a process consolidated from real world simulation practice of expert modellers may provide discipline for productivity and quality and liberty and flexibility for creativity. A number of simulation modelling processes have been reported in the literature for example Robinson (2004), Law and Kelton (2000), Shannon (1998), Nordgren (1995), however, they are based on author’s personal experience of simulation model development. No such process has been reported in the literature that entails a simulation modelling process based on an empirical study of expert modellers’ contexts and practices. It would be interesting to consolidate a process from real world practices of expert modellers and compare it with the processes reported in literature. 6. CONCLUSION Studying the simulation contexts and practices of exerts helped understanding the way they develop their models modellers. Most of the participants do not seem to have a very well defined and a formal simulation modelling process. However, most of them seem to have some key steps or stages in their process of simulation model development. Generally a three phased process has been identified from the participants which can be named as problem definition, model development, and model usage. This study identifies some general trends in the simulation model development practice of the participants. It would be hard to generalise the results across the business process modelling and simulation community, however, it gives us some indication as to how people develop their models when their models are small/medium and their model’s complexity is low/medium and when models developed for short-term use. This study does not provide a uniform view of simulation practice in business and industry but some trends and indications on which future studies can be built to further underpin our understanding of the simulation practice real world. Conducting studies in each niche (e.g. defence, manufacturing, healthcare, retail, logistics etc.) of simulation modelling will help further understanding the state-of-the-art and state-ofpractice in discipline specific area. Moreover, in-depth studies of various aspect of simulation modelling process (e.g. problem understanding, model design, documentation) will help understand and improve simulation practice. Furthermore, the findings from this study also encourage us to consolidate a simulation modelling process based on the empirical data collected from expert simulation modellers, which will be reported in future publications.

“In most professions, competent work requires the disciplined use of established practices. It is not a matter of creativity versus discipline, but one of bringing discipline to the work so that creativity can happen.” However, it seems that generally the simulation modellers are more interested in the end product and less in the process of creating that product. In simulation, where the world is driven by time constraints, commercial pressures, and competition, weakness in the modelling process may bring up many issues. Therefore, Gass (1987) suggested: “We need to get away from the crutch that modelling is an art. Guidelines need to be proposed, methodologies for validation and evaluation need to be formalized and applied; and the concept that modelling is a profession with standards must be brought into education and on-the-job training activities of the coming generation of analysts.” Eriksson (2003) suggest that a model’s quality is questionable if it is constructed without a disciplined approach. It can be argued, therefore, that the creative principles of simulation modelling can be incorporated in a disciplined framework for simulation model development. A disciplined simulation modelling process that provides room for creative aspects of simulation is likely to produce good simulation models

REFERENCES Ahmed, R., Hall, T., Wernick, P., Robinson, S. and Shah, M., (2008). Software process simulation modelling: A survey of practice, Journal of Simulation Modelling 2 (2) : 91- 102 Ahmed, R, and Robinson, S. (2007). Simulation in Business and Industry: How simulation context can

296

affect simulation practice? Spring simulation (SpringSim) multi-conference, Business and Industry Symposium, 24-29 March 2007, Norfolk, USA Arthur J D, Sargent R G, Dabney J D, Law A M, Morrison J D (1999). Verification and validation (panel session): what impact should project size and complexity have on attendant V&V activities and supporting infrastructure? Proceedings of the 31st Winter Simulation Conference, pp: 148 - 155 Banks J. (2001). Panel Session: Education For Simulation Practice --- Five Perspectives. Proceedings of the 33rd Winter simulation conference Arlington, Virginia pp. 1571 - 1579 Chwif, L., Barretto, M.R.P.and Paul, R.J. (2000), On simulation model complexity. Proceedings of the 32nd Winter Simulation Conference archive. Dec. 10-13, 449 - 455 Cochran, J.K., Mckulak,,G.T. and Savory, P.A. (1995), Simulation project characteristics in Industrial settings. INTERFACES. 25:104-113 Cugola G, Ghezzi C. (1998). Software Process: A Retrospective and a Path to the Future. Software Process Improvement and Practice. Vol. 4:101-123 David A. (2001). “Model Implementation: A State of the Art.” European Journal of Operational Research. Vol. 134:459-480 Eriksson DM, (2003). A Framework for the Constitution of Modelling Processes: A Proposition. European J. of OR, vol. 145, pp202-215 Ferguson P., Humphrey WS., Khajenoori S. Macke S. and Matvya, A. (1997). Results of Applying the Personal Software Process. Computer. Vol. 30(5):24 -31 Foss, B.A., Lohmann, B. and Marquardt, W. (1998). A Field Study of the Industrial Modeling Process. Journal of Process Control, Vol. 8(5/6):325-338 Gass, S.I. (1987), Managing the Modelling Process: A Personal Reflection. European Journal of Operations Research. 31:1-8 Hollocks, B.W. (2001), Discrete-event simulation: an inquiry into user practice. Simulation Practice and Theory, 8:451-471 Humphrey WS. (1997). Introduction to the Personal Software Process. Addison-Wesley Publications, Harlow, UK. Humphrey WS. and Kellner MI (1989). Software Process Modelling: Principles of Entity Process Models. 11th Int. Conference on Software Engineering 15-18 May. Johnson-Laird PN. (1988 ). Freedom and Constraint in Creativity, in R.J. Sternberg (ed) The Nature of Creativity: Contemporary Psychological Perspectives. Cambridge: Cambridge University Press. 1988 pp 202-219

Kellner IM., Madachy R. and Raffo D. (1999). Software process simulation modelling: Why? What? How? J. of Systems and Software, Vol. 46(2/3):91-105 Kneller G. (1965). The Art and Science of Creativity. Holt, Rinehart and Winston Inc. London Law, D.R. (1998), Scalable means more than more: a unifying definition of simulation scalability. Proceedings of the 30th Winter Simulation Conference, 781-788 Law AM. and Kelton WD. (2000). Simulation Modeling and Analysis. 3rd ed. McGraw-Hill, New York. Nance RE. and Sargent R.G. (2002). “Perspective on the evolution of simulation”, Operations Research, INFORMS, 50(1):161-172 Melão N. and Pidd, M. (2003), Use of business process simulation: A survey of practitioners. Journal of Operational Research Society 54:2-10 Nordgren WB. (1995). Steps for Proper Simulation Project Management. Proceedings of the 27th Winter Simulation Conference Arlington, Virginia, United States. pp: 68-73 Murphy, S.P., and Perera, T.D. (2001), General applications: Simulation practice: key enablers in the development of simulation. Proceedings of the 33rd Winter Simulation Conference Page E H, Nicol D M, Balci O, Fujimoto R M, Fishwick P A, L'Ecuyer P and Smith R (1999). Panel: Strategic directions in simulation research. Proceedings of 31st Winter Simulation Conference, pp: 1509-520 Paul RJ., Eldabi T. and Kuljis J. (2003). Perspectives on Simulation in Education and Training: Simulation Education is no Substitute for Intelligent Thinking. Proceedings of the 35th Winter Simulation Conference: New Orleans, Louisiana. pp. 1989 1993 Pidd M. (1996). Tools for thinking: modelling in management science. Chichester, John Wiley & Sons Ltd., Powell SG. (1995). The Teachers Forum: Six Key Modelling Heuristics. INTERFACES, Vol. 25(4):114-125 Robinson S. (2002). Modes of Simulation Practice: Approaches to Business and Military Simulation. Simulation Modelling Practice and Theory, Vol. 10(8):513-523 Robinson S. (2004). Simulation: The Practice of Model Development and Use. Wiley, Chichester, UK Robinson, S., Nance, R.E., Paul, R.J., Pidd, M. and Taylor, S.J.E. (2004), Simulation Model Reuse: Definitions, Benefits and Obstacles. Simulation Modelling Practice and Theory. 12:479-494 Salt, J.D. (2006), Modes of practice in military simulation. Proceedings of OR 48, 11-13 Sept. 2006. UK

297

Shannon RE. (1975). Systems Simulation: The Art and Science. Prentice-Hall. Shannon RE. (1998). Introduction to the Art and Science of Simulation. Proceedings of the 30th Winter Simulation Conference Washington, D.C., United States pp:7-14 Simonton DK. (2002). Creativity as a Constrained Stochastic Process. in R.J. Sternberg, E.L. Grigorenco, & J.L. Singer (eds) Creativity: from Potential to Realization Washington, D.C. American Psychological Association. pp 83-101 Taylor, S.J.E., Lendermann, P., Paul, R.J., Reichenthal, S.W., Straýburger, S. and Turner S.J. (2004), Panel on Future Challenges in Modeling Methodology. Proceedings of the Winter Simulation Conference, 319-327 Willemain TR. (1994). Insights on Modelling from a Dozen Experts. Operations Research, Vol. 42 (2):213-222 Willemain TR. (1995), Model Formulation: What Experts Think About and When. Operations Research, Vol. 43(6): 91

298

MODELLING RESILIENCE IN CLOUD-SCALE DATA CENTRES John Cartlidge (a) & Ilango Sriram(b)

Department of Computer Science University of Bristol Bristol, UK, BS8 1UB (a)

[email protected], (b) [email protected]

There are two important reasons why this is the case. Firstly, there is no uniform definition of what a cloud computing infrastructure or platform should look like: where Amazon uses virtualization (DeCandia et al. 2007), Google uses MapReduce (Dean and Ghemawat 2008). Secondly, it is a hard problem: a realistic simulation tool should include real network models (fibre channel, Gbit ethernet), disk models (disk arrays, solid-state, caching, distributed protocols and file systems), queueing models for web servers, etc. As such, while it is our long-term goal to develop a set of simulation tools that can be used to aid the development of cloud DCs, as an initial step we present a tractable problem using a simplified model. DCs for cloud computing have now reached such a vast scale that frequent hardware failures (both temporary and permanent) have become a normal expectation. For example, if a DC contains 100,000 servers and the average commodity server life expectancy is 3 years, we expect a server to reach the end of its natural life every 15 minutes; considering temporary failures and failure of other components makes failures occur even more frequently. Thus, when a job is submitted to the cloud, the physical hardware available at the start of the job cannot be guaranteed to be there at the end:

ABSTRACT The trend for cloud computing has initiated a race towards data centres (DC) of an ever-increasing size. The largest DCs now contain many hundreds of thousands of virtual machine (VM) services. Given the finite lifespan of hardware, such large DCs are subject to frequent hardware failure events that can lead to disruption of service. To counter this, multiple redundant copies of task threads may be distributed around a DC to ensure that individual hardware failures do not cause entire jobs to fail. Here, we present results demonstrating the resilience of different job scheduling algorithms in a simulated DC with hardware failure. We use a simple model of jobs distributed across a hardware network to demonstrate the relationship between resilience and additional communication costs of different scheduling methods. Keywords: cloud computing, simulation modelling, data centres, resilience 1. INTRODUCTION Cloud computing—the online utility provision of hardware and software computing infrastructure and applications—necessitates the demand for data centres (DC) on an ever-increasing scale. The largest now fill purpose-built facilities approaching one million square feet.1 Already, DCs are so large that manufacturers (including IBM, HP, Sun) do not have the capability to build and destructively test models on the scale of the final production systems. Hence, every day, massively parallel, tightly-coupled, complex and sometimes heterogeneous data centres are put to service having undergone insufficient pre-testing; while it is still possible to test individual node servers and other standalone hardware, the complex interactions between the components of the DC under normal and abnormal operating conditions are largely unknown. Whereas in other engineering domains this problem has been addressed with robust industry-standard simulation tools—SPICE for integrated circuit design (Nagel 1975), or computational fluid dynamics for the aeronautics industry—a well established realistic (rigorous) simulation framework of cloud computing facilities is lacking. 1

With such high component failure rates, an application running across thousands of machines may need to react to failure conditions on an hourly basis (Barroso and Hölzle 2009)

To avoid frequent job failures, redundancy is necessary. The cloud computing design paradigm builds on achieving scalability by performing scale-out rather than scale-up operations, i.e., increasing resources by using additional components as opposed to using more powerful components. For this reason, jobs are generally split into parallel tasks that can be executed by (potentially) several services. For resilience purposes, the tasks can be multiply copied and run in parallel threads on different hardware (Hollnagel, Woods, and Levson 2006). Thus, as long as a “backup” copy exists, individual task failures will not degrade a job's overall resilience. However, redundancy inherently generates extra work, requiring more space, greater computational

http://www.datacenterknowledge.com/special-report-the-worlds-largest-data-centers

299

where test environments are no longer larger, or even of the same order of magnitude, as the final products. Cutting edge DCs are believed to have more than half a million cores,2 but even one order of magnitude less would make a physical test environment too expensive to be practical. Hence, it is difficult to impossible to test the chosen configurations before going into production, which can lead to costly errors. This highlights the need for predictive computer simulations to evaluate possible designs before they go into production: with simulation studies it is possible to rapidly evaluate design alternatives. However, for simulating cloud-scale computing DCs there are currently no well-established tools. The literature includes some early-stage cloud simulation models. For a consumer centric view of the cloud, there is CloudSim (Buyya, Ranjan, and Calheiros 2009). CloudSim's design goal is to compare the performance of services on a cloud with limited resources against their performance on dedicated hardware. To aid the vendor perspective, we have previously developed SPECI (Simulation Program for Elastic Cloud Infrastructures) for modelling scaling properties of middleware policy distribution in virtualised cloud data centres (Sriram and Cliff 2011). This paper explores aspects of resilience modelling that we aim to develop as a component in a set of simulation tools for data centre designers.

Figure 1: Example hardware: (a) HP C7000 chassis holding 16 blade servers; (b) Sun Pegasus C48 server rack, containing 4 chassis × 12 blade servers. effort and increased communication costs. There is clearly a trade off here: how much redundancy and how to schedule redundancy—where to physically locate copies of the same code in the DC to minimise the chances of failure—versus increased communication cost and computational effort. In this paper, we conduct an initial foray into the analysis of this trade off, using a simple simulation model to analyse the relationships between scheduling, redundancy, network structure and resilience. In Section 2 we introduce cloud-scale data centres and the problem of failure resilience. Section 3 outlines the simulation model we use, before detailing our experimental set-up in Section 4. Section 5 presents the results of our experiments, which are discussed in Section 6. In Section 7 we outline our future plans to extend this work, before summarising our conclusions in Section 8. 2.

2.2. Failure, Resilience and Communication Cost As economies of scale drive the growth of DCs, there are such a large number of individual independent hardware components that the average life expectancy will imply that component failure will occur continually and not just in exceptional or unusual cases. This expected near permanent failing of components is called normal failure. For practicable maintenance, failed components are left in situ and only replaced from time to time; it may also be imagined that entire racks are replaced once several servers on it have failed. However, despite normal failure, resiliency must be maintained. Furthermore, the cloud design paradigm of solving jobs using smaller tasks or services that are typically spread across several physical components further increases the risk of normal failure affecting any given job. As cloud vendors seek to provide reliable services, requiring the maintenance of guaranteed levels of performance and dependability, resilience has become a new non-functional requirement (Liu, Deters, and Zhang 2010). To this end, cloud applications such as BigTable, Google's massively parallel data storage application, have in-built management systems for dealing with failures (Chang et al. 2008). Hardware failure can occur anywhere in the physical hierarchy of the data centre: power outages can disable an entire DC; faulty cooling system behaviour can force an aisle to be shutdown to avoid overheating; racks, chassis and blades have individual power

BACKGROUND

2.1. Cloud Data Centres Cloud Computing transitions DCs from co-located computing facilities to large resources where components are highly connected and used in an interlinked way. Computations are broken down into services, allowing for easier scale-out operations. From the physical perspective, DCs are structured regularly in a hierarchical design: a warehouse scale DC is made up of aisles of racks, each rack being a vertical frame to which a number of chassis can be mounted; each chassis containing an arrangement of thin computer mother-board units: the blade-servers that make up the DC's computing infrastructure. Each blade server in turn hosts Virtual Machines (VMs) running cloud services. Figure 1 shows example chassis and rack components. With Cloud Computing, the level of interconnectivity and dependency between services across the DC is so high that Barroso and Hölzle (2009) coined the term “warehouse-scale computers”. This introduces various aspects of complexity to DCs. Firstly, many of the protocols in place scale worse than linearly, making conventional management techniques impractical beyond a certain scale as complex interactions between services lead to unpredictable behaviour. Secondly, DC design has reached a stage

2

300

http://www.zdnet.com/blog/storage/googles-650000-core-warehouse-size-computer/213

HP C7000 chassis and 4-chassis IBM rack shown in Figure 1. 3.2. Jobs, Tasks and Redundancy We assume that all jobs to be run in the cloud can be parallelized into T independent task threads. We make this simplifying assumption on the basis that one of the major draws of cloud infrastructures is the elasticity of rapid scaling and de-scaling through parallelization. In our model, J jobs are run on the DC, with each job, J, consisting of T independent parallel tasks. While tasks can be parallelised, they are not entirely independent otherwise they would constitute a new job. Thus, tasks must periodically communicate with each other, passing runtime data when necessary. To pass runtime data, tasks within a job communicate at fixed time intervals. Normally, if any one task within a job fails, the entire job will fail. To mitigate this, redundancy can be introduced by running R duplicate copies of tasks in parallel. Then, job J will fail if and only if all redundant copies of an independent parallel task fail. Such redundancy introduces failure resilience. Let J denote a job consisting of T tasks, each having R redundant copies. Then, J can be written in matrix notation, with T rows and R columns:

Figure 2: Data centre tree schematic. We describe this as an h-2-2-2-3 hierarchy (2 racks per aisle, 2 chassis per rack, 2 blades per chassis and 3 services per blade). The full DC contains as many aisles as necessary. supplies which can fail; and individual VMs can suffer from instability in software and require an unplanned reboot. Thus, with growing DC scales, resources can no longer be treated as stable; and interactions no longer static but rather exhibiting dynamic interactions on multiple descriptive levels. To counter normal failure, redundancy must be introduced. This happens by spinning off parallel copies of all tasks. Thus, when any task from the original copy fails, a redundant copy is available to replace the service that has gone “missing”. Hadoop, for example, is an open-source software for reliable, scalable distributed computing and is used by Yahoo!, Facebook and others, on clusters with several thousand nodes.3 It includes HDFS file system, which as default creates 3 copies (redundancy 3).4 When considering parallel execution of tasks rather than file storage, service redundancy causes extra load through the additional execution of tasks. The execution load grows linearly with the numbers of redundant copies, but in addition, there will be some form of load associated with parallel threads periodically passing runtime data that we describe as communication cost. This paper uses a simulation model of parallel job execution to explore the trade-off between resilience and communication cost as failure, redundancy and scheduling types vary. For model simplicity we focus on computational redundancy and ignore disk and I/O redundancy. 3.

⎛ j1,1 ⎜ ⎜ j 2,1 ⎜  J =⎜ ⎜ j t,1 ⎜  ⎜ ⎝ jT ,1

4



j 2,2   

j1,r



j 2,r   

j t,2



j t,r





 



 

jT ,2

jT ,r

j1,R ⎞ ⎟ j 2,R ⎟  ⎟ ⎟ j t,R ⎟  ⎟ ⎟ jT ,R ⎠

(1)

Job failure occurs when all tasks in a given row fail. More formally:

{

}

fails( J ) ⇔ ∃t ∈ T, ∀r ∈ R, fails( j t,r )

(2)

Throughout this paper, we denote experiments running J jobs, each with T tasks and R redundancy as a {J, T, R} configuration, with sum total tasks #T = J x T x R. 3.3. Scheduling Algorithms Jobs and tasks can be placed onto a DC network in an infinite variety of ways; using schedules that range from the simple to the complex. In this work, we are interested in deriving general relationships between job scheduling methods and the effects they have on communication cost and resilience. Since we cannot hope to assess the relative behaviours of every scheduling algorithm, to aid analytical tractability, we selected a small subset purposely designed to be simple. The intention is not to test intelligent, complicated, realworld algorithms, but rather to tease out general behaviours of these simple algorithms so that we can equip ourselves with better knowledge to design intelligent industrial algorithms in the future. To this end, we define the following three scheduling algorithms:

SIMULATION MODEL

3.1. Network Tree Hierarchy We model the interactions between networks of VM cloud services that exist in a hierarchical tree-structure (refer to Figure 2). Network structure is configurable and we use several tree hierarchies. Throughout this paper, however, unless otherwise stated assume a fixed hierarchy h-8-4-16-16. That is, each aisle has 8 racks, each with 4 chassis containing 16 blades, with each blade running 16 cloud services. This structure was chosen to model realistic hardware, such as the 16-blade

3

j1,2

http://wiki.apache.org/hadoop/PoweredBy http://www.hadoop-blog.com/2010/11/how-to-change-replication-factor-of.html

301

• •



Random: Uniformly distribute tasks across the DC, independent of job or redundancy group. Pack: Use the minimum amount of DC hardware to run all jobs. Place tasks from consecutive redundancy groups for all jobs on consecutive DC services. Cluster: Place all tasks belonging to the same redundancy group on the smallest piece of hardware that they fit (e.g., on one blade).5 Uniformly distribute redundancy groups across the DC.

Table 1: Communication Costs Communication Relative Cost Inter-Service CS = 100 Inter-Blade CB = 101 Inter-Chassis CC = 102 Inter-Rack CR = 103 Inter-Aisle CA = 104

Figure 3: Job scheduling for a {J=2, T=3, R=2} configuration on an h-2-3-8 hierarchy subset. Top: Random uniformly distributes the #T=12 tasks across the DC. Middle: Pack schedules tasks onto the minimum physical hardware set, in this case 2 blades on 1 chassis. Bottom: Cluster schedules full job copies onto the same physical hardware, while uniformly distributing copies across the DC.

Figure 4: Communication network costs. Top: tasks communicate with the nearest copy of every other task. Bottom: when a task fails, communicating tasks find the nearest alternative. When task 2 fails, communication costs increase from 6CS to 4Cs+2CB. Refer to Table 1 for cost values.

Figure 3 shows a schematic example of each scheduling algorithm. Random, top line of figure, assigns tasks to the DC using a random uniform distribution over all DC services. Random schedules tasks independently, taking no account of which job or redundancy group a task belongs. Conversely, Pack preserves geographical co-location of individual tasks according to job and redundancy groupings, middle. Tasks are sequentially scheduled using neighbouring services until each hardware is filled. Finally, Cluster uses a combined approach, bottom. Similar to Pack, Cluster places all tasks belonging to a job redundancy group on the same physical hardware. However, redundancy groups themselves are uniformly distributed across the DC. In aggregate, these trivial scheduling algorithms form a simple strategy spanning-set from which we aim to tease out general rules for improving failure resilience.

Figure 5: Hardware failure. Top: initial communication networks resulting from alternative scheduling methods. Middle: individual service failure produces minor restructuring of communication networks, including the addition of more costly inter-blade edges. Bottom: blade failure produces major network reconfiguration, while Random and Cluster recover Pack results in job failure. the DC. Intuitively, cost increases with physical distance between tasks, increasing in magnitude each time it is necessary to traverse a higher layer in the DC tree (Figure 2). From Table 1, tasks communicate with themselves and other tasks on the same service with zero cost. For tasks on different services on the same blade the communication cost is CS=100; between different blades CB=101; etc. These costs were chosen to give a qualitatively intuitive model of costs: clearly, communication costs between tasks running on the same physical chip are

3.4. Network Communication Costs As explained in Section 3.2, the model assumes that tasks within a job need to communicate at fixed time intervals, passing runtime data between parallel threads. Table 1 shows inter-task communication costs within 5

In the case that no hardware has enough free space to fit the entire task-group, deploy as many tasks as possible on the hardware with the largest free space, then deploy the remaining tasks as “close” (lowest communication cost) as possible.

302

many orders of magnitude lower than between tasks located in different aisles of a data centre. While more accurate estimates of relative costs are possible, we believe the simple relationship defined in Table 1 adequately serves the purposes of this paper, since we are only interested in qualitative differences between scheduling algorithms, rather than accurate quantitative relationships.

rate. Communication costs increase with physical distance; see Table 1. Network Utilisation: The DC is effectively infinite in size, enabling us to ignore the dynamics of full utilisation. Failure: Failure can occur at any level in the hierarchical tree. Failure events are drawn from an exponential distribution.

3.5. Hardware Failure Hardware failures directly affect task communication. Figure 4 highlights a schematic example of the effects of a single service failure. Initially, tasks communicate with the “nearest” copy of every other task; where “nearest” is defined as the least costly to communicate. Top: communication takes place between tasks running on the same blade. Middle: after task 2 fails on blade 1, tasks 1 and 3 on blade 1 begin inter-blade communication with the “nearest” alternative copy of task 2. The resulting network communication load increases from 6CS to 4Cs+2CB. Within the model, hardware failure can occur at any level in the physical hierarchy tree. Figure 5 demonstrates example effects of failure on the underlying communication network. Tasks form an initial communication network, top. Individual service failure, middle row, results in some restructuring of the communication network; with the addition of more costly inter-blade links for Random and Cluster. Hardware failure of an entire blade server, bottom row, has a more profound effect on the network. While Random and Cluster find a new rewiring, there no longer exists a full task-set for Pack, thus resulting in job failure.

4.2. Configuration Hierarchy Tree: Unless otherwise stated, all experimental runs use an h-8-4-16-16 tree hierarchy. These are realistic values based on current consumer hardware (refer to Section 3.1). Where alternative tree architectures are used, we use the notation h-5 and h-10 as shorthand for h-5-5-5-5 and h-10-10-10-10, respectively. DC size: To approximate unlimited resources, we scale the size of the data centre, |DC|, to equal twice the size needed to run all jobs, that is:

DC = 2×# T = 2 × J × T × R

(3)

In an alternative configuration, data centre size is fixed. Under these conditions, set:

DC = 20 × J × T

(4)

Communication Costs: Communication costs are set equal to Table 1 Refer to Section 3.4 for a discussion. Scheduling: Jobs are scheduled using the algorithms Random, Pack and Cluster (as detailed in Section 3.3). Hardware Failure: We set the proportion of hardware that will fail, fhw, during the length of a simulation run to 1%, 5% or 10%. Note, however, that a failure event will cascade down the hierarchy tree, such that failure of a chassis will cause all blades and services running on the chassis to fail. Thus, the overall proportion of a DC that will fail during a simulation run will be larger than the value of fhw. These failure rates may appear to be high. However, it is our intention to model resilience under extreme conditions that cannot be observed readily in operational environments. When a hardware failure event occurs, a discrete distribution is used to select the type of hardware failure. The relative probability of a given type of hardware, htype, failing is calculated as the relative proportion of that hardware in the data centre, htype hall . Although this

4. EXPERIMENTAL DESIGN Using the model described in Section 3, we perform a series of empirical experiments to observe the effect that different job scheduling algorithms have on resilience and communication cost in a data centre with hardware failures. 4.1. Assumptions To keep the model tractable we make some simplifying assumptions. Time: Simulations have a fixed time length. Jobs are scheduled before the simulation clock begins, then run for the entire length of the simulation. Jobs: Jobs consist of a set of tasks that can be run in parallel with no inter-dependencies or I/O requests, other than periodic passing of runtime data. Consider, for example, computationally intensive batch jobs such as overnight computation of market data for financial institutions, or CFD simulations for the aeronautics industry or Met Office. For all tasks comprising a job, if at least one copy of the task succeeds, then the job completes successfully; refer to equation (2). Communication Cost: Tasks within a job need to communicate with a copy of all other tasks at a constant

distribution is simplistic, it provides the intuitive result that the more common a type of hardware, the more likely it is to fail. 5. RESULTS Here, we present simulation results for all scheduling experiments. Figures plot mean values of repeated trials, plus or minus 95% confidence interval. Thus, where error bars do not overlap, differences are

303

Figure 6: Resilience of scheduling algorithms in a fixed-size data centre with tree hierarchy h-8-4-16-16. As hardware failure increases, top to bottom, resilience falls; as tasks per job increases, left to right, resilience falls. Overall, Cluster is more resilient than Random and Pack across all conditions. Error bars show 95% confidence intervals. The dotted horizontal line plots the mean percentage of DC services surviving at the end of a run: the resilience that jobs with T=1 and R=1 will tend toward. and all job sets they occur often enough to stop Pack from reaching SJ=100%, regardless of R. For all algorithms, we see that as the number of tasks per job, T, is increased from T=10, left, to T=100, right, more redundancy is needed to maintain a given level of resilience. This is intuitive. Since task failure results in job failure, the greater the number of tasks per job, the greater the chances of any one job failing; hence, the greater the number of redundant copies needed to counter this failure. Similarly, when the probability of hardware failure, fhw, is increased from 0.05, top, to 0.10, bottom, to maintain resilience redundancy R must be increased. Once again, this is intuitive: as failure increases, so too does the likelihood of job non-completion. Overall, across all conditions, Cluster is the most resilient. With low values of R, Cluster and Pack outperform Random. When R ≥ 7, Cluster and Random outperform Pack. Further, there is no condition under which Cluster is significantly outperformed by either Random or Pack. Yet, there are several conditions under which Cluster significantly outperforms both alternatives. Thus, results suggest that Cluster is the most robust strategy. Interestingly, the default number of redundancies used in Hadoop's HDFS, R=3, appears to be a reasonable choice when T=10. As the number of tasks increases, R=3 does not suffice under our conditions.

statistically significant. The simulation experiments were run in parallel and distributed across a cluster of 70 linux machines and the number of repetitions varies between 30 repetitions to over 100 repetitions. Confidence intervals remain relatively large due to the stochastic nature of the failure process: particular failure events can have widely ranging effects. It should be noted that occasionally the entire DC fails during simulation. When this occurs, the run is rejected so these catastrophic outliers do not skew results. This is reasonable since DC failure is a direct result of random hardware failure and is independent of the scheduling algorithms under test. Hence, all plots display summary data from trials where the entire DC did not fail. 5.1. Resilience Figure 6 shows simulation results for J=10 jobs using fixed DC size, equation (4). The proportion of successful job completions, SJ, is plotted against number of redundant task copies, R, for each algorithm: Random, Cluster, and Pack. In each graph, we see the intuitive result that success, SJ, increases with redundancy, R. However, whereas Random (blue circle) and Cluster (red square) reach 100% success under all conditions except bottom-right, Pack (green triangle) reaches a maximum in the range 80%-90% at approximately R=5 and then plateaus, with fluctuation, as R is increased further. As shown schematically in Figure 5, Pack schedules all tasks to fit on the smallest hardware set possible. However, this tactic of “putting all your eggs in one basket” is vulnerable to specific hardware failure events that may take out the entire set of tasks. Although such events are rare, across all runs

5.2. DC Architecture Figure 6 displays a clear relationship between increased redundancy, R, and increased resilience, SJ. However, the resilience graphs for Pack exhibit “dips”, for example at R=8 (top-left and bottom-right) and R=7

304

Figure 7: Pack scheduling using h-8-4-16-16, h-5 and h-10, from left to right, respectively. Resilience dips when jobs exactly fit on underlying hardware.

Figure 8: Cluster scheduling using h-8-4-16-16, h-5 and h-10, from left to right, respectively. Cluster is largely insensitive to underlying hardware architecture. comparison, when R=4 or R=6, task group copies will be more unevenly distributed over hardware, giving greater resilience to failure. Figure 8 plots results for Cluster under the same conditions as Figure 7. Here, we see that Cluster is largely unaffected by the underlying tree structure of the data centre.

(bottom-left), which raises some questions. Are these “dips” merely statistical aberrations, or are there underlying dynamics causing specific R values to result in lower SJ? To address this issue, a series of experiments were performed using alternative DC tree structures (Figure 2) to observe the effect that hardware hierarchy has on resilience. Three configurations were tested: h-8-4-1616, h-5 and h-10. In each case, data centre size |DC| was variable; refer to equation (3). For all hierarchy trees, results were qualitatively similar to those displayed in Figure 6, suggesting that the general behaviours of each scheduling algorithm are largely insensitive to the underlying hardware hierarchy configuration. However, Pack does display some idiosyncratic sensitivity to hierarchy. Figure 7 plots SJ against R for Pack under each tree hierarchy. Left: SJ increases with R until R=5, but then fluctuates around 80%, with a minimum at R=7. When DC hierarchy is changed to h-5, centre, Pack has a minimum at R=5. Finally, with hierarchy h-10, right, there is a minimum at R=10. This evidence suggests that Pack is sensitive to the underlying physical hierarchy of the DC. In particular, if all redundant copies of a job fit exactly onto one hardware unit—blade, chassis, rack, etc.—then a failure on that hardware will take out all copies of the entire job. Hence, with an h-5 hierarchy, for instance, R=5 results in poor resilience for Pack. Under these circumstances, each job, JT,R, contains 50 tasks, which exactly fit onto 2 chassis. Thus, two neighbouring chassis failure events will take out the entire job. In

5.3. Jobs To see how results scale with an increase in jobs, we ran experiments with J=100 jobs, using hierarchy h-8-4-1616 with variable data centre size, |DC|. Results are qualitatively similar to those of Figure 6. Pack outperforms Random with low R, but plateaus around SJ=80%. Random has poor resilience when R is low, but outperforms Pack as R approaches 10. Finally, Cluster has best resilience overall. Results for other failure rates and number of tasks, T, are also qualitatively similar, indicating that resilience is insensitive to J. 5.4. Communication Costs While it is important for scheduling algorithms to enable job completion, it is also important they do not induce prohibitive cost through wasteful communication. In this section, we explore the mean communication cost, CJ, for each successfully completing job. Results suggest that the relationships between algorithms are largely robust to variations in hierarchy-tree, number of jobs, J, number of tasks, T, and hardware failure rate, fhw. Thus, in this section we consider only two conditions: variable |DC|; and fixed |DC|. Each time an h-8-4-16-16 architecture is used.

305

Figure 10: Communication costs per successful job, normalized to R=10. Clear-faced markers represent variable-sized DC; filled markers show fixed-size DC. We see that communication costs scale linearly with redundancy for all algorithms.

Figure 9: Costs per successful job, CJ, using fixed |DC| with h=8-4-16-16.

Figure 9 displays communication costs per successful job, CJ, using fixed |DC|. 95% confidence intervals are drawn, but are generally very small. With tasks uniformly distributed throughout the DC, Random produces the greatest communication costs per job. Conversely, with tasks placed as close together as possible, Pack has the smallest communication costs per job. For Pack and Random, an increase in redundancy, R, leads to a proportional increase in cost, CJ. When R=10, CJ is approximately 10 times greater than the value at R=1. For Cluster, however, the story is different. When R=1, Cluster produces smaller CJ than Pack, since Cluster guarantees all job copies are placed on the lowest branch of the hardware tree that they fit; with Pack, however, depending upon number of tasks, T, and the underlying tree-hierarchy, some jobs will occasionally be split across hardware (see schematic Figure 3, for example), thus incurring greater communication costs. When redundancy is increased to R=2, communication costs, CJ, becomes a magnitude greater than Pack. As Cluster distributes job groups across the network, when an individual task fails, new communication links to alternative copies are likely to be long-range and costly. In contrast, Pack places all clones near each other, so communication with alternatives does not radically increase costs in most cases (refer to Figure 5). Interestingly, with fixed |DC| mean communication cost per successful job, CJ, remains constant when R ≥ 2. Since Cluster distributes job copies uniformly across the data centre, the mean distance or cost for communication between tasks in different redundancy clusters is inversely proportional to R. Hence, additional redundancy reduces the mean communication cost a task must pay to communicate with an alternative clone, thus making CJ invariant under changes to R. It should be noted that the same is not true for Random, however. Unlike Cluster, since Random distributes all tasks uniformly independent of redundancy group, the majority of communication paths are inter-hardware and costly. Hence, doubling R will approximately double CJ. When using a variable-sized data centre—equation (4)—results for CJ against R are similar to Figure 9 for Random and Pack. For Cluster, however, CJ is no

longer invariant to R and instead increases proportionally as R increases. As |DC| increases with each increase in R, the mean length between communicating tasks remains stable. Thus, as the number of tasks increases so too does overall communication costs. Figure 10 plots normalized communication cost for each scheduling algorithm against redundancy, R. Clear faces show fixed |DC| (not including Cluster) re-plotted from Figure 9. Coloured faces show data from the equivalent set of runs using variable |DC|. In all cases, with all algorithms, there is clearly a linear relationship, suggesting that communication costs rise in direct proportion to R. Note, however that this is not the case for Cluster under fixed |DC| (not plotted): here, communication costs are invariant in R. 5.5. Summary of Findings The main findings can be summarized as follows: 1.

2.

3.

4.

The network hierarchy tree has little effect on the resilience of scheduling algorithms (except in the case of Pack, where particular tree configurations have negative impact on particular levels of redundancy). Cluster is the most resilient scheduling algorithm from the selection modelled. In contrast, Pack is a non-resilient high-risk algorithm. Pack is the most efficient algorithm, Random the least. Cluster generates intermediate costs, but scales well under fixed data centre size. Overall, Cluster is the most practical algorithm, effectively combining the efficiencies of Pack with the resilience of Random.

6. DISCUSSION The aim of this work is to build an understanding of the general relationships between scheduling, resilience and costs (rather than perform a detailed analysis of any particular algorithm), the results presented support our basic endeavour to use simulation models as a methodological framework to design and test tools for elastic cloud-computing infrastructures. We do not

306

of failure. Conversely, a Random distribution of tasks throughout the DC leads to greater resilience, but with a much elevated cost. Clustering tasks together in cohesive job-groups that are then distributed throughout the DC, however, results in a beneficial trade-off that assures resilience without prohibitive costs. This work provides a teasing glimpse into the powerful insights a cloud simulator can provide. Given the grand scale of the challenge, this work has naturally raised many open questions and introduced scope for future extensions.

suggest that Random, Pack or Cluster are practical jobschedulers that should (or would) be used in industry, but rather that these purposely naïve algorithms provide a simple base-line set of strategies that enable us to tease out fundamental relationships between density, clustering and spread of jobs; and the impact each has on resilience and communication cost. By using simulation to better understand how these concepts interact, we gain access to a powerful off-line test-bed for algorithm development that provides a design route towards more robust and efficient cloud-computing services. The simulation model we have used makes some simplifying assumptions that should ideally be relaxed. However, despite this, the model is powerful enough to highlight how underlying complex interactions, such as between scheduling and the shape of the hierarchy-tree, can affect resilience. This is a promising indication of the value of pursuing the goal of creating an extensible simulation framework for cloud computing.

ACKNOWLEDGMENTS Financial support for this work came from the EPSRC grant:6 EP/H042644/17 (for J. Cartlidge) and from Hewlett-Packard's Automated Infrastructure Lab, HP Labs Bristol (for I. Sriram). The authors would like to thank Prof. Dave Cliff and the sponsors for their support and interest in this topic. REFERENCES Barroso, L. A. and Hölzle, U., 2009. The datacenter as a computer: An introduction to the design of warehousescale machines, Synthesis Lectures on Computer Architecture, no. 1, pp. 1–108, 2009. Buyya, R., Ranjan, R., and Calheiros, R. N., 2009. Modeling and simulation of scalable cloud computing environments and the cloudsim toolkit: Challenges and opportunities, in International Conference on High Performance Computing & Simulation, HPCS ’09, pp. 1–11, June. Chang, F., Dean, J., Ghemawat, S., Hsieh, W. C., Wallach, D. A., Burrows, M., Chandra, T., Fikes, A., and Gruber, R. E., 2008. Bigtable: A distributed storage system for structured data, ACM Transactions on Computer Systems, vol. 26, no. 2, article 4, pp. 1–26, Jun. Dean, J. and Ghemawat, S., 2008. MapReduce: Simplified data processing on large clusters, Communications ACM, vol. 51, no. 1, pp. 107—113. DeCandia, G., Hastorun, D., Jampani, M., Kakulapati, G., Lakshman, A., Pilchin, A., Sivasubramanian, S., Vosshall, P., and Vogels, W., 2007. Dynamo: Amazon's highly available key-value store, SIGOPS Oper. Syst. Rev., vol. 41, no. 6, pp. 205-220. Hollnagel, E., Woods, D. D., and Levson, N., 2006. Resilience Engineering: Concepts and Precepts. Ashgate. Liu, D., Deters, R., and Zhang, W. J., 2010. Architectural design for resilience, Enterprise Information Systems, 1751-7583, vol. 4, no. 2, pp. 137–152, May. Nagel, L., 1975. Spice2: A computer program to simulate semiconductor circuits, University of California, Berkeley, Tech. Rep. UCB/ERL-M520. Sriram, I., and Cliff, D., 2011. SPECI2 - simulation program for elastic cloud infrastructures, to appear in SummerSim’11.

7. FUTURE WORK Here, we outline potential future extensions: 1.

2.

3.

4.

5.

6.

More realistic modelling assumptions: the introduction of sequential task interdependencies, heterogeneous jobs and services, full-DC utilisation, etc. Model verification and validation using realworld data. Retroactive validation of results through testing on real-world DCs. Introduction of other scheduling algorithms from industry and the development of novel algorithms using evolutionary computation as an automated design tool. Monitor individual failure events rather than failure over time, to observe how the system changes when failure occurs, and what exactly takes place at this lower level of description. Compare the effects of scale-up versus scaleout: If the resource usage increases, what does it mean for the resilience if more services are used, rather than more powerful ones? Introduce migration of services to the scheduling algorithm. This allows a task to be cloned when a parallel instantiation fails, and the clone can then be migrated towards the other tasks belonging to that job.

8. CONCLUSIONS We have presented a simulation model for testing the effects that different scheduling algorithms have on the resilience and communication costs of jobs running “in the cloud” of a large scale data centre (DC). Modelling the data centre as a tree-hierarchy of physical machines and jobs as a collection of parallel tasks that can be cloned, we have demonstrated the effects that different job-scheduling algorithms have on network resilience and communication cost. As intuition would expect, Packing all tasks together in a small area of the DC greatly reduces communication cost but increases risk

AUTHORS BIOGRAPHY Dr John Cartlidge is a Research Associate in cloud computing. His research interests include simulation modelling, automated trading and electronic markets, and evolutionary computing. Ilango Sriram is a final year PhD student soon to defend his thesis on simulation modelling for cloud-scale data centres. 6

307

http://gow.epsrc.ac.uk/ViewGrant.aspx?GrantRef=EP/H042644/1

                $" #   !  &% '%# ( ()  %*#  %   + )    ,   - )..# . ,      /001'. .' )1  .   *'1 . .' ' 1 ). .'' 1 . .'

                                                   

                                                                   

                  !               "        "     

)          ) .  $()+&$*')*',-,*+       ,  -  )           )       22        .7   '  )   8            .

 %'&$(#&% )       2   )  

     )  22         )   )  ) ) )*.)   )      )  )       )       )  . 3         *     )     2      * )     )       )           .,  )                 * )   *' *     )  

   ) )      5046. 7 

  )  8    )    2               9   *' )  )    '        )           * ) )   *.  ) 

        )        . )        *                 )        )     )       .7             *    )                   )  + 2 ' -      : 9 *   3 ) 5;2006.  )   )     *

    +  '       ) %  )*     2       )          C%.   ;//2

   9   *   = >+  '  *       )    )       *)  . 0      

   )  )    )*)  *?

309

      4   .=0> 

B



 





. 7.         

&)  )     '  )  ) )  =  >     *?



. 5.       2'.

   B/  B  $  B /.A  4 B# 4    4/.A =A>

     B   BA  B   BA  *)    I B/    $ I /.A .  )    *)  ) 

) )     )    )  *)  =  > )      ) )        =    >.))  )     ) *)       4F/)        )  *?    )     4/.A.      

    =     >      =4> %     &   B  4 : .F)*)  ) )     *)      )     )  *)      )         )       )  *  )    .       )   2* )  )          7   )   ) *  '=  > )  2* )) !. )   * )    *)   ) !   *)  )            )   =  >     ) ! * ) * )   . )    *   )     ) )  *)     

  .,   =4>)     )       *?            B 4   

=;>

 /./A  B/  0/  

4  /./H  0  0/ ;  [email protected]   B 4    *)    IH/

   I/./A  IB/   I0. ;

 ! "     )     )   )  '  )    '       *)   )   '      B/2   .  . 8#    2'9    :

 '  ()

 )      ) !   )         '    * )    

 



310







. ;  2'

:!   ) *         * ) *   )* : .0/.  =>  ) *  ' . /0   





 )       ) *      ) 

         *  )       )  . /     * ?      )  *    )       *           . /3,               

 

2  ?;//0/

  =>     )  *   '?    )  *    )*  : . 0B *       *    0G)      *       )       )  2  ? 4/



  * ) * )    )         )       )  #     )      .

=>  * 



. /4$         

311



7 : . 0B  )* )  )  N   ) 

  * ) * )              )     . :  )       )      )       )         )  )    ) ** )     . )      *  )         )   4/

.)   : .0G        )       )  *   '  )  . /7           *          . : . 0A )* )       )     *.) *  ' =       ,   )  *             )      77)*)    )    .  * )    ) *  )     . )   04F   *      )          )      . )  BFA   *   )  

   .   77.   )          :   O  ?  DC%32@/0E  C   ?;/2G///

    ,  *  :   ?B/')*   



 . /5               .

 : .0G)*)  ) )   )  3  .

: .0F)*)    ?=>   S => )    .  

. /6       $+ .



 > *@?*'+*%'*,(),  ? ,     )       )         2            2   * *  

  . )           *   )    )* : .0@.

312

=>      => )     . /8    

: . 4/=> )* )  !        )  #   2' * ! *                  = > )         ! = > )      )       = > )   '    )       B/2   9.  *  )     )  *   )    . & ) ;/     )  *  )     3 *  BA       @F   )      . )   ) )         B4U *  )       . > #&%#)(,&%, =>,  

 =>- '  

7 )                =7>   J 2      = 3>L *   )   !



= >7 ! . /;2'



)    ) ! )*  777. -#

.!+7:77CE +

!

 

    O  2   2!  ?AF'  23    ?[email protected]'  2#   ?/./0 D 2#  ' ?04/

 2   ?    !  2   ?    * )?0.A 3    2+,?3  2T@   2

  ?BFA  *   : 2   ?4/  

:  

#        

#B%&C)*$1*+*% )  *' *    %  7       Q    )     + .  '*.*'*%#*,

  

: . 4/ )* )  9         )    3 * )! .       =>7  =>3 !)   

    

= >#     9= >    9

    

= >     9=>:  ) . 0 x1 (t )

x2 (t ) @ { ª¬id (t ) iq (t ) º¼ T

T

u { Rl

is the power factor of wind turbine, which is defined as: Cp O, E

(5)

The electromagnetic torque of permanent magnet generator in d,q coordinate system is

2.2. Wind Turbine Model According to the Bates theory, the mechanical power captured by turbine is: Pwt

*G

2.4. Model of PMSG To suppose the permanent magnet synchronous motor is ideal, when we analyze its basic electromagnetic relations. So it fulfills:

frequency part 'v t is made up of Gaussian white

1 1 'v t  e t Tw Tw

K

electromagnetic torque of the generator. As mentioned above, the formula (4) and (5) constitute the basic lowfrequency model of the wind power conversion system.

describes long-time scale and low-frequency changes, it is usual to assume to be a Weibull distribution; and 'v t is high-frequency component, and the high-



i

Where J l is the total inertia of drive train, i is the gear box ratio, for the direct-driven case, its value is 1, K is the efficiency of the transmission shaft, *G is the

Where v t is low-frequency component, which

'v t

(4)

C* O Cp O O .

2.1. Wind Speed Model Wind speed v t which is a non-statistical random

v t  'v t

0.5SU R 3 v 2 C* O

Where C* O is torque coefficient which is defined as

In Figure 1, the wind turbine and generator are connected directly, Drive Train represents the shaft, and there is no gearbox between them. For this reason, they have the same speed :l .

v t

Pwt :l

* wt

*G

The state model of the generator can be expressed:

12.5 116 0.22(  0.4E  5)e O

O

341

1 ­ ª º ( Rx1  p ( Lq  Ls ) x2 :l ) °• « » L L  d s °x « » ° « » 1 ( Rx2  p( Ld  Ls ) x1: h  p) m : h ) » ° « L L  s ° ¬ q ¼ ° 1 º °ª 0 ®« » ° « Ld  Ls » °« 1 » 0  °« » Lq  Ls ¼ °¬ ° ° °¯ y { *G p) m x2

s

of grid and converter, ) is the flux, and is a constant m for the permanent material, p is the pole pairs of the generator.

The LPV theory was firstly proposed by Professor Shamma, its dynamic characteristics depend on the adjustable parameters which are measured in real time(Shamma,J. and Athans,H. 1991). Since these parameters can reflect the nonlinearity of the system, LPV system can be applied to describe non-linear system. Design gain scheduling controller using linearization method to make controller gain change with the parameters. Refer to the references (Inlian Munteanu, Nicolaos Antonio Cutululis and Antoneta Iuliana Bratcu 2005; 2008), the LPV model of the system can be expressed: ˜

Ax(t )  Bu (t )  Le(t )

y (t )

Cx(t )

C ( U (t )) x t

ª- V  V T * * « M - X U t  Y * « « Y 0 0 « « LT U t 0 0 « « 0 C U t V 0 « 0 T 0 «¬ V

LPV CONTROLLER DESIGN

x t

y (t )

K U t

0 0

-J f I 0

º » » » »0 » » * »  X U t »¼ * * * *

(10)

R U t V 1

according to the formula (5) , (8)

˜

J T ':l

'* wt 

1 *G

K * wt

'*G

˜

Where JT

Jl :l *wt , ':l

(11) ˜

' :l :l

˜

, ': l

˜

˜

:l  :l , ˜

˜

:l is the stable value of the :l , also for the : , ': l l

and '* wt . According to the high-frequency pulsation wind speed modeling method in the reference [9] and formula (2), there is

disturbance input e t to the control output y s less than a given performance index, that is  Jf

.

%v 

then the designed state feedback controller is u t

-J f I

* * * *

the inequality (10) has feasible solution, then the state feedback controller gain matrix which dependent on the parameters is

suppress wind disturbance and improve the system dynamic performance, the controller is designed based on the LPV dynamic model, making the H f norm of the closed-loop transfer function Tez ( s ) from the

f

* * *

where M V T AT U t  RT U t BT U t  X U t . If

The formula (8) shows that external interference e t exists in the LPV model. In order to effectively

Tezf s

(9)

V and R U t satisfying formula(10) for all the parameters, then the parameters of closed-loop (9) are quadratic stability and meet the given H f performance.

inductance of the stator in d,q coordinate system, i and d are the stator current, is the equivalent inductance L i

3.

( A(r (t ))  B (r (t )) K r (t ) ) x t  Le t

The affined parameters of this LPV model depend on U (t ) , and the controller matrix is solved according to theorem 1. Theorem 1(Junling Wang 2008) For the LPV model described by formula(8) and a given positive constant, if there are continuously differentiable symmetric positive definite matrix X U t , symmetric positive definite matrix Y , matrix

(7)

Where R is the stator resistance, L and L are the d q

q

x t

K U t ˜ x t

1 e %v

Tw

(12)

and according to formula (4) and low-frequency submodel,

the closed-loop system is obtained:

342

¸

¸

¸

%( wt  H ¸%8l 2  H %v

(13)





H depends on the low-frequency operating point of the

system, its value is C cp O

dC p O

J

O C cp O C p O



1

':l

H 1¬ H %( wt t  žžž  ­­­ %( wt t %8l t

žŸ J T Tw ®­ Tw

Formula (11) and (14) constitute the highfrequency sub-model of the conversion system, for this model and according to the theory above, u t '* G

Name R Tw JT *G max

T

'* wt º¼ is

y t  %M t  C S t

x t

vector,

is

L

ª¬0

2  J

Tw º¼

T

,

,

B

ª 1 *G º « » « J T * wt » « » « J *G » «¬ J T * wt »¼

  2 C¡ ¡¢ 2  H



Table 1:Experiment Parameters Value Name Value 2.5 m ȡ 1.25kg/m3 21.4286s Ș 0.95 0.5632 kg*m2 Cpmax 0.476 40Nm Ȝopt 7

The experiment results of power factor of the wind turbine C p and tip speed ratio O (lam) are shown in

defined to be the output vector, and the matrixes obtained by formula (14) are

1 JT º ª 0 A « » ¬J Tw J JT 1 Tw ¼

8l

In MATLAB simulation environment, find the solution of matrix in theorem 1 with LMI and Simulink toolbox, and build the general simulation diagram of the system, then download the simulation to the dSPACE to do the real-time simulation experiment. The parameters of the experiment are shown in Table 1.

(14)

H (G 2H  %(G t e t

J T I ( wt Tw

state

*Gref

Figure 2: Gain Scheduling Control Structure Based on LPV

2,541,330 No. OLD Nodes:>676,223 Levels: 105

It can be appreciated that making the implementations presented in this article the new TLS algorithm gives a better makespan. It can also be appreciated that in all the presented cases the new algorithm outperforms the previous ones. 5. CONCLUSIONS AND FUTURE WORK The scheduling of industrial systems is a challenging problem due to the combinatorial nature present in most of them. The experiments with a timed coloured Petri net model of a CNC machine shows that the makespan is improved when the nodes are evaluated on a smallestfiring-time basis. The implementations presented in this article are crucial in order to develop an approach such as the time line search algorithm which uses the CTSS as the search space for performing the optimization of the makespan of industrial models. In order to have the capacity to evaluate big state spaces a garbage collection algorithm for the CTSS is needed. This algorithm is being part of the current research of the authors.

362

SIMULATION AND MODEL CALIBRATION WITH SENSITIVITY ANALYSIS FOR THREAT DETECTION IN THE BRAIN Keegan Lowenstein(a), Brian Leventhal(b), Kylie Drouin(c) , Robert Dowman(c), Katie Fowler(b), Sumona Mondal(b) (a)

(a)

Clarkson University, Department of Computer Science (b) Clarkson University, Department of Mathematics (c) Clarkson University, Department of Psychology

[email protected] , (b)[email protected], [email protected], [email protected], (c) [email protected] , [email protected]

behavioral reaction times for threatening than nonthreatening stimuli. Interestingly, the bias in attentional capture has been more difficult to demonstrate than the ability of threats to hold attention (see Bar-Haim et al. 2007; Bishop 2008; Cisler et al. 2009; Wyble et al. 2008). The neurophysiological underpinning of the attentional bias towards threat has also been the focus of a number of studies. Although several key brain areas have been implicated, such as the amygdala and insula, (Öhman 2000, 2005; Phelps & LeDoux 2005), a detailed understanding of how these areas interact with the brain areas involved in perception, response generation, and attention is lacking. One approach to investigating the interactions between these brain areas is to combine experimental work with computational modeling. In this approach the computational models provide rigorous tests of hypotheses generated by the experimental work, and importantly, should provide novel predictions that can be tested in future work (e.g., Yeung et al. 2004). Little work has been done applying this approach to studying the attentional bias towards threats, and the few that have relied on qualitative fits of the experimental data rather than quantitative fits (Armony & LeDoux 2000; Dowman & ben-Avraham 2008; Wyble et al. 2008). In this study we describe our efforts at applying a connectionist model to quantitatively fit behavioral and brain activation data obtained in our studies of the attentional bias towards threats to the body (somatic threats). An important focus of this work involved comparing model architectures simulating the different functional interactions between the brain areas thought to be involved in threat detection and orienting. To accomplish this we applied optimization techniques and sensitivity analyses to allow more rigorous, quantitative comparison of the different architectures and to explore the properties of the parameter space. We explain the experimental setting in section 2 and follow in section 3 with the modeling. We present two different brain architectures and the model calibration results in section 4 with the follow-up sensitivity analysis in section 5. We end with a discussion of future directions.

ABSTRACT In this study we use optimization techniques and sensitivity analyses to provide more rigorous, quantitative connectionist models of the functional interactions between the brain areas involved in detecting and orienting attention towards threat. A toolkit has been developed using flexible neural network modeling with automated parameter estimation. A sensitivity analysis is provided within the framework to identify significant model parameters and better understand model dependencies. These studies emphasize the importance of fitting the models to behavioral reaction time and brain activation data. They also show that the specific architecture of the model, and the numerical precision in the model parameters is important in determining an acceptable fit of experimental data, and that it is not the case that any model will work given the appropriate set of connection strength parameters. Keywords: Neural Network, Minimization, Analysis of Variance

Least-Square

1. INTRODUCTION Our survival depends in part on being able to detect a threatening stimulus that occurs outside the focus of attention and to redirect attention towards the threat so it can be dealt with (Bishop 2008; Corbetta & Shulman 2008; Norman & Shallice 1989). The ability of threats to capture and hold attention can also have an impact on mental health. For example, hypervigilance towards threat is thought to play an important role in the etiology and maintenance of many anxiety disorders (Bishop 2008; MacLeod et al. 2004; Mogg & Bradley 2004). Yet despite its importance to survival and mental health there remain significant gaps in our understanding of this fundamental cognitive process (Corbetta & Shulman 2002; Öhman 2000; Phelps & LeDoux 2005). Several studies have shown that threats are better at capturing and holding attention than non-threatening stimuli (Bar-Haim et al. 2007; Bishop 2008; Cisler et al. 2009; Öhman 2000, 2005). The enhanced ability of threats to capture attention is evidenced by faster

363

to capture attention is greater when they are presented outside the focus of attention, and is consistent with the smaller reaction time validity effect for threatening targets observed in the reaction time data.

2.

EXPERIMENTATION In our somatic threat studies (Dowman 2007a, 2007b) subjects performed two tasks: a visual color discrimination task and a somatic intensity discrimination task alternating in random order within the same session. The visual discrimination task consisted of indicating whether a red or a yellow LED was lit, and the somatosensory discrimination task consisted of indicating whether a high or low intensity electrical stimulus was delivered to the sural nerve at the ankle. A symbolic cue given at the beginning of each trial signaled which of the two tasks was forthcoming. The target stimulus was correctly cued on a randomly determined 75% of the trials (validly cued condition) and incorrectly cued on the remaining 25% of the trials (invalidly cued). The subject was instructed to focus his/her attention on the cued target stimulus, but to respond to the target regardless of whether or not it was correctly cued. Note that in the validly cued condition the target stimulus was presented within the subject’s focus of attention, and in the invalidly cued condition the target was presented outside the focus of attention. Two different sural nerve electrical stimulus intensities were used. In Dowman (2007a) both were strong and threatening (one at pain threshold and the other moderately painful), and in Dowman (2007b) both were weak and non-threatening. The attentional bias towards the somatic threat was evidenced in our experimental studies by the reaction time difference between the validly and invalidly cued conditions (validity effect) being smaller for the threatening somatic than the non-threatening somatic or visual target stimuli (Dowman & ben-Avraham 2008). (Reaction time differences due to stimulus intensity and sensory modality precluded a direct comparison between the threatening and non-threatening target stimuli). The smaller validity effect is consistent with the idea that threat is better able to capture and shift attention than non-threatening stimuli. Electrophysiological measurements obtained during these experiments revealed three brain areas that appear to play an important role in detecting and orienting attention towards somatic threats. That is, for the threatening sural nerve target stimuli these brain areas exhibited greater activation when they were presented outside the focus of attention (invalidly cued) than when they were presented within the focus of attention (validly cued) (Dowman, 2007a, 2007b; Dowman & ben-Avraham 2008). The electrophysiological data suggest that somatic threats are detected by somatic threat detectors located in the dorsal posterior insula. The threat detector activity is in turn monitored by the medial prefrontal cortex, which then signals the lateral prefrontal cortex to shift attention towards the threat (Dowman & benAvraham 2008). The greater activation of the somatic threat detectors in the invalidly than the validly cued condition suggests that that the ability of somatic threats

3.

MODELING We further examined the threat detection and orienting hypothesis using artificial neural network modeling (Dowman & ben-Avraham, 2008). The model was based on the work of J.D. Cohen and co-workers on response conflict (Botvinick et al. 2001; Yeung et al. 2004). The response conflict modeling studies, in conjunction with behavioral and functional imaging measurements, have provided convincing evidence that the medial prefrontal cortex is involved in monitoring situations that require a change in attentional control (e.g., response errors, response conflict, unattended threats) and signals the lateral prefrontal cortex to make the change. We modified the Yeung et al. (2004) model by replacing the response conflict component with threat detectors. We compared several different model architectures in order to test the different physiologically feasible functional interactions between the brain areas responsible for detecting and orienting attention towards somatic threats. The architecture that provided the best qualitative fits of the reaction time and brain activation data is shown in Figure 1. mPv

1.5

ARv ●

-.12

● ARs

.16

1.5

mPs

.16

-.12 ● Rs Rv ● 2.0

1.4

1.4

Mv 2.0 Thv ●

2.0

Ms 2.0

Sv

Ths

Ss ●

.16

.16

-0.5

-0.5 Av ● -.12 ● As ●

inhibitory excitatory

Figure 1: Artificial Neural Network Model of the Threat Detection and Orienting Hypothesis. The model includes two stimulus-response pathways corresponding to the two tasks, where the SsMs-Rs nodes and their connections simulate the somatosensory intensity discrimination task, and the SvMv-Rv nodes and connections simulate the visual color discrimination task. The S and M nodes correspond to brain areas involved in early and late sensory processing, respectively, and the R nodes correspond to brain areas involved in the response. Note that the

364

model does not attempt to simulate the discrimination task performance in each stimulus modality, but rather only the reaction time differences for threatening vs. non-threatening target stimuli. The lateral prefrontal cortex areas controlling attention are simulated by attention nodes, one for each of the visual and somatosensory sensory nodes (Av and As, respectively), and one for each of the visual and somatosensory response nodes (ARv and ARs, respectively). The threat detectors for the visual and somatosensory systems are simulated by the Thv and Ths nodes, respectively, and the medial prefrontal cortex is simulated by the mPs and mPv nodes for the visual and somatosensory systems respectively. The activation for each node was computed over 55 cycles. During the first 5 cycles external inputs were added to the As or Av nodes to simulate the allocation of attention as directed by the cue. The validly cued condition was simulated by adding an external input of 1.0 to the somatosensory attention node (As) and 0.0 to the visual sensory attention node (Av). The invalidly cued condition was simulated by adding an external input of 0.0 to the somatosensory attention node (As) and 1.0 to the visual sensory attention node (Av). During the stimulus cycles (cycles 6-10) external inputs were added to the somatosensory sensory node (Ss) to simulate the presentation of the somatosensory target stimulus. A threatening somatosensory stimulus was simulated by also adding external input to the somatosensory threat detector node (Ths) during the stimulus cycles. Due to the symmetry in the model, targets were only presented on the somatosensory side. In the remaining 45 cycles, activation was allowed to spread through the model. For a model with M nodes, the activation levels of the nodes were computed using the following activation function: ‫ܣ‬௜ ൌ

ଵ ଵା௘ ሺరషಿ೔ ሻ

where c is the cycle at which the activation level of the response node equals 0.20, 20 is an estimate of the number of milliseconds per cycle (based on the brain activation data), and 500 is a constant that accounts for perceptual and decision processes that are not accounted for by the neural network model (Dowman & benAvraham, 2008). Dowman & ben-Avraham (2008) tested a number of different model architectures simulating different functional interactions between the brain areas thought to be involved in detecting and orienting attention towards somatic threats. As noted above, the architecture shown in Figure 1 demonstrated the best qualitative fits with the experimental reaction times and brain activations. Interestingly, this architecture led to the prediction that the attentional bias towards somatic threats will only be observed when the threat is presented outside the focus of attention (invalidly cued) and not when it is presented within the focus of attention (validly cued). As noted earlier, we could not directly test this hypothesis using the sural nerve stimuli because of the stimulus intensity confound. However, this prediction was recently confirmed in a study using pictures of somatic threats (Dowman et al. 2010), where the neutral and somatic threat target stimuli were matched for stimulus intensity and hue. Dowman & ben-Avraham (2008) used the same set of connection strengths for all of the architectures (see Figure 1). These connection strengths were based on those published by Yeung et al. (2004), and were modified manually to provide acceptable qualitative fits of the data. The comparisons were straightforward given that many of the architectures could not simulate the direction of change in both the reaction time and brain activation data. It is possible of course, that had we chosen a different set of connection strength parameters that another architecture would have fit the data better. Therefore, a much better approach for model architecture comparison would be to use optimization techniques to find the best fit connection strength parameters for each of the architectures. Optimization techniques have the added advantage of allowing us to search for the best quantitative fit of the experimental data, something that is not feasible when the connection strengths are adjusted manually. It is also important to perform sensitivity analyses to explore the parameter set. Of particular interest is determining whether the fit is dependent on a small range of connection strength values, or whether a wide range of combinations produce a good fit. Together, the optimization techniques and sensitivity analyses will provide a more rigorous quantitative comparison of the different architectures. Importantly they will allow us to determine if the architecture is important in fitting the data, or whether any architecture can be made to fit the data given the right set of connection strength parameters. Clearly the former outcome is of much greater interest in using the models to help determine the functional interactions between these brain areas.

ǡ (1)

where A is a column vector containing M elements which represent the activation levels for all nodes in the model for the ith cycle. ܰ௜ was defined as: ܰ௜ ൌ ܰ௜ିଵ ൅ ሺܹ ‫ܣ כ‬௜ିଵ ሻ െ ሺܰ௜ିଵ ‫ߜ כ‬ሻǡ (2) N is also a column vector of size M, ߜ is a scalar decay constant, and ܹ is an M x M weighted connection matrix. ܰ௜ିଵ represents the value of N during the cycle prior to i. Similarly, ‫ܣ‬௜ିଵ represents the value of A during the preceding cycle. The product of the weighted connection matrix and the last known activation values ሺܹ ‫ܣ כ‬௜ିଵ ሻ accounts for the input to each node due to its incoming connections. ‫ܣ‬଴ and ܰ଴ were both null vectors initially. The reaction time was defined as the cycle where the response node activation equaled 0.2. Reaction time was converted to milliseconds using the following function: ܴ݁ܽܿ‫ ݁݉݅ܶ݊݋݅ݐ‬ൌ ʹͲܿ ൅500, (3)

365

somatic target stimuli. For modeling purposes we approximated the stimulus intensity confound-free invalidly cued threatening somatic target reaction time by multiplying the increase in the invalidly cued reaction time relative to the validly cued condition for the threatening somatic target (i.e., [invalidly cued – validly cued]/validly cued) to the validly cued nonthreatening somatic target reaction time. Also, owing to the uncertain relationships between the scalp potentials used to measure the brain activations, the underlying brain activity, and the activation function in the model, we used the percent change in the electrophysiological measurements in the invalidly cued condition relative to the validly cued condition (i.e., {[invalidly cued – validly cued]/validly cued} * 100). As described in detail in Dowman & ben-Avraham (2008) the electrophysiological measures of the threat detector activation also include activation of the adjacent sensory area. Hence, this activity was modeled by combining the activations of the Ss and Ths nodes. The scalp potential measurements do not provide acceptable isolation of the medial and lateral prefrontal cortex activities, hence they were not included in the modeling studies. The original model was able to provide excellent fits of the reaction time data (least-square error ~ 1.0e9), but could not also fit the activation data (least-square error = 1.0). The failure of the original model to fit reaction time and activation data was because it could not account for the lack of change in the Ss node across the validly and invalidly cued conditions for the nonthreatening somatic targets, as was originally pointed out by Dowman & ben-Avraham (2008). Rather, in that model the Ss node was smaller in the invalidly than the validly cued conditions. Our experimental studies have reported that brain areas involved in later sensory processing do show this attention effect (Dowman 2007a). Hence, we altered the model architecture by connecting the sensory attention nodes (As) to the middle layer (Ms), where the latter simulates the later stage of sensory processing. This model is shown in Figure 2 below. The result was a much better fit of the reaction time and activation data. Specifically, of the 20 best-fit parameter sets obtained from the Nelder-Mead algorithm, three gave a least-square fit to within the measurement error (i.e., least-square error ~ 1.0e-5, modeled reaction times within 5 milliseconds). Dowman & ben-Avraham (2008) also compared architectures where the threat signal from the medial prefrontal cortex (mPs) to the response attention node (ARs) to one that has the threat signal going to both the sensory (As) and response attention nodes. This architecture is shown in Figure 3. The latter architecture is more consistent with the known anatomical connections between the medial and lateral prefrontal cortices (Miller & Cohen 2001). Dowman & benAvraham (2008) could not find any difference between the two architectures. We re-ran this simulation using the best-fit connection strength parameters for each architecture to see if this would make a difference, and

4.

MODEL CALIBRATION Our previous effort involved manually adjusting connection strength parameters to provide acceptable fits, and then using these parameters to compare the different model architectures. This process was slow, tedious and at best led to rough qualitative fits of the data. More recently we developed a MATLAB ®-based toolkit that provides automated calibration of connection strengths using optimization techniques (Lowenstein 2010). The optimization involved minimizing ଵ

‫ܬ‬ሺܹሻ ൌ  σ௉௜ୀଵ ቀ ௉

௘೔ ି௠೔ ଶ ௡೔

ቁ , (4)

where W is the matrix of connection strengths, P corresponds to the number of statistics that are being fit, ݁௜ is the experimental value of a given statistic, ݉௜ is the corresponding modeled value (which depends on W), and ݊௜ is a normalization factor which ensures that each statistic contributes equally to the cost function. For our purposes, this normalization can be accomplished by setting ݊௜ equal to ݁௜ . Within the toolkit, the Nelder-Mead simplex method is used for the minimization of Eq. (4) (Nelder & Mead 1965). Nelder-Mead has previously been shown to be effective in parameterizing connectionist models (Bogacz & Cohen 2004). A benefit of NelderMead is that no gradient information is needed and minimization is based solely on function evaluations using a simplex that changes at each iteration based on the best point found. The function to be minimized can be non-differentiable, non-convex, or even discontinuous. This is an attractive feature of the toolkit because it allows for a general framework for the model calibration. Thus changes made to the simulation tool itself will have little, if any impact on the calibration process. For each model, the five optimal parameters sought were asr, at, in, smr, and tmar (see Figure 2). NelderMead is well known to be a local optimization method that can be highly dependent on an initial simplex. Consequently, multiple optimizations are usually required to better search the design space. Here 20 optimization runs were obtained for each model. To find the starting values for each Nelder-Mead optimization run, 1000 connection strength parameter sets were randomly chosen and the fits calculated. The set with the lowest J(W) value was used as the starting values. 4.1 Numerical Results First we determined whether the original architecture (Figure 1) provides a good, quantitative fit of the reaction time and brain activation data. As noted above, the stimulus intensity confound prevented us from directly comparing behavioral reaction times obtained for the non-threatening and threatening

366

the response attention node. The same result was obtained here when trying to quantitatively fit the reaction time and threat detector/sensory activation data (least-square error ~1.0e1).

indeed it did: the architecture sending the threat signal to both the sensory and response attention nodes provided a noticeably better quantitative fit of the reaction time and activation data.

Table 1: Best Fit Model Parameters and Least-Square Error for Architecture in Figure 3 Parameter asr at in smr tmar J(w)

Optimal Value 0.5640 -0.2119 -0.4327 0.0256 0.2073 2.0702e-6

Table 2: Comparison on Experimental and Model Reaction Times.

Figure 2: Modified Architecture to Sensory Attention Applied to the Late Sensory Processing Stage.

Experimental Condition

Reaction Time (milliseconds)

Valid Non-Threat Exp.

694.2

Valid Non-Threat Model

694.8

Invalid Non-Threat Exp.

826.1

Invalid Non-Threat Model

826.1

Valid Threat Exp.

694.2

Valid Threat Model

693.4

Invalid Threat Exp.

770.5

Invalid Threat Model

770.7

Interestingly, we consistently found that almost all of the architectures that we tested could provide excellent fits of the reaction time data when the best fit connection strengths were used (least-square errors < 1.0e-8). However, only the 2 architectures described here provided acceptable fits of the reaction time and brain activation data. Clearly, the brain activation data appears to provide critical constraints on the model.

Figure 3: Architecture to Account for the Threat Going to Both the Sensory and Response Attention Nodes For this model, 7 of the 20 optimization runs produced acceptable fits. The least-square errors for these parameters were an order of magnitude smaller than the previous model (2.0702e-6 vs. 6.8900e-5, respectively) (Table 1), where the modeled reaction times were within 1 millisecond of the experimental data (Table 2) and the modeled percent change in brain activations equaled the experimental data. We also examined one of the architectures that Dowman & ben-Avraham (2008) reported was unable to fit the reaction time and activation data. In this architecture the threat signal from the medial prefrontal cortex was sent to the sensory attention node instead of

5. SENSITIVITY ANALYSIS The optimization results described above show that the model architecture is critical in obtaining fits of the reaction time and brain activation data. We next sought to determine the range of connection strength parameters that produced acceptable fits of the data. The mean + SD of the connection strength parameters for the architecture producing the best fit of the data (see Figure 3) is shown in Figure 4. The 7 optimization runs that produced acceptable fits (least-square errors = 2.1–7.5e-6) were all tightly clustered around the same values. This was not the case for the 13 runs that

367

resulted in unacceptable fits (least-square error = 6.69.1e-3). Recall that the starting values for the NelderMead optimization were determined by 1000 iterations of randomly selecting parameter values and computing the least-square error, and using the values that produced the best fit as the starting point in the optimization. This strategy reduces the probability that the optimization will always converge on the same local minimum. Hence, the tight coupling of acceptable fit parameters around the same values strongly suggests that range of best-fit connection strengths is very narrow. This result was confirmed with a sensitivity analysis. When developing and studying mathematical models, it is common in practice to perform a sensitivity analysis to gain a deeper understanding of the model behavior, regardless of whether optimization is part of the design process. Analysis of variance (ANOVA) is one approach to studying the impact of changes in model parameters on model output. Specifically, ANOVA may reveal that some parameters have little effect on the overall model while others have a profound effect. In such cases, certain insignificant parameters can be set to a reasonable value while optimization can be done to fit the sensitive parameters and thereby reduce the problem size for the leastsquares problem. This approach can also determine the specificity of the connection strength parameters. That is, does each connection have to be within a tight range for the model to work, or can changes in one connection be offset by a change(s) elsewhere in the model. This analysis can have a significant impact on interpreting the functional significance of the connection strength values. A benefit of ANOVA is that only sets of parameters and output are required as opposed to needing any derivative information. ANOVA compares the ratio of the variation between sample means to the variation within each sample. The starting point for the procedure is to sort each parameter into groups. Analysis is done by considering changes in a response (here the leastsquares error) as the group changes. Specifically, ANOVA is a hypothesis test with null hypothesis, H0 : μ1 = μ2 = μ3 = · · · = μk, where k is the number of experimental groups. Each μ represents the mean of the single parameter, often called a factor, that is being found by the values in each experimental group. When rejecting the null hypothesis, the alternative hypothesis states that at least one mean is different from another, however it does not specify which one. The experimental groups are different equally spaced intervals for a single variable. The ANOVA examines the source of variation by finding the sum of squares of deviation from the mean for each of these groups. Using a statistical F test, the procedure is able to determine whether or not at least one mean is deviating from the others. The F test will produce a p-value; If this value is below a significance of 0.05 then the null hypothesis is rejected.

The model calibration experiments described above revealed that small changes in even the third decimal place of the connection strengths could strongly impact the overall least-square error and result in a poor model fit. For the sensitivity study presented here, tight bounds were place on each parameter based on the best point found. We provide the details of the sensitivity analysis for the architecture in Figure 3, since it provided the overall best fit to the experimental data. For the sensitivity analysis, we chose the best of the 7 parameter sets that provided acceptable fits. The bounds are shown in the second and third columns of Table 3 below. For the analysis, each parameter was divided into 8 equally spaced groups and 500 values of each parameter were chosen via a Latin hypercube sampling, giving 2,500 parameter sets. The Kolmogorov-Smirnov normality test was applied to the response variable (here the least-squares error) and we found the data was not generated from a normally distributed population. Thus, the nonparametric ANOVA method, the Kruskal-Wallis test, was used to calculate the corresponding p-values, shown in the last column of Table 3. Three parameters had values close to zero (asr, at, and in) indicating that they are significant in the modeling process. However, smr has a p-value of 0.052, which is very close to our level of significance 0.05, and we still can consider it to be a sensitive parameter. The parameter tmar was identified as insensitive and this was also evident in the values identified by the optimizer for the seven best parameter sets found during optimization. For the significant parameters, the standard deviation was always less than 0.03 but for tmar it was 0.07, indicating that a range of values would still lead to reasonable fitting to the data. These results intuitively make sense because tmar and smr are feed-forward connections, the rest are bidirectional. Clearly the positive feedback associated with a bidirectional connection will make it much more sensitive to change than a feed-forward connection. Furthermore, at is particularly sensitive since it is largely and only responsible for the brain activation fit. Table 3: Lower and Upper Bounds for Parameter Study on Architecture in Figure 3 Parameter asr at in smr tmar

Lower Bound 0.5 -0.3 -0.5 0.02 0.2

Upper Bound 0.6 -0.2 -0.3 0.03 0.29

pvalue ≈0 ≈0 ≈0 0.052 0.493

To this end, an interval plot can provide a deeper insight into how the response values are distributed. We show these for smr and in in Figures 5 and 6. Here, the shaded, red dots correspond to points values of J(W)

368

while the average for each group is shown with a blue '†' . For smr, the average values are relatively constant, with small fluctuations across the groups, but there is actually a broad range of response values within a group. For in, the average values change significantly across the groups, which is expected since in was identified as a sensitive parameter. It is important to note that of the 2,500 parameter sets randomly chosen from within the bounds given in Table 3 the leastsquare errors were greater than 1.0e-2, which is four orders of magnitude greater than the best-fit values. The sensitivity analysis for in produced an unexpected result. Sensitivity analysis is often used to guide the starting parameter set values for optimization. That is, the factor (parameter range) showing the best fit values is chosen in the optimization. However, sensitivity analysis suggests that the best fit point(s) would lie within factor 7 whereas the optimal value for in actually falls within factor 2 (-0.4327). The interval plots (Figure 6) reveal that the neighborhood around this point is considerably small thus requiring high accuracy in the optimization process. This implies that caution must be applied when interpreting the sensitivity analysis results for models with a very narrow range of best-fit values. An important clue that the sensitivity analysis results may not provide meaningful information on selecting starting points and/or bound constraints for the optimization was that even the best J(W) values were four orders of magnitude greater than the optimal value.

Figure 5: Range of Response Values Across Groups for smr

Figure 6: Range of Response Values across Groups for in

ACCEPTABLE FIT (n=7) UNACCEPTABLE FIT (n=13)

2.0

6.

DISCUSSION We have developed a flexible toolkit to develop and test artificial network models of the brain mechanisms for detecting and orienting attention towards threats to the body. Using optimization techniques were able to provide excellent quantitative fits of behavioral reaction time and brain activation data. These studies demonstrate that the model architecture is critical in producing good quantitative fits of the reaction time and brain activation data. Indeed, of the several models examined by Dowman & ben-Avraham (2008), only 2 provided acceptable fits. However, essentially all of the architectures could fit the reaction time data given optimal set of connect strength parameters. Clearly, including the brain activation data is critical in obtaining meaningful results in this type of work. The sensitivity analysis suggests that only a very narrow range of connection strength parameters will fit the data. This implies that for fits of reaction time and brain activation at least, it is not case that the acceptable fits are an artifact of having a large number of parameters to fit the data. These results strongly suggest that the sensitivity analysis should not be used to determine the starting parameter values and ranges when the range of optimal values is very narrow. Future

CONNECTION STRENGTH

1.5 1.0 0.5 0.0 -0.5 -1.0 -1.5 IN

AT

ASR

SMR

TMAR

CONNECTION PARAMETER

Figure 4: Mean + SD Connection Strength Parameters For the Best-Fit Architecture.

369

Dowman, R., Quin, J., & Sieg, E. (2010) Mechanisms Underlying the Capture of Attention by Somatic Threats. American Psychological Society 22nd Annual Meeting, Boston MA May 27-30. Dowman, R., & ben-Avraham, D. (2008) An artificial neural network model of orienting attention towards threatening somatosensory stimuli. Psychophysiology, 45, 229-239. Koster, E.H.W., Crombez, G., Verschuere, B., Vanvolsem, P., & De Houwer, J. (2007) A timecourse analysis of attentional cueing by threatening scenes. Experimental Psychology, 54, 161-171. Lowenstein, K. (2010) A Toolkit for Developing Neural Network Models of How the Brain Detects Threat. Honors Thesis. Clarkson University. MacLeod, C., Campbell, L., Rutherford, E., & Wilson, E. (2004) The causal status of anxiety-linked attentional and interpretive bias. In: Cognition, Emotion and Psychopathology. Theoretical, Empirical and Clinical Directions (J. Yiend, Ed.) New York: Cambridge University Press, 172-189. Miller, E.K., & Cohen, J.D. (2001). An integrative theory of prefrontal cortex function. Annual Reviews of Neuroscience, 24, 167-202. Mogg, K., & Bradley, B.P. (2004). A cognitivemotivational perspective on the processing of threat information in anxiety. In: Cognition, Emotion and Psychopathology. Theoretical, Empirical and Clinical Directions (J. Yiend, Ed.) New York: Cambridge University Press, 68-85. Nelder, J.A., & Mead, R. (1965). A Simple Method for Function Minimization. Computer Journal, 7,308313. Norman, D.A., & Shallice, T. (1986) Attention to action. In: R.J. Davison, G.E. Schwartz, & D. Shapiro (Eds.) Consciousness and Self-Regulation. Advances in Research and Therapy. Plenium Press: New York, pp. 1-18. Öhman, A. (2000) Fear and anxiety: evolutionary, cognitive, and clinical perspectives. In: M. Lewis & J.M. Haviland-Jones (Eds.) Handbook of Emotions, New York: Guilford Press, pp. 573-593. Ohman, A. (2005) The role of the amygdala in human fear: automatic detection of threat. Psychoneuroendocrinology, 30, 953-958. Phelps, E.A., & LeDoux, J.E. (2005) Contributions of the amygdala to emotion processing: From animal models to human behavior. Neuron, 48, 175-187. Wyble, B., Sharma, D., & Bowman, H. (2008) Strategic regulation of cognitive control by emotional salience: A neural network model. Cognition and Emotion, 22, 1019-1051 Yeung, N., Botvinick, M.W. & Cohen, J.D. (2004). The neural basis of error detection: Conflict monitoring and the error-related negativity. Psychological Review, 111, 931-959.

work will include understanding the interaction of parameters. Future experimental studies are aimed at testing predictions derived from the model. Of most interest is the prediction that the attentional bias towards threats are only seen when the threat is presented outside the focus of attention (Dowman et al. 2010). We are also performing modeling studies to determine if the model can simulate the attentional bias towards threats that have been reported using other experimental paradigms (e.g,. Koster et al. 2007). ACKNOWLEDGMENTS This research work was supported by National Science Foundation UBM (Undergraduate BiologyMathematics) Grant DBI-0926568 and the Clarkson University Honors Program. REFERENCES Armony, J.L., & LeDoux, J.E. (2000) How danger is encoded: Toward a systems, cellular and computational understanding of cognitiveemotional interactions in fear. In: M.S. Gazzaniga (Ed.) The New Cognitive Neurosciences, Cambridge Massachusetts: The MIT Press, pp. 1067-1079. Bar-Haim, Y., Lamy, D., Pergamin, L., BakermansKranenburg, M.J. & van LJzendoorn, M.H. (2007) Threat-related attentional bias in anxious and nonanxious individuals: A meta analytic study. Psychological Bulletin, 133, 1-24. Bishop, S.J. (2008) Neural mechanisms underlying selective attention to threat. Annuals of the New York Academy of Sciences, 1129, 141-152. Bogacz, R. & Cohen, J.D. (2004) Parameterization of connectionist models. Behavior Research Methods, Instruments, & Computers. 36, 732-741. Botvinick, M.M., Braver, T.S., Barch, D.M., Carter, C.S., & Cohen, J.D. (2001) Conflict monitoring and cognitive control. Psychological Review, 108, 624652. Cisler, J.M., Bacon, A.K. & Williams, N.L. (2009) Phenomenological characteristics of attentional biases towards threat: A critical review. Cognitive Therapy and Research, 33, 221-234. Corbetta, M. & Shulman, G.L. (2002) Control of goaldirected and stimulus-driven attention in the brain. Nature Reviews Neuroscience, 31, 201-215. Corbetta, M., Patel, G., & Shulman, G.L. (2008) The reorienting system of the human brain: From environment to theory of mind. Neuron, 58, 306324. Dowman, R. (2007a). Neural Mechanisms Of Detecting and Orienting Attention Towards Unattended Threatening Somatosensory Targets. I. Modality Effects. Psychophysiology, 44, 407-419. Dowman, R. (2007b). Neural Mechanisms Of Detecting and Orienting Attention Towards Unattended Threatening Somatosensory Targets. II. Intensity Effects. Psychophysiology, 44, 420-430.

370

SIMULATION AND MODELLING OF THE FLAT-BAND VOLTAGE FOR BELOW 200nm SOI DEVICES 1

C. Ravariu1, F. Babarada1 Politehnica University of Bucharest, Faculty of Electronics and Telecommunications, Splaiul Independentei 313, 060042, Bucharest, Romania, E-mail: [email protected] and [email protected]

analysis between the classical model, new model and the simulation results are presented, as novelty.

ABSTRACT The nowadays SOI technologies frequently offer below 200nm, even up to tens of nanometre, for film on insulator. The flat-band voltage is one of main parameter in the electrical characterization of the SOI devices. The conventional models for this voltage were established for thicker structures, with 0.5...2Pm Si-film thickness and 23ȝm buried oxide thickness. The electric charge from the buried oxide was ignored because the interesting conduction occurs in the vicinity with the front oxide. The pseudo-MOS transistor is a dedicated device for the electrical characterization of SOI wafers and works with a buried channel. The downscaling consequences of the SOI sizes on the flat-band voltage modelling were studied in this paper, with applications on the pseudo-MOS device.

2. The analytical model The definition of the flat band voltage is related to the compensation of the positive electric charges from the buried insulator in order to bring the film surface potential to zero volts. Adapted from the SOI-MOSFET to the pseudo-MOS transistor, this parameter is classically expressed as, [8]:

VFB

Q it2 Q it    ) MS 2C s 2qH Si N A

(1)

where NA [cm-3] is the doping concentration in substrate, Qit [e/cm2] is the surface electric charge density, ĭMS [V] is the metal-semiconductor work function, Cs=Hox/xox [F/cm2] is the specific oxide capacitance, HSi/ ox is the dielectric permittivity of Silicon, respectively oxide, xox is the BOX (Buried Oxide) thickness and q=1,6x10-19C is the elementary electric charge. Firstly, in the analytical model ĭMS =0V will be assumed.

KEY WORDS Modelling, simulation, SOI, buried interface, devices

1. Introduction The algorithm of modelling and simulations of physical processes spread over a large spectrum of applications, [1], [2]. The SOI structures represent a promising candidate for the nanodevice implementation, as classical [3] or novel architecture, [4]. This miniaturization is possible only if it is accompanied by proper models for the developed electronics devices. Some device parameters lost or change their classical meaning for new technologies. For example, the threshold voltage cannot be defined in a SON Transistor, [5], but it still arises in the transfer characteristics of a pseudo-MOS device, [6].

At the interface Si/SiO2 usually exists a positive electric charge, Qit, due to the presence of two kinds of charges: the interface charge Qt representing the electrons trapped on the fast surface states and the fixed charge Qf representing an excess of the ionic silicon solved in oxide and frozen at the Si/SiO2 interface during the end of the annealing. The global charge is noted in this paper by Qit = Qt + Qf. Its sub-components can be: Qt1=109y1010e/cm2, Qt2=1010y1011e /cm2, Qf1=1010e/cm2, Qf2=1012e/cm2=102 e/nm2; where the index “1” is used for the upper SOI interface and “2” for the bottom SOI interface. Frequently, the effect of Qt on VFB is neglected. For example, the contribution of Qt charge is just 0,01V in VFB value for the density of states 1010eV-1cm-2 in a bulk MOSFET with NA=1015cm-3 and xox=100nm.

An excellent device for the electrical characterization of the SOI wafers is the pseudo-MOS transistor. This device represents an up-side-down SOI-MOSFET, with a backgate command and electrical conduction through the film bottom, [7]. This paper comparatively presents some new analytical models, versus some simulation results, regarding the flat-band voltage of a pseudo-MOS transistor.

Therefore, in this paper we will work with the total positive electric charge Qit1, Qit2 were considered, fig. 1. The flat-band voltage represents that gate voltage, which reduce to zero volts the potential in the Si-film, equivalent

A new reference model will be confronted with others, analytically deduced in the next paragraph. A comparative 371

with V(0)=0 and E(0)=0 in fig.1. By integration of the Poisson’s equation, yields: VFB

· § Q it1 · § qN A 2 Q ox1  Q ox 2 ¨ ¸ ¨ x d ¸¸ ¨ C ¸  ¨ 2H x d  H Si s ¹ © Si ¹ ©

The traditional Lim and Fossum model completely ignores the second interface, considering Qit2=0 and also the depletion of the substrate, xd2=0, [6]. Hence, VFB is -Qit1/Cs. This happened at the beginning of the SOI structures, with Micronics sizes. Obviously, “Qit” is a model parameter in (1) and hasn’t a physical meaning. It is named “the global charge from BOX”, but from eq. (1) it must be measured in [C/cm2], being a superficial electrical charge density. A first disagreement between models (1) and (7) consists in different values of Qit, and Qit1+Qit2. Considering additionally the electric charge from the second interface Qit2z0, from the limit conditions the accurate model is (7). The classical model (1) systematically underevaluates the flat-band voltage value. Additionally, Qit from the first and second ratio in eq. (1) hasn’t quite the same values. Another correction concerns the “2” factor that is missing in the model (7), first fraction at denominator, due to an average value assigned to Qit. In fact, either interface comprises fixed charges Qf and interface trapped charges, Qt. Consequently, the model (7) can be detailed as:

(2)

The first parenthesis represents the potential drop over the buried oxide and the second parenthesis is the potential drop over substrate. The notations correspond to the fig. 1, where xd is the width of the depleted region in substrate. Neutral Qit2 Si-film Qit1 S ND + Buried + n + OXide + +

xfilm

Substrate

-------

NA p

xd2

xS2-xd

xox

-xfilm 0

xox

V G< 0

G

xox+xd

x

VFB

Fig. 1. The analyzed SOI structure with positive fixed charges in BOX and negative ions in substrate.

Q it1 Ÿ E ox

H Si E SB ( x ox )  H ox E ox ( x ox ) E SB ( x ox )

Q it1 H ox

(3)

Q it 2 Ÿ

Q it1  Q it 2 H Si

(4) VFB

From Gauss’law for x  ( x ox , x ox  x d ) results: E SB ( x )



Q  Q it 2 qN A ˜ ( x  x ox )  it1 H Si H Si

Q it1  Q it 2 qN A

(5)

(6)



Q f 1 Q it1  Q it 2 2  Cs 2H Si qN A



Q it1 / 2 Q it1  Q it 2 / 2 2  Cs 2qH Si N A

(9)

In the following simulations, a reverse way was investigated: the interface global charge densities were selected for different pseudo-MOS transistors and the flatband voltage was extracted from definition. The scope was to accomplish the best fitting between VFB simulated and VFB analytical.

By replacing xd from (6) in (2), the final expression of VFB is obtained: VFB

(8)

In this way, two targets are reached: the problem of “2” missing at denominator of first ratio of model (7) is solved and a better agreement between simulations and the analytical model is obtained; the second ratio from model (7) overestimate the flat-band voltage, while the second ratio from model (9) brings the analytical values closer to the simulation results. The insight for Qit1 must be Qf1+Qt1 and for Qit2 must be Qf2+Qt2.

From the limit conditions: ESB(xox+xd)=0 in (5), the xd expression results: xd

Q f 1  Q t1 Q f 1  Q t1  Q f 2  Q t 2 2  Cs 2H Si qN A

where Qt1, 2 respectively are the electric charge densities due to the electrons captured on the fast-states from the Si-film/BOX and BOX/Substrate interfaces. The correct value of the fixed charge density, Qf 1, 2 must be extracted from VFB parameter after the Qt1, 2 subtractions from VFB in eq. (8). In the spirit of the classical model (1), model (8) could be corrected by averaging:

The limit conditions give the electric field: H ox E ox (0)  H Si E Si (0)



(7)

372

In this case a global values Qit1, 2 values were established, as is modelled in (9).

3. Simulations The simulated SOI structure had: ĭMS1=ĭSM2=-0,32V, as is shown in fig. 2, ĭs-s=0, selecting the same p-type semiconductor as film and substrate with NA=5x1015cm-3.

Therefore, the discussion regarding the subtracting of the Qt1, 2 from the global density Qit 1, 2 resting just at theoretical level. However the simulation can reveal some discrepancies between the classical model (1) and the proposed models (7) and (9).

The interface charge densities were chosen accordingly with some typical experimental results, [9]. The total front charge Qit1=5u1010e/cm2 placed at x=0 in fig.1 and the total bottom charge, Qit2=5x1011e/cm2 placed at x=xox in fig.1, was selected. None charge in the front oxide was select, in order to be focused just on the buried oxide.

Figure 3 a presents the current flow density through the Si-film in the case of biased structure at: VS=0V, VD=0,5V, VG=-1,5V. The conduction prevails through the film bottom as is expected. Figures 3,b and c provide the adopted method for the extraction of the simulated flatband voltage, VFBsim. The gate voltage was increased in modulus till the film potential becomes zero. Then, the potential graph was translated with -0.32V value, correcting the metal-semiconductor work function, in order to extract the simulated flat-band voltage, VFBsim=1.92V, affected just by the surface electric charges densities, Qox1, 2.

The simulations started with an SOI structure having xfilm=200nm and continue to 50nm, xox=400nm, xSB=750nm. Figure 2 presents the simulation results for a pseudo-MOS with 200nm. In these conditions, a holes distribution still arises along the structure.

Fig.2. The holes concentration in 200nm Si-film structure.

(a) (b) (c) Fig.4. The 2nm Si-film structure at VS=0V, VD=0.5V, VG=-1.33V: (a) potential distribution, (b) holes concentration near source, (c) holes concentration near drain. Accordingly with fig. 4.a, at VG=-1,33V applied on the back gate, the hole concentration p>1016cm-3>NA. Hence a lower flat-band voltage is searching. Finally, VFB sim =-0.85V for previously mentioned Qit1, 2 values.

6. Discussions

(a) (b) (c) Fig. 3. (a) Detail of the current density in the 50nm SOI Si-film biased at VS=0V, VD=0,5V, VG=-1,5V; (b) the potential distribution across the SOI near the source contact; (c) the potential distribution across the SOI through the middle.

For the investigated SOI structures, with xfilm=50nm, xox=400nmnd SOI with xfilm=2nm, xox=4nm, the same amount of positive interface charge density was used for both structures in order to provide a comparison. These 373

contributions were centralized in the table 1. Here can be compared some situations simulated and computed for different sizes. The notations are: VFB sim for the simulated value of VFB, VFB(1) for the value deduced with the model (1), VFB(7) for the value deduced with the model (7), VFB(11) for the value deduced with the model (9).

In conclusion the charge placed at the bottom interface BOX/Substrate has a maximum influence on VFB parameter extraction in the thin SOI films and it is partially annihilated by the negative inversion layer formed at the substrate surface during the device operating, in thicker SOI films.

xfilm xox Qit1 Qit2 VFB(1) VFB(7) VFB(11) VFBsim (V) (nm) (nm) (ecm-2) (ecm-2) (V) (V) (V) 50 400 2·1010 5·1011 -0.191 -4.51 -1.31 -1.82 50 400 5·1010 5·1011 -0.501 -5.55 -1.84 -1.95 50 400 2·1010 1012 -0.191 -16.3 -4.32 -3.92 2 4 2·1010 1012 -0.007 -15.9 -4.14 -1.95 2 4 5·1010 1012 -0.042 -16.9 -4.63 -2.11 2 4 1·1010 1011 -0.002 -0.18 -0.056 -0.01 Table 1: Comparisons for structures with 50nm and 2nm film thickness.

Acknowledgment. This work is supported by the POSDRU /89/1.5/S/62557, PN2 no. 12095, 62063.

References [1] Antonio Cimino, Francesco Longo, Giovanni Mirabelli, A General Simulation Framework for Supply Chain Modeling: State of the Art and Case Study, International Journal of Computer Science Issues, Volume 7, Issue 2, No 3, pp 1-9, March 2010. [2] Michael Affenzeller, Stefan Wagner, Stephan M. Winkler, Effective allele preservation by offspring selection: an empirical study for the TSP, International Journal of Simulation and Process Modelling 2010 - Vol. 6, No.1 pp. 29 - 39. [3] F. Babarada, et al., MOSFET Modelling Including Second Order Effects for Distortion Analysis, IASTED Proc., Applied Simulation and Modelling 2006, Rhodes, Greece, pp. 506-510. [4] C. Ravariu, ey al., Modelling and simulation of a nanostructure for a single electron technology 5th International Mediterranean implementation, Modelling Multiconference, EMSS, Briatico, Italy, 16-19 Sept, 2008, ISBN 978-88-903724-0-7, pp.312-315. [5] J. Pretet, S. Monfray, S. Cristoloveanu and T. Skotnicki. Silicon-On-Nothing MOSFETs: performance, short channels effects and back gate coupling. IEEE Trans. Electron Devices, vol.51, no.2, pp. 240-245, 2004. [6] C. Ravariu, A. Rusu, Parameters extraction from some experimental static characteristics of a pseudo-MOS transistor, Bucharest, UPB Scientific Bulletin, ISSN 1454-234X, Series C, vol. 70, no. 1, pp. 29-34, 2008. [7] S. Sato, K. Komiya, N. Bresson, Y. Omura, S. Cristoloveanu, Possible influence of the Schottky contacts on the characteristics of ultrathin SOI pseudo-MOS transistors, IEEE Transactions on Electron Devices, vol 52, no.8, pp. 1807-1808, 2005. [8] H.K.Lim, J.G.Fossum, Threshold voltage of thin-film silicon-on-insulator (SOI) MOSFET’s, IEEE Trans. Electron. Devices, vol. ed-30, no.10, October, 1993. [9] CEA-Leti R&D, 20nm Fully Depleted SOI process, EUROSOI, Newsletter, October, vol XXVI, 2010.

Table 1 highlight that the analytical model (1) always underestimate the flat-band voltage, considering all the time just the first interface charge, Qit1. The inclusion of the second interface charge Qit2, with some correct limit conditions but in the depletion approximation, systematically overestimate the flat-band voltage accordingly the model (7). The best model at thick or thin sizes is the analytical model (9). For thinner SOI films, a lower interface area results, within a lower quantity of negative ionic charge in substrate, in order to fulfill the flat-band conditions. In the ultra-thin SOI structures, the substrate isn’t inverted, being in incipient depletion regime. Hence, the depletion approximation used in the deduction of the model (7), is more justified in ultra-thin SOI films than in thicker films. In SOI nanofilms the components Qf1, Qf2 change the balance of importance on VFB parameter. In thick BOX, some values like Qit1=1010e/cm2, Qit2=1012e/cm2, influence the potential of Si-film mainly via Qit1 parameter. In the case of some nanometres thickness of film and BOX and a device area=10x10nm2 the prior charges densities are: Qit1=1012e/cm2=0.01electrons/ device area – in probability terms quite negligible and Qit2=1012e/cm2=1electron/ device area – has a strong activity through a 2-5 nm thickness of buried oxide.

6. Conclusion The classical model (1) induces high errors in thin SOI films because it entirely ignores the back charge interface that was true at thick BOX. The model (7) accurately deduced by Poisson equation integration systematically overestimated the flat-band voltage, because it use the depletion approximation in substrate and ignore the inversion layer arisen at the substrate surface. Simulator that proved accumulation of electrons at the substrate surface surprises the superposition. Therefore, the best model is (9), based on the averaging of the known interface charge, Qit1 and Qit2. 374

SIMULATION OF HUMAN BEHAVIOR IN SITUATION OF EMERGENCY Samira Benkhedda, Fatima Bendella, Karima Belmabrouk SIMPA Laboratory, Department of Computer Science, University of Science and Technology of Oran, Algeria. [email protected], [email protected], [email protected]

It represents the complexity of a phenomenon through the interaction of a single set of entities called agents. Each agent can: - Communicate with other agents to exchange information. - Perceive and act on all or part of the environment - Apply knowledge, skills and other resources to perform their individual personal goals. The objective of the multi-agents simulation is : -to Infer the nature of the functioning of the entities of a complex system. -System Analysis. Simulator knows two phases of use : - A research phase in which the simulator acts as an incubator model. - An operational phase: once the model is validated, the simulator becomes a tool in the field.

ABSTRACT Multi-agent systems are used to study the complex natural and social phenomena. In this context, simulations are based on agent-oriented representation to describe the characteristics of the situation that all actions can be performed by actors. These tools have been used in situations of crises that are usually very difficult to manage, because of their complexity or the damage on help. In this paper. We propose a new approach for simulating multiple agents in a medical emergency based on practical reasoning of human and using the notion of simulation of complex systems. Keywords: Multi agent systems, simulation, simulation of complex systems, medical emergency

2.1. Simulation of complex system Simulate a complex system is to model its components, their behaviors and interactions between them and with their environment and then run the model obtained numerically. A feature of these systems is that one cannot predict the evolution of the modeled system without going through this phase of simulation. The approach "experimental" simulation makes it possible to reproduce and to observe complex phenomena (eg biological or social) in order to understand and anticipate their evolution [3]

1. INTRODUCTION In a Multi-Agent System (MAS), agents interact and cooperate to perform a task or to achieve a common goal. In this paper we present our simulation of a medical emergency that is ‘heart attack’, because if no action is taken immediately in emergency assistance and if do not act soon, the victim's life is in danger in a short term. We describe the features which we've featured in our simulation, by determining the scientific value of the project and identifying our agents, their roles and interactions between them.

3. 2.

CONCEPT OF MULTI-AGENTS SIMULATION The multi-agent simulation is a simulation that employs the concept of multi-agents systems in the conceptualization, specification and implementation. A multi-agents system simulated living in a simulated environment; the multi-agents simulation directly represents the people, their behavior, their actions in the environment and their interactions. The multi-agents simulation is Interactions, Agents and Environment [1]. Ferber [2] notes that the multi-agents simulation allows the study of complex systems.

MULTI AGENT SIMULATION SYSTEM FOR RAPIDLY DEVELOPING INFECTIOUS DISEASE MODELS IN DEVELOPING COUNTRIES (IDESS) [4]

3.1. Model Overview and context IDESS (Infectious Disease Epidemic Simulation System) is a system able to build a simulation model to detect infectious diseases from the existing data in a geographical area. IDESS is characterized by: - The ability to create a simulation model for any location worldwide.

375

- Uses existing data to generate the simulation model. - The ability to view the results in several ways. -From a software engineering perspective, it can change the behavior of agents and interactions with others.

the plan and around them there is no agent who knows the plan, they move according to two strategies: always running to the west or east still running, if they meet they descend the stairs, for if cons in their neighborhood, an agent knows the plan, then they follow it. When agents encounter a fire, they change their direction, increasing their speed and find another way out. The strategy to choose another way is to take the stairs or the nearest fire is not yet declared. Other agents that meet these agents will learn the information lights and follow them.

3.2. IDESS Implementation They have implemented the approach in a multi-agent system application-specific. At the heart of this simulation, they used the agent person (Person Agent PA) that acts like a normal person; the PA interacts with other PA in the model simulation and the environment in terms of Agent City (Town Agent TA). PA and TA have parameters and interactions that are associated with each agent. TA agents vary according to their consciences on the changes in the population of the city. The agent TA has connections with other agents TA and PA. The interaction between the TA may change if a containment strategy was invoked by isolating and TA officer concerned. IDESS was used to quickly build a simulation model based on agents that can be used in the investigation of the spread of disease in a given geographic. This approach is flexible in its ability to model any geographic location by processing the unpredictable nature of the spatial location where an outbreak of an infectious disease will occur. This system is dynamic because you can edit and add new agents and information to the model.

4.2. IFI Implementation This work aims to build a simulation that is closest to reality as possible so the authors used the BDI architecture (Belief, Desire and intention) because pedestrians must communicate and reason to find the way to the exit, and then Agents are getting smarter. In this model, the environment is the building of the IFIs and the fires are considered an agent. There are two types of agents: 1) Agents who know the plan of the building. 2) Workers who do not know the plan of the building. Each agent type has a different algorithm, but they have the same process to observe, to update the world (the state building and state of the stairs that wants to spend). If an agent is affected by the fire that minimal heat, it will be hurt, and the heat of fire is greatest, the agent will die. The application is flexible; we can change the building plan by modifying the xml file. You can also change the number of agent in each room, the number and position of obstacles, fire, and the percentage of agent type and percentage rate of agent. The reactions of the agents respond to environmental changes, they are reasonable, they can avoid congestion, avoid fire and teach others. The reasoning of the agents is following the reasoning of real people. The application meets one important limitation: The information exchanged between agents are small, there is little communication. If agents communicate and exchange much more information, the model becomes more real.

4. AGENT-BASED SYSTEM FOR THE EVACUATION OF THE BUILDING IN CASE OF FIRE (IFI) [5] 3.1. Model Overview and context This system is designed to simulate agent-based building evacuation in an emergency, and more specifically they simulated the building of the IFIs in the case of the fire. In this work, they modeled and simulated fire. They built the model with the map of the IFI building (3 floors with 18 rooms, two staircases, and 3 outputs). The initial state is that all agents are in the rooms. When the program starts, an emergency occurs. All agents try to leave the room. They determine the nearest door to get out of them. When moving they have to avoid obstacles. If they want to pass the door is full (many agents want to spend, there is more room) they have to go to another door if available. After leaving the room, the agents determine the direction to move. With agents who know the plan of the building, they measure the distance between them is the stairs. They take the stairs closest to you. The strategy is that all agents who know the plan always use the shortest path to the exit. In the event that there are many agents who pass the stairs, congestion occurs, some agents waiting at the tail end will choose another staircase. If agents do not know

5. MULTI-AGENTS SIMULATION IN THE CASE OF HEART ATTACK 5.1 The heart attack [6] A heart attack is a serious problem caused by a blood clot in a coronary artery or one of the chambers of the heart. Cardiac arrest is the sudden stoppage of the heart that pumps more blood. 5.2. Signs of heart attack [6] Defibrillation and cardiopulmonary resuscitation (CPR) quickly can save lives. Prompt treatment to break up clots can greatly increase the chances of survival of the person who suffers of heart attack. Since prompt

376

treatment can make a difference, it is important to know the early signs of heart attack. In case of heart attack, you may experience one or more of the following: • Discomfort in the center of the chest that lasts more than 5 minutes or comes and goes. It takes the form of an uncomfortable pressure, tightness, a feeling of heaviness or pain. • Discomfort in other parts of the body, such as pain or discomfort in one or both arms, neck, at the jaw or stomach. • Shortness of breath is often accompanied by chest discomfort but can occur before the discomfort. Among other signs, call the cold sweats, nausea and a sensation of floating. Women who are a heart attack may not experience the usual symptoms, which can delay their care. Among the symptoms: an atypical or unusual pain in the chest, abdominal pain, nausea, shortness of breath and unexplained fatigue.

• The emergency agent and the doctor agent: aid workers are expert-like agent, they are reactive agents. • The no-expert agent: an agent's type BDI (Belief, Desire and intention), the agent has beliefs about the world in which it operates, it must meet the desires by making intentions).

5.3. Mobile-Learning [7] Mobile-learning is a logical extension of e-learning. In this sense it refers to the provision of courses or learning objects through mobile devices such as PocketPC, cell phones or the PalmPilot Users will have lessons in reduced format, but the main advantage of such a solution is accessible at any time of day and from any location with a network is nearby. Today the knowledge of students in mobile technology (accessibility, ease of use, speed of adaptation) makes the m-learning possible.

Figure1: Architecture of Complex SIMUL.

5.4.4. Why a BDI agent type? This architecture is the most valued. In theoretical point of view, a BDI agent can perform any type of task, The architecture allows the agents to solve complex problems. In this architecture, agents have a feature that allows evaluating the utility of each action. Contrary to the cognitive architecture of agents, the agents ask random actions when they were not able to achieve their goals. BDI agents can determine the action (or actions) to be performed to get as close as possible goal. This means that when an agent can achieve its goal by setting an action, then it will select the action as close as possible to the goal or action which will achieve this goal as quickly.

5.4. Complex SIMUL The main objective of our project (Complex SIMUL) is to simulate an emergency based on the simulation of complex systems. Complex SIMUL includes reactive agents that will work in a complementary way and deliberative agents (BDI type) that will incorporate the concept of practical reasoning in humans. Our system includes three different types of agents (figure1):

5.4.5. The operation of our Succourer Agent like a BDI agent (figure2) Our agent includes an event queue (the actions of the agent to whom he rescued victim) by storing the internal events of the system, the beliefs (knowledge of the agent), a library of plans (know-how of the agent ), a battery of desires (goals of the agent) and a stack of intentions (instantiated plans to achieve goals). The BDI interpreter cycle begins by updating the event queue and beliefs of the agent. It then activates new desires by selecting the plans of the library, so our agent has the opportunity to decide by running the first selected action stack of intentions, and so on. This agent must: 1. Observe the environment agent and the victim agent for the possible risks. 2. Report a warning (phone call). 3. Select a plan (and it depends on his desires).

5.4.1. Victim Agent It is a passive actor, it can take three different cases (a patient with a blue face, red or a deceased victim), it is reactive, and it reacts to the actions of the agent rescuer (move, help him ...). 5.4.2. Environment Agent it is the agent that identifies the location of the accident, it is reactive type and it is dynamic (the road, the greenery ...). 5.4.3. Succourer Agent It is the active player of our emergency; there are three types of agents:

377

4. Select a plan by running the corresponding plan of action (the ABC of first aid). Our no-expert agent is going to learn during the simulation using his mobile phone through the MLearning. During the M-Learning, our BDI agent will learn along the ABC of first aid. We used the concept of priority for emergency agents, the doctor agent has the highest priority then the emergency agent but no-expert agent will take the lowest value of priority zero boots (0) then after every successful in emergency and in the same case this value will increment to 1 and so on. The priority of the no-expert agent is less than or equal to the priority of the emergency agent but the last two priors are always less than that of the doctor agent. The BDI agent will protect the victim agent and his self from danger (the middle of a road) then it will use the ABC of first aid.

CARDIOPULMONARY RESUSCITATION (CPR) The CPR is a combination of chest compressions with rescue breathing done on the victims is believed to be in cardiac arrest. When cardiac arrest occurs, the heart stops pumping blood. CPR can support a small amount of blood flow to the heart and brain to "buy time" to restore normal heart function [9]. 8. CONCLUSION In this paper, We have tried to show the methodology for the design of our simulation, we have implemented a multi-agent simulation and an application of MLearning which is flexible and intuitive enough to allow the rescuer to the first aid to save the life of a victim of a heart attack because the speed of treatment is very important in this case. We currently implement the agents of our simulation and integrate them into a learning application.

REFERENCES [1] Wafa Ketata, Wided Lejouad and Chaari. 2007. UneOntologie pour la réutilisation des Interactions dans un Système Multi-Agents. JFO, 18-20 octobre 2007, Sousse, Tunisie. [2] J. Ferber, 1999. Multi-agent systems. Reading MA : Addison-Wesley. [3] Alain Boucher, NGUYEN Nhu Van , 2007. L’interaction dans simulation multi agents. Hanoi. [4] Dean Yergens, Julie Hiner, Jörg Denzinger et Tom Noseworthy. Multi Agent Simulation System for Rapidly Developing Infectious Disease Models in Developing Countries. The Second International Workshop on Multi-Agent Systems for Medicine,Computational Biology, and Bioinformatics. [5] NGUYEN Thi THUY Nga et HO Tuong Vinh, Juillet 2009. SMA pour la simulation à base d’agents d’évacuation de bâtiment dans les cas d’urgence. Hanoi,. [6] McKesson Provider Technologies, 2006. La crise cardiaque : Signes annonciateurs précoces. [7] J, MUESSER., 2005. Conception et réalisation d’un objet pédagogique pour périphérique mobile. projet profesionnel IUP PSM 3éme année.

Figure2: The BDI agent of Complex SIMUL.

6. THE ABC OF FIRST AID The priorities of first aid are… A AIRWAY B BREATHING C CIRCULATION (and bleeding) We will explain only the first operation and the Cardiopulmonary resuscitation CPR Airway The airway of an unconscious person may be narrowed or blocked, making breathing difficult and noisy or impossible. This happens when the tongue drops back and blocks the throat. Lifting the chin and tilting the head back lifts the tongue away from the entrance to the air passage. Place two fingers under the point of the person’s chin and lift the jaw, while placing your other hand on the forehead and tilting the head well back. If you think the neck may be injured, tilt the head very carefully, just enough to open the airway [9].

378

[8] B.David, juillet 2006. Mobile-learning pour des activités professionnelles. cours architecture informatique, école d’été du CNRS « EIAH ». [9] http://tilz.tearfund.org/Publications/Footsteps+1120/Footsteps+18/The+ABC+of+first+aid.htm.

AUTHORS BIOGRAPHY Dr. Fatima Bendella is a Senior Lecture in the Department of Computer science in USTO; she got an engineer diploma in computer science at the University of Oran in 1988, a magister of USTO in 1995 and a doctorate in 2005. She directs several theses of magister and doctorate in the application of multi-agent systems in software development. She is responsible of many research projects and a national research project (PNR) approved in May 2011.

379

Simulation, Optimisation and Design a Platform for in-vivo Electrophysiological Signals Processing F. Babarada1, C. Ravariu1, J. Arhip2 1

University Politechnica of Bucharest, Faculty Electronics Telecommunications and Information Technology, DCAE, ERG, Bucharest, Romania 2 S.C. Seletron Software si Automatizari SRL, Bucharest, Romania

Abstract— The paper presents a hardware solution of the in vivo electrophysiological signals continuous processing and using a data vector acquisition on PC. The originality of the paper comes from some blocks proposal, which selective amplify the biosignals. One of the major problems in the electrophysiological monitoring is the difficulty to record the weak signals from deep organs that are covered by noise and the cardiac or muscular strong signals. An automatic gain control block is used, so that the high power skin signals are less amplified than the low components. The analog processing block is based on a dynamic range compressor, containing the automatic gain control block. The following block is a clipper since to capture all the transitions that escape from the dynamic range compressor. At clipper output a lowpass filter is connected since to abruptly cut the high frequencies. The data vector recording is performing by strong internal resources microcontroller including ten bits A/D conversion port. Design of analogical blocks is assisted by electronics circuit’s simulation and optimization.

frequency of 60Hz and amplitude of 1 – 3 x EMG signal, [4]. Ɣ Motion artefact: It has two main sources: electrode/skin interface and electrode/cable, having a frequency range of 0 – 20Hz. It is reducible by a proper circuitry and set-up. Ɣ Inherent instability of signal: All electronics equipments generate noise and the amplitude is somewhat randomized, being in correlation with the discrete nature of the matter. This noise has a frequency range of 0 – 20Hz and cannot be removed. In this situation, a gastric signal for instance, recorded at skin level, is hundreds times lower than the parasitic signals. The most accurate solution is the invasive one, straight to the target organ, using microelectrodes, [5]. Unfortunately, the majority of organs are inaccessible without a surgical act which adds two great disadvantages: the health-state in danger and high costs. This paper presents an analog processing and digital recording system for low power electrophysiological signals, with the possibility to use them in medical applications like ECG, EGG, EMG etc. For low contact electrodes area, the noise introduced by the electrodes begins to be most significant. As the results of modelling of the ensemble source-electrode, it is recommended for the amplifier to be implemented by a low noise and distortions, transimpedance amplifier stage followed by one low passing filter. For very low electrophysiological signals it is necessary a differential amplifier because it has a high common mode rejection of parasitic signals characteristic [6].

Keywords— Simulation, Design, Health care, Compressor technique, Electrophysiological signal.

I. INTRODUCTION

The common techniques from the human electrophysiology are non-invasive, with electrodes placed on the tissue (e.g. metallic electrodes in contact with gastric mucosa in electro-gastro-graphy [1] or at cutaneous level in the classical electro-cardio-graphy ECG [2]). Therefore, a main problem arises when the electrodes are placed onto skin: the useful weak signals are buried in high level parasitic signals. The non-invasive electrophysiological methods suffer from noise, collected by the surface electrodes. There are many types of noise to be considered: Ɣ Inherent noise in electronics equipment: It is generated by all electronics equipment and can’t be eliminated. It is only reduced by high quality components using. It has a frequency range: 0 – several thousand Hz, [3]. Ɣ Ambient noise: The cause is the electromagnetic radiation, with possible sources: radio transmission, electrical wires, fluorescent lights. It has a dominant

II. THE ELECTROPHYSIOLOGICAL SIGNALS PROCESSING

As in the case of many concepts from engineering, automatic gain control was also discovered by natural selection. A. The automatic gain control

380

Automatic gain control (AGC) is an adaptive system found in many electronic devices. The average output signal level is feedback to adjust the gain to an appropriate level

V1 10Vdc

0

8 6

10Vdc

0

7

V+

VEE

T1

Vcaa

5 R8 22k

VCC R9 22k

Vccaa V2

D7 D1N4148

D4 D1N4148

R12 47k

D6 D1N4148

R11

0

22k C13 0.47u

R10 4.7k

0

VEE

+

OVC

Ro-scale 18k LT1115/LT OUT T2 2 -

0

4

VEE

0

0

VCC U22

VCC 3

V

R13 10k

D8 D1N4148

1

T2

7 V

V-

0

LT1115/LT OUT 3 OVC +

0U12

Level Vcaa 6 8

LT1115/LT OUT T2 2 -

R35 1k

R14 100k VEE

5 Rf-scale 6 8

2.2k

T1

VCC

-

5

OVC

Rmeter 50

1

10k

+

1

2

7

U10 VCC

R6

C12 0.47u

7

OVC

R7 10k VEE

R5 100k

R4 1k

D1N4148

V+

0

+

5

3

V-

7

VCC

8 6

D3 D1N4148

4

T2 OUT

Out AGC

4

VEE

VCC U13

D5 6 8

T1

LT1115/LT 3

-

5

V-

OUT T2

4

10k

U9 Vcaa

-

OVC

V+

4

2

V-

5

R2

+

LT1115/LT 2

T1

OVC

8 6

V+

T2 OUT

D2 D1N4148

V-

10k VEE

1 T1

2 VOFF = 0 VAMPL = 0.04 LT1115/LT FREQ = 1k 3 +

V+

0

V-

4

VEE

R3

1

D1 D1N4148

1

10k V

T1

3

V+

7

VCC U11

R1 V3

for a range of input signal levels. AGC algorithms often use a proportional-integral-differential controller.

R36 10k

0

Fig. 1 The automatic gain control In order to display the output level of the dynamic range compressor signal we added the stage composed with operational amplifier U22 that adapt the connection and the adjusting of zero and end of scale for a linear display with LEDs bargraph. The fig. 2 presents the beginning action of the automatic gain control for the imput amplitude signal 20mV.

The basic components of compressor are the U9, U10 integrated circuits, which biases the D1, D2 diodes, fig. 1, at their I-V curve knee. The input voltage is in the range of 10 to 300mVp and the output voltage is in the range of 5 to 10mVp. The voltage command of AGC is in the range of 300 to 600mVdc. The resistor R3 allows the circuit to be balanced and adjust the output voltage so it does not produce distortion in the output when gain reduction is active. In order to provide the voltage command of AGC (Vcaa) we choose a feedback configuration design. This design contain the amplifier, composed by U11, R4, R5 with the amplification around 101, the full wave rectifier, composed by D3, D4, R6, R7, U12 which bring the signal to the absolute value and the positive voltage detection realized with D6, C12 connected trough half voltage divider R8, R9. The voltage over the condenser C12 is exactly the voltage command of the automatic control amplifier Vcaa. Discharging of the condenser C12 is made through the diode D7. This diode is opposite polarized by a voltage greater than Vcaa, respectively the voltage produced by diode D5, which is not reduced by half and loaded at the absolute value the condenser C13 through resistances R10 and R11. At reduction of the input signal amplitude the voltage Vcaa remains constant until C13 is discharged by R11 and R10. Thus at transient simulation at 1kHz, the amplification remains constant 5ms and then increase in time of 15ms. To reduce the temperature dependence of the automatic gain control we used the IC stage realized with U13, the diode D8 and the resistance R12, achieving the circuit from fig. 1. The voltage command Vcaa can be adjusted from resistive divider composed with resistances R13 and R14.

2.0V

0V

-2.0V

-4.0V 0s V(D3:1)

V(V3:+)

5ms V(R9:2)

10ms

15ms

20ms

25ms

30ms

Time

Fig. 2 Simulation of output voltage CCAA and the detected voltage VCAA for the input signal amplitude of 20mV.

B. The clipper A clipper circuit was added to catch all the transitions that escape out of the dynamic range compressor. This is done with two diodes D10 and D11, connected in parallel, fig. 3. Each diode is reverse biased so that they do not drive until they reach a certain amount of tension. This voltage is set using a resistive divisor consisting of R20 and R21 to a value of approximately 1V and may be adjusted to set the threshold for clipper. It is applied to the uninverted input of the operational amplifier U18 and from the output of U18 to the inverted input of the second operational amplifier U19. Between the AGC system and the clipper, an adapting gain block is installed. If this block is operating at unity

381

gain, it is practically transparent and the clipper is operating

at threshold.

C15 0.01u

7

R20 2.7k

T2 OUT

-

D9 3

0

+

D1N4148

0

OVC

5

R25 2 10k

T2 OUT OVC

7

+

6 8

Vnet

10k

LT1115/LT 3

VEE R28 8 6

R27

1k

10k

5

0

U19 VCC 7

0

-

5

T1

V-

VEE 8 6

U18 VCC

Clipper Level

V+

7

10k LT1115/LT

OVC

1

V-

T1

2 C16 100p

+

LT1115/LT OUT T2 2 4

R23

10k

3

1

VEE

R24 10k

T1

R22 10k

R21 10k

R18

10k

D1N4148 D10

10k

VCC

Clipper Output D1N4148 D11 R26

4

R17

VCC U17

R19 Clipper Input

6 8

V-

4

VEE

5

V+

-

OUT T2

1

2

OVC

4

U15 VCC

+

V+

LT1115/LT

7

0

3

5

T1

T2 OUT OVC

+

V+

LT1115/LT 3

VCC U16

8 6

V-

0.01u

HF Expanding

1

VEE

1

2.2k

-

T1

2

V-

C14

V+

Vccaa

4

R16 1k R15

Fig. 3 The clipper Adjusted, the upper audio frequencies are increased and resulting the family of curves from fig. 4. The reason for introducing the resistance R22 and diode D9 is for the clipper output signal temperature compensation. C. The lowpass filter Since the clipper circuit can create higher harmonic to the output, we add a filter to cut frequencies over 3KHz. Must be a lowpass filter, with a flat response and an abrupt shape from the maximum passing frequency. Therefore we choose a filter of order five Chebyshev with 0.2 dB wave amplitude in passing band, fig. 6. The low pass filter circuit was optimized in order to have the best lowpass filter transient response.

C18 0.02u

0.016u

C19 1800p

LT1115/LT OUT T2 2 -

0

4

VEE

6 8

10k

+

OVC

10k C20 0.03u

C21 LT1115/LT OUT T2 730p 2 -

0

VEE

T1

OVC

5 6 8 V

1

C17

0

+

10k

VCC U21

R33 3

T1

10k

V

V-

10k

R32 5

4

3

7

7

VCC U20

R31

V+

R30

1

R29 Vnet

V+

When the resistor R16 in the schematic fig. 3, is set to zero ohms, this circuit is simply a unity gain block. As R16 is

V-

Fig. 4 The clipper driver frequency response

Fig. 6 The lowpass filter Fig. 7 Frequencies response of the whole chain of signal processing

Frequency response of the entire chain corresponds with the block level simulations and makes the designed behaviour. Thus the compressor frequency response is smooth over 500KHz, the driver stage of clipper emphasizes high frequencies and low-pass filter cut abruptly the frequencies over 3KHz, fig. 7.

III. THE PC INTERFACE

The electrophysiological signals acquiring begins from the source of the bioelectric signals coupled with the electrodes, amplification, processing, analog-digital conversion and data storage in some file format. For the electrophysiological studies, the data storage is necessary, for a long time, as vector data storage. Acquisition and data storage are performed by an 8-bit microcontroller series AVR (Atmel), namely ATMega32 on a development board that has its own power source, a real

200V

150V

100V

50V

0V 1.0Hz V(VCCAA)

10Hz V(VNET) V(U21:OUT)

100Hz V(V3:+) V(VCAA)

1.0KHz

10KHz

100KHz

1.0MHz

Frequency

382

time clock circuit, an EEPROM memory, a LED display and a serial interface adapter (RS232 or RS485), fig. 8. This microcontroller has strong internal resources, allowing data acquisition and digital conversion through a 10-bit ADC, provided with eight inputs multiplexer and its own high accuracy reference voltage reference [6, 7]. Different interesting voltages are collected to internal DAC by means of microcontroller port A, ADC0 to ADC7. The conversion of analog data to a digital vector is synchronized by an internal clock which allows for choose different sampling rates. A conversion cycle starts by clearing the memory locations for the measured values. After that, every input is converted in a 10 Bits word and temporary stored into the internal RAM memory then the next input is also converted, and so on. At this moment we have an eight 10 Bytes words representing a sample of the analog entry signal. This word is now completed with the conversion time, extracted from the external “Real Time Clock” (RTC), the U10 chip.

IV. CONCLUSIONS

Usually, only non-invasive electrophysiological methods can be accepted, in respect with the tissue particularities. The paper was focused on signal processing because the electrophysiological signals have a high dynamic range and can be easily covered by the artefacts noise. The presented vector data collection, processing and recording have the possibility to use many input channels that give the possibility to simultaneously test different versions of source-electrodes-amplifier blocks. Later this facility can be used to multipoint measuring or to increase the resolution. A specific data vector recording was presented, with the advantage of development for new remote methods in electrophysiology.

ACKNOWLEDGMENT This work was supported by projects 62063, 12095 financed by Romanian National Authority for Sic. Research.

REFERENCES 1.

2. 3.

Fig. 8 The digital recording module using the microcontroller with integrated Port-A analog-digital converter

The vector obtained looks like: 5AYYMMDDHH mmsshhV0V1.....V7, [8]. The whole record is now 24 bytes long and it is stored into the external flash memory U2. This memory has 65536 bytes allowing for over 2700 records. At a rate of 20 samples/second that means there is enough space for more of 2 minutes of records. The acquired data can be extracted by a serial link, the chip U3 providing for the RS232 specification, including hardware handshake by RTS-CTS pair. The communications parameters have been choose to meet the MODBUS specification, as is: 1 START bit, 8 data bits, 9600 Bauds, 1 Even parity bit, 1 STOP bit. During the recording process time information and recorded values are displayed cyclic.

4.

5.

6.

7. 8.

383

KvČtina, J.; Varayil, J.E.; Ali, S.M.; Kuneš, M.; Bureš, J.; Tachecí, I.; Rejchrt, S. & Kopáþová, M. (2010). Preclinical electrogastrography in experimental pigs. International Journal of Interdiscip. Toxicol., Vol.3, No.2, (June 2010), pp. 53-58. Al. Rusu, N. Golescu, C. Ravariu, (2008) Manufacturing and tests of mobile ECG platform, IEEE Conf. Sinaia Romania, 2008, pp 433-436 Bogdan, D.; Craciun, M.; Dochia, R.I.; Ionescu, M.A. & Ravariu, C. (2009). Circuit design for noise rejection in electromyography, Proceedings of INGIMED 2009 2nd National Conference on Biomedical Engineering, pp. 76-81, Bucharest, Romania, ICPE-CA Publisher, November 12-14, 2009 Merletti, R. & Parker, A.P. (2004). Electromyography: Physiology, Engineering, And Non-invasive Applications, IEEE Computer Society Press, New York, USA B. Firtat, R. Iosub, D. Necula, F. Babarada, E. Franti, C. Moldovan, (2008) Simulation, design and microfabrication of multichannel microprobe for bioelectrical signals recording, IEEE Int. Conf., Sinaia, Romania, 2008, pp 177-180 F. Babarada, J. Arhip, (2009) Electrophysiology Signal Data Vector Acquiring, Congress of Romanian Medical Association, Bucharest, 2009, pp 82 Rustem Popa, (2006) Medical Electronics. Matrix House, Bucharest ATMega32 data sheet, http://www.atmel.com/dyn/resources/prod_ documents/doc2503.pdf

A SIMULATION-BASED FRAMEWORK FOR INDUSTRIAL AUTOMATED WET-ETCH STATION SCHEDULING PROBLEMS IN THE SEMICONDUCTOR INDUSTRY a)

Adrián M. Aguirre, a)Vanina G. Cafaro, a)Carlos A. Méndez*, b)Pedro M. Castro

a) INTEC (Universidad Nacional del Litoral - CONICET), Güemes 3450, 3000 Santa Fe, Argentina. b) UMOSE, Laboratório Nacional de Energia e Geologia, 1649-038 Lisboa, Portugal *

[email protected] which must be respected by robots for all transfer movements. Another constraint adding more complexity to the system operation is that baths must process wafer lots one by one, during a predefined period of time, avoiding the overexposure in the chemical ones, which can seriously damage or contaminate the wafer lot. In spite of this, wafers can stay longer than its processing time only in water baths. So, a zero wait (ZW) and local storage (LS) policy must be strictly satisfied in every chemical and water bath, respectively. As a direct consequence, an effective schedule of material movement devices and baths along the entire processing sequence will provide a better utilization of critical shared-resources and, at the same time, an important reduction in the total processing time. In the last years, different methods have been developed to achieve convenient solutions to this challenging problem. Main approaches to large-sized problems lie mainly on heuristic and meta-heuristic methodologies, such as the ones presented by Geiger et al. (1997) and Bhushan and Karimi (2004). In these works, tabu search (TS) and simulated annealing (SA) procedures, together with other different algorithms, were developed to provide a quick and good-quality solution to the job sequence problem and also, a feasible activity program for the robot. A more recent approach under the concepts of Constraint Programming (CP) was developed by Zeballos, Castro and Méndez, (2011) to handle the sequencing problem of jobs and transfers in the AWS. This method could obtain better results than the ones reported by Bhushan and Karimi (2004) for industrial problem instances in a reasonable CPU time. To the best of our knowledge, efficient systematic solution methods need to be developed to represent and evaluate the complex dynamic behaviour of the AWS. Thus, a discrete event simulation environment becomes a very attractive tool to analyze the impact of different solution schemes in the system. In this work, a modelling, simulation and optimization-based tool is developed to validate, test and improve the daily operation of the AWS, allowing an easy evaluation of different operative schemes and possible alternative scenarios. To do this, a discrete event simulation model was developed by using most of the tools and capabilities that are available in the Arena simulation environment. The principal aim is to provide a highly dynamic and systematic methodology to reach the best feasible schedule of limited resources by testing

ABSTRACT This work presents the development and application of an advanced modelling, simulation and optimizationbased framework to the efficient operation of the Automated Wet-etch Station (AWS), a critical stage in Semiconductor Manufacturing Systems (SMS). Lying on the main concepts of the processinteraction approach, principal components and tools available in the Arena® simulation software were used to achieve the best representation of this complex and highly-constrained manufacturing system. Furthermore, advanced Arena templates were utilized for modelling very specific operation features arising in the process under study. The major aim of this work is to provide a novel computer-aided tool to systematically improve the dynamic operation of this critical manufacturing station by quickly generating efficient schedules for the shared processing and transportation devices. Keywords: Discrete-event simulation, Semiconductor Manufacturing System (SMS), Automated Wet-Etch Station (AWS), Arena Software. 1. INTRODUCTION Semiconductor wafer fabrication is perhaps one of the most complex manufacturing systems in the modern high-tech electronics industry. Wafer facilities typically involve many production stages with several machines, which daily perform hundreds of operations on wafer lots. Moreover, different product mixes, low volume of wafer lots and hot jobs are some of the typical issues arising in this type of system. Wet-Etching represents an important and complex operation carried out in wafer fabrication processes. In this stage, wafer’s lots are automatically transferred across a predefined sequence of chemical and water baths, where deterministic exposure times and stringent storage policies must be guaranteed. Hence, automated material-handling devices, like robots, are used as shared resources for transferring lots between consecutive baths. An important process restriction is that each robot can only transport a single wafer lot at a time and it cannot hold a wafer lot more than the exact transfer time. Due to the lack of intermediate storage between consecutive baths, this condition can be considered as a non-intermediate storage (NIS) policy in every bath,

384

different measures of effectiveness and performance rates for the system. Thus, the paper is organized as follows: Section 2 introduces the major features of the problem addressed. Then, Section 3 describes the proposed solution method, highlighting its advantages in comparison with other existing methods and tools as well as the main objectives of this work. Later, the simulation structure is explained in detail in Section 4. A brief description concerning the simulation tool is presented. Software integration and principal interfaces between different tools are discussed. A detailed analysis regarding external and internal logic of the model and the implementation of this solution in a discrete-event simulation environment is also presented. In Section 5, an alternative solution strategy is tested using several examples, with the main idea of validating the model and, at the same time, comparing results of different solution methods. Finally, the solutions generated and the comparative study results are reported in Section 6. Conclusions and future work are stated at the end.

For this problem, it is assumed that each wafer lot, also called job, i (i=1,2…N) has to be processed in every bath j (j=1,2…M), by following a predefined processing sequence. In addition, it considers that a single robot (r=1) is available, which has to perform all the transportation activities in the system. Consequently, the problem to be faced corresponds to the scheduling of N jobs in M baths, in a serial multiproduct flowshop, with ZW/LS/NIS policies. The use of a single shared robot with finite load capacity for the wafer movement between consecutive baths is explicitly considered in this work. 3. PROPOSED SOLUTION METHODOLOGY This work introduces an efficient discrete-event simulation framework, which faithfully represents the actual operation of the automated Wet-etch Station (AWS) in the wafer fabrication process. The main advantage of this computer-aided methodology is that it permits to systematically reproduce a highly complex manufacturing process in an abstract and integrated form, visualizing the dynamic behaviour of its constitutive elements over time (Banks et al. 2004). The proposed simulation model represents the sequence of successive chemical and water baths, considering the automated transfer of jobs. Based on a predefined job sequence, which is provided by an optimization-based formulation, the model structure allows the evaluation of many different criteria to generate alternative efficient schedules. The major aim here is to efficiently synchronize the use of limited processing and transportation resources. This methodology allows also evaluating and improving the operation and reliability of baths and robot schedules. What is more, simulation runs permit addressing industrial-sized problems with low computational effort. As a result, a basic model is generated to achieve an effective solution to the whole AWS scheduling problem. It becomes also very useful for making and testing alternative decisions to enhance the current process performance.

2. PROBLEM STATEMENT The AWS scheduling problem provides a complex interplay between material-handling limitations, processing constraints and stringent mixed intermediate storage (MIS) policies (Figure 1). We can summarize major features of the system in the following way: -Material-handling devices (robot) can only move one wafer lot at a time. No intermediate storage is allowed between successive baths. So, NIS policy is applied between consecutive baths. -Waiting times are not allowed during the transportation of a wafer lot. -Robots and baths are failure-free. -Setup times are not considered for robots. -Every bath can only process one wafer lot at a time. -A ZW storage policy must be ensured in chemical baths whereas LS policy is allowed in water baths.

4. THE SIMULATION-BASED FRAMEWORK In order to formulate a computer-aided representation to the real-world Automated Wet-Etch Station (AWS) described above, it was decided to make use of the simulation, visualization and analysis tool set provided by the Arena discrete-event simulation environment (Law et al., 2007, Kelton et al., 2007). The simulation model developed in Arena Software provides an easy way to represent the AWS by dividing the entire process in specific sub-models (Initializing, Transfer, Process and Output). For each sub-model, the detailed operative rules and strategic decisions involved are modelled using the principal blocks of Arena Simulation Tool and, at the same time, a set of visual monitoring objects is used to measure the

Figure 1: Automated Wet-etch Station (AWS) process scheme.

385

utilization performance of all baths and resources in the system. Additionally, the model allows working with a user-friendly interface with Microsoft Excel for simultaneously reading and writing different data. In next sections, we will describe these features in detail.

chemical and water bath and a job sequence provided by a MILP model, which is considered as an initial alternative solution. Then, the discrete-event simulation model generates as many entities as wafer lots are to be scheduled. Here, the logic behind the automated transfer of jobs is performed in order to generate a feasible schedule for the robot activities. The subsequent simulation module is the Transfer sub-model, which defines the needed delay time to transfer a wafer lot to the next bath. This module is used to explicitly simulate the time spent to transfer the jobs between the input buffer to the first bath, between successive baths, according to the predefined sequence, and also between the last bath to the output buffer. Only after the transfer is finished, the bath from where the wafer comes is released. It should be noted that a transfer can be only executed if the robot and the destination bath are both available. In order to simulate the process itself, one Process sub-model for each bath is defined. There is a different logic depending on the type of bath (chemical or water). The wafer residence times in chemical baths must be controlled strictly (as soon as chemical bath finishes, the wafer must transferred to the succeeding water bath). While holding time in water baths is allowed. Thus, for every baths, the logic performs the following tasks: (i) reports the time at which the process begins and ends; (ii) seizes the following bath after the delay time finishes; (iii) performs the transfer to the following bath, only if the robot and the destination bath are empty. It is important to notice that the logic driving in the Process sub-model permits to easily identify why and when a given wafer's lot is discarded. Basically, it may occur because the robot and/or next bath are not available. This allows making a detailed analysis about the behaviour of the system, executing, if necessary, the corresponding adjustments when unexpected events occur or when different strategies are tested in the way to improve the process performance. So, Process submodels permit to evaluate and also validate the feasibility of the internal logic algorithm proposed in the Initializing Process of the system, identifying the possible causes of infeasibility to be corrected.

4.1. Software integration The simulator allows an easy communication with Excel spreadsheets. Thus, this tool permits reading, writing and processing important data for the simulation model. Figure 2 illustrates the data flow between Excel and Arena. Both tools support Visual Basic for Applications (VBA) that can be used to move data between them. As shown in the figure, a hybrid solution framework is proposed on these tools. The Mixed integer linear programming (MILP) model provides an initial solution that is written in Excel as input data of the Arena’s model. Using that input data, Arena simulation software runs the process model to generate many important statistics that are collected by Excel files as output data. The procedure of reading and writing data is used to dynamically generate a solution schedule by updating the start and finish times of every job in each bath and, simultaneously, determine the status of each job in every stage of the system.

Figure 2: Information exchange between Excel – Arena – MILP Software 4.2. Proposed simulation model As shown in Figure 3, the entire logic of the simulation model is divided into four main modules (input, transfer, process and output). The first module is the Initializing sub-model. The initializing process receives as input data the processing time of each job at each

Figure 3: Partial size view of the in-progress simulation model generated in the Arena environment.

386

The last module is the Output buffer. The logic of this sub-model represents the final stage of each job. At this module the final processing time (Makespan) of each job is reported. It is the ending point for entities created at the input module. Here, the model reports if the current job has been successfully finished or has been discarded.

On the other hand, if the LS rule is applied to a water bath j, inequality (2) must be satisfied.

ts( i , j 1) t ts( i , j )  tp( i , j )  S ( j ) j

ts( w, j ) t ts( w 1, j 1)  S ( j 1) j 1,2,3...M  1; w 1...N

(3)

So, any transfer of a job processed in the pw position, at bath j, has to wait the ending of the transfer of the job located in the pw-1 position at the succeeding bath j+1 to be processed. In the next section, we will explain the transfer comparison algorithm developed to solve the described problem. Only one robot is considered to be available for the execution of the transfers in the system. 4.2.2. Generation and evaluation algorithm for transfers This algorithm is mainly based on the major ideas of the JAT (Job-at-a-time) algorithm, developed by Bhushan and Karimi (2004). The JAT algorithm always prioritizes the transfers related to jobs that were previously inserted in the system, following a predefined processing sequence. For transfers related to the same job, they are executed according to the fixed sequence of baths to be visited (j=1…M+1). So, based on processing constrains (1)-(3) and assuming that all the jobs follow the same processing stages, no job in the pw position may leave the system before the one located in the pw-1 position. This means that all jobs will be processed in the different baths following the same p sequence, what is known as "flowshop permutation schedule". Our algorithm, as the JAT algorithm, selects a job to be processed and then generates (Generation Process) and evaluates (Evaluation Process) all the transfers for this job, one at a time, before going to the next job of the sequence. The principal difference between the proposed algorithm and the JAT algorithm is the evaluation procedure used for the system transfers. In the proposed evaluation process, every selected transfer is compared with all the transfers previously inserted into the system. Thus, a detailed schedule of the robot operations is defined. The aim of this process is to avoid that any transfer previously inserted (w´,j´) can be performed (for pw´” pw

ts( i , j )  tp (i , j )  S ( j ) j 1,3,5...M  1; i 1...N

(2)

Let p = {p1, p2, p3,…,pN} define a permutation processing sequence N different jobs. pw represents the w-th position of a job i (i=1…N) in the processing sequence. It means that the job processed in the w-th position will be always before the job processed in the w+1 position in the sequence p. Due to the NIS policy in the transfers and constrains on finite load capacity of the baths and the robot, the equation (3) is to be defined.

4.2.1. Advanced internal logic for the robot The principal aim of modelling the internal robot logic is to explicitly represent the finite capacity of transportation resources for transferring jobs between consecutive baths. The sequence and timing of transfers will depend on the stringent storage restrictions to be satisfied in the baths (ZW / NIS / LS) as well as on the availability of a transportation resource to carry out the transfer. Since there is only a single robot to do all the job movements, the sequence in which the transfers will be performed needs to be clearly defined. Transfers related to a particular job can never overlap because they are carried out after the corresponding processing stages finish. Consequently, no pair of transfers of the same job may be performed simultaneously. Therefore, the sequencing problem of transfers must only be focused on the comparison of transfer activities of different jobs in order to determine a feasible robot schedule. For that reason, a complex internal logic for the robot was embedded in the simulation model to compare and update the start (ts(i,j)) and end times (te(i,j)) of transfers (i,j). The aim is to define the earliest time at which each transfer can be executed. This logic permits to sequence the different transfers in a correct way, generating a feasible schedule for the robot and a nearoptimal solution for the whole system, considering a predefined sequence of jobs. By using this logic, the transfers related to a given job are sequentially inserted according to the order in which they will be processed at every different bath (j=1,2,3,…,M+1). Then, the transfers are compared successively with all the transfers that were previously inserted into the schedule (according to a predefined processing sequence). The application of strict storage policies such as ZW and LS in the baths and the NIS rule in the robot significantly complicates the solution of the problem. Enforcing a ZW policy in the chemical baths j implies that the start time of the transfer to the water bath j+1 must strictly satisfy equation (1).

ts( i , j 1)

2,4,6...M ; i 1...N

(1)

For that reason, the value of ts(i,j) allows directly determining the value of ts(i,j+1). Here tp(i,j) represents the processing time of job i in bath j while ʌ(j) denotes the transfer time for every job from bath j-1 to j.

387

change (iter) and the number of transfers loaded onto the system (transf) are updated and the next transfer from the list (w,j) with j=j+1 if j < M+1; or w=w+1 and j=1 if w < N, is taken for the comparison. The algorithm ends when there are no more transfers to be compared in the system (j=M+1 and w=N). The simplified logic proposed is summarized in Figure 4. Next, the Generation Process is explained more in detail as well as the procedures of Initialization, Comparison and Update of the Evaluation Process, all of them generated by our algorithm.

and for all j´) between the starting time (ts(w,j)) and the ending time (te(w,j)) of the inserted transfer (w,j) During this iterative evaluation process the transfer times are initialized (Initialized Process), then they are compared with all the other transfers times (Comparison Process) and finally, they are updated (Updating Process). This loop is repeated successively for a given transfer, until all the comparisons, with the previously inserted transfers, do not introduce new updates at the compared transfer times. So, the comparison and updating processes end. Then, the transfer is evaluated and loaded onto the system with its respectively times [ts(w,j), te(w,j)], the counter number of iteration without

Figure 4: Pseudocode of the Generation and Evaluation Algorithm for transfers

388

Generation process To apply the logic in the simulation model it was necessary to define each transfer as a particular new entity in the system, together with the entities associated to the jobs in the system. Consequently, a given job “i” will have associated a certain number of transfers and/or entities (i,j) corresponding to the quantity of baths into the system j=1…M+1. Therefore, to start the processing of a given job i, all its j transfers must be pre-loaded into the system. Going back to equations (1) and (2), we can notice that the treatment of the transfers must be done in successive pairs. In order to define the start and end time of the transfer, it is necessary to correctly arrange the successive transfers in the robot, without overlap with any other transfer in the system. So, infeasible schedules are avoided. For that, it is necessary to define a set of attributes [ts(i,j),te(i,j)] [ts(i,j+1),te(i,j+1)], for each transfer (i,j) in order to define a correct sequence of transfer over time, avoiding infeasible solutions for the future transfer at the same job i (i, j+1).

Instead, for w = 1, the initial values of the attributes are defined following equation (6) and (7).

ts( w, j )

ªte( w 1, j 1) º « » Max «te( w 1, j  2 )  S ( j )  tp( w, j ) » «te » ¬ ( w, j 1)  tp( w, j 1) ¼

2,4,6...M ; w 2...N

(4)

(5)

(7)

If

te

Then

[ts ( w, j )

( w, j )

Else_If

ts

Then

[ts ( w, j )

( w, j )

d ts( w´, j´) › ts( w, j ) t te( w´, j´)

ts ( w, j ) ] š [te( w, j )

te( w, j ) ]

te( w´, j´) ] š [te( w, j )

ts ( w, j )  S ( j ) ]

 te( w´, j´) š te( w, j ) ! ts( w´, j´)

Update Procedure: This procedure is used to generate the earliest time at which the analyzed transfer (w,j) can be executed, in relation with the transfers previously inserted (w´, j´) and taking into account the resource constrains. As result, the efficient assignment and the detailed program of the robot is determined. The procedure tries to recalculate the value of the attributes [ts(w,j),te(w,j)] and [ts(w,j+1),te(w,j+1)] from the (w,j) transfer fulfilling the equations (1) and (2). As result of the comparison process, the attributes [ts(w,j),ts(w,j+1)] will be updated according to equation (9).

ts( w, j )  S ( j ) j 1,2,3...M  1; w 1...N

2,3...M  1; w 1

As can be seen, the updating process consists in delaying the start time of transfer (w,j) when overlapping with (w´,j´) are observed. Initially, it is necessary to compare the attributes [ts(w,j),te(w,j)] vs. and then [ts(w,j+1),te(w,j+1)] vs. [ts(w´,j´),te(w´,j´)] [ts(w´,j´),te(w´,j´)]. Thus, we try to ensure that if ts(w,j) • te(w´,j´), then by equation (1) and (2) ts(w,j+1) • te(w´,j´), else if te(w,j+1) ” ts(w´,j´) then te(w,j) ” ts(w´,j´).

There, te(w,j) is calculated for all baths j with the equation (6).

te( w, j )

( j ´)

( w, j ) z ( w´, j´);w 2...N ; w´d w; j, j´ (8)

ªts( w, j 1)  S ( j 1)  tp( w, j 1) º Max « » ¬«te( w 1, j 1) ¼» j

j´ 1

¦S

Comparison Procedure: Once transfers are initialized to w=1, they are loaded in the system by updating the subset of charged transfers V. In Vthere are all the transfers (w,j) that have been previously compared and assigned to the robot in a correct way. The value Vk represents the k-th transfer analysed and initialized into the system according to the priorities described above. The comparison procedure is applied to the pw position with w > 1. During this iterative procedure, the inserted transfer (w, j) is compared in pairs with a transfer (w´,j´) of the subset ı, already assigned to the robot (being the pw´ position < pw, that means w´< w). If analyzing the attributes (w,j) ([ts(w,j),te(w,j)] and [ts(w,j+1),te(w,j+1)]) of the transfer with the ones already inserted (w´,j´) ([ts(w´,j´),te(w´,j´)]) there is any overlap between the values of them, then the algorithm will update them for avoiding overlaps (see Equation (8)). Otherwise, the attributes will not be updated. That means that transfer (w,j) does not overlap with (w´,j´ ).

While for water baths (baths with even number), equation (5) is applied:

ts( w, j )

j´ 1

For the first transfer in the system (w=1 and j=1), the initial value is equal to zero (ts(1,1) = 0).

Initialized Procedure: the initialization procedure consists on determining the lower value at which the transfer can be initialized, assuming that there are not limitations of resources. So, we can determine the initial value ts(w,j) for each transfer using the following equations (4)-(5). For chemical baths (baths with odd number), equation (4) is applied:

j 1,3,5...M  1; w 2...N

j ´ j 1

¦ tp( w, j´)  j

Evaluation process After defining all the attributes of the inserted transfer, we proceed to determine an initial value.

ts( w, j )

j ´ j 1

(6)

So, for any job w > 1 the initial state of the attributes in the system is determined: [ts(w,j),te(w,j)]; [ts(w,j+1),te(w,j+1)].

389

If

ts

Then

ts( w, j 1)

the ı sequence without updating attributes, that is that the algorithm iteration number (iter) is greater or equal than the ı set cardinality (iter • transf), then the last transfer is loaded into the system with its respectively times, and the number of elements of ı set are updated (transf = transf + 1). The iteration counter is initialized (iter = 0) and the robot is assigned to the (w,j) transfer during the time between the interval [ts(w,j),te(w,j)]. The next transfer will be (j = j+1) if j100%, as seen in figure 8, the increase in Rel.OEE is negligible. Therefore this experimental formula is meaningful only for value of 0,05%≤ ≤100%. If ≤0,05% buffer can be considered null, therefore condition of complete dependence occurs, and than OEE = OEE(station 1) * OEE (Station 2)

value 0,5% 1% 2% 3% 4% 5% ≥ 6%

90,0% 85,0% 80,0%

Simulation configurations

Figure 9: R-square index between Analytic value and simulation results The formula (5) or (6) allows the buffer size without need for any further simulation. Two different application of the formula (6) to real industrial cases are summarized in table 4.. Table 4: Formula parameters of two example cases

is a experimental coefficient that allow to take into account that the behavior of the curve is not a real hyperbole. The depends by the value of j and value of they are reported in Table 4. Table 4:

95,0%

Value

Cases

MTTR A

MTTR B

T a (Sec)

PA

PB

Case 1 Case 2

20

30

7

0,9

0,8

18

23

15

0,85

0,8

Starting with case 1 we want to show how the Rel.OEE change depending on buffer size, and what are the value assumed by the equation parameter (table 5).

value 0,005 0,01 0,019 0,028 0,037 0,039 0,05

When the buffer size is zero the two station are completely dependent. Therefore the value of the OEE will be the equal to the product of the two performance index (Pa*Pb=0,8*0,9=0,72). With a buffer of 3 units Mj is 1,33%, therefore once chosed the right Kj value (0,01) the formula output is Rel.OEE=92,39%, corresponding to a OEE of 73,9%. Hence, a buffer of three unit increase the OEE of the system of 1,9%. For a buffer of 12 units, corresponding to an Mj of 5,19%, once selected the right Kj (0,039) the Rel.OOE is 92,87%, corresponding to an OEE of 74,3%. It means a further increase in OEE of 0,4%. Further increase in buffer size generates increase in OEE.

Equations (5) and (6) are the same, but their representation want to evidence two different aspects: Equation (5) represents the inverse relation is 100% the buffer between OEE and . When is maximum, the second term is null, and the first is null because REl.OEE = 100%, therefore maximum buffer cause maximum REL.OEE.

Table 5: Result description of Formula (6) for Case 1 Buffer size

Mj

Kj

Equation (6) is easier to use in order to calculate the Rel.OEE(j) for a chosen size of buffer.

0

0%

3

1,30%

For each analyzed configuration the analytic results of the analytic relation (equation 5 and 6) have been compared with simulation result.

12

The obtained R square index are vey high, and vary, from case to case within [0,945;0,993]. The relative ANOVA output on the significance of regression test

430

Rel OEE

-

OEE

90,00%

72,0%

0,01

92,39%

73,9%

5,19%

0,039

92,87%

74,3%

22

9,51%

0,05

95,24%

76,2%

32

13,83%

0,05

96,88%

77,5%

42

18,15%

0,005

99,77%

79,8%

The same consideration can be done for the case two, and results are briefly shown in Table 6.

REFERENCES Chiadamrong, N. &. (2003). Using storage buffer to improve unbalanced asynchronous production flow line’s performance. International Journal of Manufacturing Technology and Management , 149-161.

Table 6: Result description of Formula (6) for Case 2 Buffer 0

5.

Mj

Kj

-

-

Rel OEE

Oee

85,00%

68,0%

1

1,28%

0,01

88,42%

70,7%

4

5,12%

0,039

89,15%

71,3%

6

7,67%

0,05

90,98%

72,8%

15

19,18%

0,05

96,84%

77,5%

25

31,97%

0,005

99,84%

79,9%

D. Battini a, A. P. (2009). Buffer size design linked to reliability performance: A simulative study. Computers & Industrial Engineering 56 (2009) , 1633–1641. Gershwin, S. G. (1995). Efficient algorithms for transfer line design. MIT Laboratory for Manufacturing and Productivity Report LMP-95-005.

CONCLUSION AND FURTHER RESEARCH

The goal of this study have been to deliver an operative tool that allows an effective, but easy sizing of the buffer in flow shop industries also considering all the necessary information regarding the OEE trend. The wide range of values that have been simulated allows to include in the study a significant amount of different production systems and to evidence some analogies in the behavior between them. Pharmaceutical sector, where authors have already applied this study, is also included. the simulation range. Big effort in simulation analysis in conjunction with deep knowledge of the physical problem of the buffer design allow the introduction of a analytic relation. The added value of the analytic relation is the possibility to assess immediately, without the need for further simulations, and with strong statistical significance the optimal buffer size for a chosen level of availability. This relation is valid within the wide range of simulated value, and that proximally will be even wider. In fact studies to obtained a further more general analytic relation has already begun. Many other area of research are possible, such as a deeper analysis on the effect of time variability on buffer size, or the introduction of a analytic method that allows, once define the required OEE; to obtain an estimation of the required buffer size en a easier way, and without recourse to the recursive computation. Further research could be also carried out by changing the statistical distribution and the method of analysis

Gutowski, T. (2005). http://web.mit.edu.

Inventory

buffer

size.

H.T. Papadopoulos, C. H. (1996). Queueing theory in manufacturing systems analysis and design: A classification of models for production and transfer lines. European Journal of Operational Research , 1-27. Hillier, F. S. (1977). On the optimal allocation of work in symmetrically unbalanced production line systems with variable operation times. Management Science, 25(8), , 721-728. Hillier, F. S. (1993). Some data for applying the bowl phenomenon to large production line systems. International Journal of Production Research , 811–822. Kingman, J. F. (1966). On the algebra of queues . London: Methuen. Lawless, J. (1982). Statistical models and methods for lifetime data. John Wiley & Sons. Lutz, C. M. (1998). Determining buffer location and size in production line using tabu search. European Journal of Operational Research , 301– 316.

Keywords: Availability, Buffer, Buffer Design for Availability, Flow Shop, Simulation.

Malakooti, B. B. (1994). Assembly line balancing with buffers by multiple criteria optimization. . International Journal of Production Research, 32(9) , 2159–2178. Papadopoulos, H. T. (2001). heuristic algorithm for the buffer allocation in unreliable unbalanced

431

production line. Computers Engineering , 261-277.

and

Operations management, Energy Management, Project Management. Vito Introna has published more than 40 articles on several leading international and national journals, and presented research results in several international conferences.

Industrial

Samuel H. Huangt, J. P. (2003). Manufacturing productivity improvement using effectiveness metric and simulation analysis. International Journal Production Research , 513-527. Slack, N. (1993). The Flexibility of Manufacturing Systems. International Journal of Operations & Production Management , Vol. 7 Iss: 4, pp.35 - 45. Spinellis, D. D. (1999). Production line buffer allocation: Genetic algorithms versus simulated annealing. Second international Aegan conference on the analysis and modelling of manufacturing systems, (pp. 12-101). Tempelmeier, H. (2003). Pratical considerations in the optimization of flow production systems. International Journal of Production Research , 149–170. Yamada, T. &. (2003). A management design approach to assembly line systems. International Journal of Production Economics, 84 , 193–204.

AUTHORS BIOGRAPHY Vittorio Cesarotti is Tenured Associate Professor at the University of Rome “Tor Vergata”, holding the Chairs of Operations Management and Quality Management His Areas of Expertise are Strategy and Organization, Operations Management, Production Planning and Control, Business Excellence, Business Performance Measurement and Improvement, Quality Management and Control, Facility Management, Project Management, Supply Chain Management and Service Science. Prof. Cesarotti has published one book, seven international case studies, and more than 60 articles on several leading international and national journals, and presented research results in several international conferences. Alessio Giuiusa is PhD Candidate in Managerial Engineering at the University of Rome “Tor Vergata”. His Areas of Expertise are Operations Management, Quality Management and Control, Project Management and Service Science. Vito Introna is Assistant Professor in Industrial plant at University of Rome Tor Vergata. His Areas of Expertise are Maintenance, Statistical process control, Quality Management and Control,

432

USING QUERY EXTENSION AND USER FEEDBACK TO IMPROVE PUBMED SEARCH Viktoria Dorfer (a), Sophie A. Blank (b), Stephan M. Winkler (c), Thomas Kern (d), Gerald Petz (e), and Patrizia Faschang (f) (a,b,c,d)

University of Applied Sciences Upper Austria School of Informatics, Communications and Media, Bioinformatics Research Group Softwarepark 11, 4232 Hagenberg, Austria (e,f)

University of Applied Sciences Upper Austria School of Management, Digital Economy Research Group Wehrgrabengasse 1-3, 4400 Steyr, Austria (a)

[email protected], (b) [email protected], (c) [email protected], (d) [email protected], (e) [email protected], (f) [email protected]

view of an exemplary keyword clustering solution consisting of three clusters. In this paper we present a new interface for PubMed search that uses various keyword clusters to extend user queries. These keyword clusters have been generated using various evolutionary algorithms as described in Dorfer, Winkler, Kern, Petz, and Faschang (2010) and Dorfer, Winkler, Kern, Blank, Petz, and Faschang (2011). User ratings of the query results (returned by PubMed for queries that were extended using keyword clusters) will then be used to improve the generation of keyword clusters and query extensions.

ABSTRACT PubMed is a search engine that is widely used to search for medical publications. A common challenge in information retrieval, and thus also when using PubMed, is that broad search queries often result in lists of thousands of papers that are presented to the user, too narrow ones often yield small or even empty lists. To address this problem we here present a new PubMed search interface with query extension using keyword clusters generated with evolutionary algorithms to obtain more specific search results. Users can choose to add various words to their query and then rate search results; this scoring is stored in a database to enable learning from user feedback to improve keyword cluster optimization as well as query extensions. We show how users can extend PubMed queries using previously generated keyword clusters, rate query results, and use these ratings for optimizing parameters of the keyword clustering algorithms. Keywords: Bioinformatical Information Retrieval, PubMed, Query Extension, Keyword Clustering 1. INTRODUCTION Search queries in PubMed (PubMed 2011) often return a huge number of papers, many of them not relevant for the user. To address this problem of irrelevance, information retrieval provides a suitable solution, namely query extension. User queries are extended with matching terms to result in fewer, but relevant search results. This method is based on clusters of keywords and / or documents; from these clusters words are selected that belong to the same clusters as the search terms defined by the user, and thus queries can be extended automatically. Figure 1 shows a schematic

Figure 1: Example of a keyword clustering solution 2. A NEW PUBMED SEARCH INTERFACE We have developed a PubMed search interface that uses previously generated keyword clusters to suggest words that fit the terms the user entered as query. Several clustering files can be selected to search for suitable extensions.

433

Figure 2: PubMed query extension using previously generated keyword clusters In Figure 2 a screenshot of the search and extension interface is provided. On the left hand side the keyword clusters containing terms or parts of terms of the user query are shown; these keywords can be selected for query extension. On the right side the chosen terms can be added to the query using ADD or OR conjunctions. More than one search query can be submitted in parallel to be able to compare results with different selections of keywords extending the original query.

D – the document confidence – the amount of keywords in the documents also present in the cluster is represented. Parameter E regards the distribution of the documents in the clusters and G considers the number of generated clusters with respect to the data set size. Į, ȕ, Ȗ, į, İ, and ȗ are weighting factors that are necessary to be able to emphasize or neglect specific features, depending on the user’s needs.

3. KEYWORD CLUSTER GENERATION As mentioned before, these keyword clusters have been generated using evolutionary algorithms, driven by mutation, crossover and selection (see Dorfer, Winkler, Kern, Petz, and Faschang (2010) for further details on the operators implemented for this problem class) using a fitness function that is used for evaluating solution candidates. This fitness function takes various features into account we consider important for keyword clustering solution candidates and is defined by the following equation (Dorfer, Winkler, Kern, Petz, and Faschang (2010)):

Figure 3: Genetic algorithm with offspring selection optimizing keyword clusters

‫ ܨ‬ൌ ߙ ή ‫ ܣ‬൅ ߚ ή ‫ ܤ‬൅ ߛ ή ‫ ܥ‬൅ ߜ ή ‫ ܦ‬൅ ߝ ή ‫ ܧ‬൅ ߞ ή ‫ܩ‬. (1)

In Dorfer, Winkler, Kern, Blank, Petz, and Faschang (2011) we have presented a detailed comparison of the performance of various evolutionary algorithms solving this keyword clustering problem (KCP), including the evolution strategy (ES) (Schwefel 1994), the genetic algorithm (GA) (Holland 1975), the

Parameter A represents the number of clusters a document is assigned to; parameter B the data coverage; C – the cluster confidence – quantifies to which amount the keywords of the cluster are also present in the documents assigned to the cluster, whereas in parameter

434

genetic algorithm with offspring selection (OSGA) (Affenzeller, Winkler, Wagner, and Beham 2009), and the elitist non-dominated sorting genetic algorithm (NSGA-II) (Deb, Pratap, Agarwal, and Meyarivan 2002). In Figure 3 we show the most important parts of an evolutionary algorithm solving the KCP (in particular, an OSGA is depicted). The evolutionary process in ESs is mainly based on mutation and selection, whereas evolution in GAs is directed by the interaction of mutation, crossover and selection; the OSGA is an extension of the standard GA and includes an additional offspring selection step. The NSGA-II is a multi-objective approach, which optimizes various objectives in parallel, as we have identified parameters B, C, and D as potentially contrary. We have shown that intense parameter settings and weighting factor tuning is necessary to receive good solutions using single-objective algorithms and that ESs and the NSGA-II perform best in terms of achievable solution quality (Dorfer, Winkler, Kern, Blank, Petz, and Faschang (2011)). The data set used to generate the mentioned keyword clusters has been taken from the 9th Text Retrieval Conference (TREC-9) from the year 2000 (Vorhees and Harman 2000), consisting of 36,890 samples based on PubMed entries including title, abstract, authors, its medical subject headings (MeSH), publication type and source.

4. USER FEEDBACK After the search has been performed the user can rate the results (see Figure 4); this rating in combination with the words used for query extension and the algorithm parameters the keyword clusters have been created with (algorithm type, algorithm parameters, documents used for clustering, …) are stored in a database in order to be able to examine the settings and algorithms which delivered the best results. Another interesting aspect in the context of biomedical information retrieval is whether there are specific keywords that, when included in the query, never lead to satisfying solutions. By analyzing user feedback these words can be identified and collected in a blacklist to warn the user as soon as such a term is included in the query. User feedback cannot only be used for the generation of a blacklist, but also to improve the parameter settings of the keyword clustering algorithms. This shall enable an iterative improvement of parameter settings leading to better and more suitable clusters which can then be used in query extension.

Figure 4: Example of search results with rating possibility

Figure 5: Analysis of algorithm parameters that were applied for forming keyword clusters using various evolutionary algorithms. The keyword clusters have been used for query extension and the search results have been rated; algorithm settings are ranked according to the user’s feedback

435

In Figure 5 we show an example of the examination of rating results and algorithm specific parameter settings to find out whether these settings yield appropriate solutions. All fields of the database can be selected for filtering, including general algorithm parameters (such as algorithm type, population size, and also fitness function specific parameters as – for example – B and D) and of course the user ratings. The previously explained keyword clusters and the evolutionary algorithms used to generate these clusters provide the basis for this analysis. Based on the results of these examinations we want to further improve the algorithm parameters to optimize the generation of keyword clusters; this shall extend our optimization algorithm with a user-driven feedback component.

After users have performed various queries and have rated the obtained results accordingly, we can now have a further look on the performance of the various evolutionary algorithms and their corresponding settings in terms of parameters and also weighting factors. In Figure 6 we can see that here keyword clusters generated by evolution strategy with specific weighting factor settings and also generated by the NSGA-II provided clusters the users ranked best. However, this does not have to be the case in all examples; one can assume that specific user needs and specific combinations of terms are more likely to be produced by a genetic algorithm, for example. This and other questions can be analyzed by applying various filters on the feedback data, beside the possibility to gain a quick insight in which algorithm performs best regardless of the specific user needs.

5. APPLICATION EXAMPLES Using the best keyword clustering files identified in Dorfer, Winkler, Kern, Blank, Petz, and Faschang (2011), generated by ES, GA, OSGA and the NSGA-II, a detailed analysis on the differences of the generated clusters and on the performance of the different algorithms with the different parameter settings with respect to user feedback can now be performed. In Figure 6 we provide an application example, in particular, we search for “urethra”. We can clearly see the differences in the clusters generated by ES and NSGA-II. As evolution in evolution strategies solving the KCP is dependent on the fitness function given in Equation 1, quite different keywords are grouped together in contrast to the clusters generated by the NSGA-II, which in this example optimizes only parameters B, C, and D. On the right bottom of the window detailed information on the used algorithm and its parameter settings to generate the cluster is given. All shown keywords can be chosen for query extension, optionally in combination; this selection does not depend on the algorithm the clusters have been generated by.

Figure 7: Application example on the analysis of user feedback As this new search tool shall be used to improve PubMed search results we give another application example here: Assuming – for the sake of simplicity – one wants get some information about the effect of smoking on the lung, a search of these two terms (“smoke AND lung”) would yield 7310 hits. In Figure 8 the words that often occur in combination with at least one of these two terms are depicted. To be sure not to have missed any paper on this topic, the query can be extended with an OR conjunction on the term “pulmonary” to retrieve also all the papers where pulmonary has been consequently used as a synonym for lung, leading to a publication list of 8384 entries.

Figure 6: Application example using keyword clusters generated by two different evolutionary algorithms

436

Alternatively, a reduction of the result space can be obtained by adding more specific terms inspired by the suggestion of the keyword clusters. Adding for example the keyword “asthma”, PubMed returns only 699 hits, adding the keyword “haemoptysis” leads to only 6 results.

generation of so-called white lists; these are lists of terms that should always be used in combination to obtain meaningful results. Users can then be warned or encouraged to include specific terms or keyword clusters of specific algorithms in their search. ACKNOWLEDGMENTS The work presented in this paper was done within the TSCHECHOW project, sponsored by the basic funding program of the University of Applied Sciences Upper Austria.

REFERENCES Affenzeller, M., Winkler, S., Wagner, S., Beham, A., 2009. Genetic Algorithms and Genetic Programming - Modern Concepts and Practical Applications. Chapman & Hall/CRC. ISBN 9781584886297. 2009. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation 6:182-197. Dorfer, V, Winkler, S.M., Kern, T., Blank, S.A., Petz, G., and Faschang, P., 2011. On the Performance of Evolutionary Algorithms in Biomedical Keyword Clustering, Proceedings of the Genetic and Evolutionary Computation Conference, 10, July 12-16, 2011, Dublin, Ireland. Dorfer, V, Winkler, S.M., Kern, T., Petz, G., and Faschang, P., 2010. Optimization of Keyword Grouping in Biomedical Information Retrieval Using Evolutionary Algorithms, Proceedings of the 22nd European Modeling and Simulation Symposium, 25-30, October 13-15, 2010, Fes, Morocco. Holland, J.H., 1975. Adaption in Natural and Artifical Systems. University of Michigan Press PubMed, 2011, Available from: http://www.ncbi.nlm.nih.gov/pubmed Schwefel, H.-P., 1994. Numerische Optimierung von Computer-Modellen mittels der Evolutionsstrategie. Basel: Birkhäuser Verlag Vorhees, E.M., Harman, D.K., 2000. NIST Special Publication 500-249: The Ninth Text REtrieval Conference (TREC-9). Gaithersburg, Maryland: Department of Commerce, National Institute of Standard and Technology.

Figure 8: Keyword clusters containing the terms “lung” or “smoke”

6. CONCLUSION In this paper we have presented a new query interface that used keyword clusters to improve PubMed queries. The used clusters have been generated using various evolutionary algorithms and different parameter settings. In addition to the extension functionality, obtained results can be rated and various analyses can be performed on this kind of user feedback. We have shown that the various algorithms generate different keyword clusters and that it is therefore reasonable to use several keyword clustering files to choose the words for extension. The provided example described the applicability for extending and narrowing the search space by the use of either synonyms or specifications, depending on the users’ needs in the context of PubMed search. 7. FUTURE WORK We are currently working on an online version for the proposed tool for improved PubMed querying. This version will enable us to address a broader user clientele and then to do further analyses on their feedback. Having now identified the best performing algorithms we plan to slightly adapt parameters and weighting factors and to perform further keyword clustering to hopefully produce even better keyword clusters. This will be an ongoing process as new keyword clusters will lead to new ratings and so on. As mentioned before we also plan to design a black list of terms or of combinations of terms and algorithms that are less likely to lead to useful search results. We are also working on the automated

437

AUTHORS BIOGRAPHIES VIKTORIA DORFER is a senior researcher in the field of Bioinformatics at the Research Center Hagenberg, School of Informatics, Communications and Media. After finishing the diploma degree of bioinformatics in 2007 she was a team member of various projects in the field of bioinformatics and software engineering. She is currently working on biomedical information retrieval within the TSCHECHOW project.

THOMAS KERN is head of the Research Center Hagenberg, School of Informatics, Communications and Media, Upper Austria University of Applied Sciences (UAS). He finished his studies in Software Engineering in 1998. After some work experience in industry he started his academic career at UAS in autumn 2000 as research associate and lecturer for algorithms and data structures, technologies for knowledge based systems, semantic systems and information retrieval. Since 2003 he has conducted several application oriented R&D projects in the fields of bioinformatics and software engineering.

SOPHIE A. BLANK received her bachelor's degree in bioinformatics from the Upper Austria University of Applied Sciences in 2011; she is currently a junior researcher at the Research Center Hagenberg, School of Informatics, Communications and Media working on biomedical information retrieval within the TSCHECHOW project.

GERALD PETZ is director of the degree course Marketing and Electronic Business at the University of Applied Sciences in Upper Austria, School of Management in Steyr. His main research areas are Web 2.0, electronic business and electronic marketing; he has also conducted several R&D projects in these research fields. Before starting his academic career he was project manager and CEO of an internet company.

STEPHAN M. WINKLER received his MSc in computer science in 2004 and his PhD in engineering sciences in 2008, both from Johannes Kepler University (JKU) Linz, Austria. His research interests include genetic programming, bioinformatics, nonlinear model identification, and machine learning. Since 2009, Dr. Winkler is professor at the Department for Medical and Bioinformatics at the Upper Austria University of Applied Sciences, Campus Hagenberg, and since 2011 he is head of the Bioinformatics Research Group.

PATRIZIA FASCHANG received her bachelor’s degree in electronic business and is a junior researcher at the Research Center Steyr, School of Management, in the area of digital economy. She has worked on several small projects in the field of marketing and electronic business and is currently working on Opinion Mining and Web 2.0 methods within the TSCHECHOW project.

438

SIMULATION OF THE VESSEL TRAFFIC SCHEDULE IN THE STRAIT OF ISTANBUL Şirin Özlem(a), İlhan Or(a), Birnur Özbaş(b) (a)

(b) (a)

Boğaziçi University, Istanbul Rutgers, The State University of New Jersey

[email protected], [email protected], (b) [email protected]

national conventions and regulations, while improving safe navigation, protecting life and environment. Within the framework of this system, vessels desiring to transit the Strait have to submit two reports to the VTS, Sailing Plan 1 (SP-1) and Sailing Plan 2 (SP-2). SP-1 includes all the information about the vessel and must be submitted at least 24 hours before the arrival. SP-2 is of vital importance for planning of vessel passages from the Strait and must be submitted at least 2 hours or 20 nautical miles (whichever comes first) prior to entry into the Strait. The VTS analyze the data in these reports and prepare a safe daily sailing traffic plan (VTS Users’ Guide).

ABSTRACT In this study, a simulation model is developed via the Arena 11.0 software to mimic the actual Istanbul Strait vessel flow under the established traffic regulations and meteorological conditions. The established practice of uni-directional daytime and two-directional nighttime traffic schedules are reflected and pilot and tugboats services scheduled in the traffic flow direction, visibility, current and storm information are also integrated into the model. The effects of factors such as pursuit distance, vessel profile, pilot availability, arrival rate and visibility over selected performance measures are investigated through scenario analysis and the most important factors are determined as arrival rate of vessels and visibility.

2. SIMULATION MODEL The first step to better understand the risks generated by the maritime traffic in the Strait is to understand and model the maritime actively in the Strait. This study aims to design and develop a simulation model to represent the actual traffic flow in the Strait with regard to the VTS rules and regulations (R&R) and policies that meteorological and geographical conditions, support services (like pilot and tugboats) and frequency, type and cargo characteristics of vessel arrivals (to make a passage through the Strait) with the aim of identifying the impact of such factors on traffic conditions, potential problems and bottlenecks for a less risky transit and overtaking allowance during the passage of vessels on Strait lanes.

Keywords: Strait of Istanbul, Maritime traffic, Simulation 1. INTRODUCTION The Istanbul Strait, 31 kilometers in length is one of the narrowest waterways in the world with only 660 meters at its narrowest point (Almaz 2006). Vessels navigating through the Strait have to make many sharp turns (between 45 and even 80 degrees) which carry high risks for the vessels in such a narrow channel (Ulusçu et Al. 2009). The Strait which is situated in the middle of a huge metropolitan area of 15 million residents, features a very heavy maritime traffic (more than 51,000 vessels annually), with more than 15,000 such vessels carrying dangerous cargo; there is also heavy local traffic including more than 2,000 passenger ferry trips daily between the two shores (Gönültaş 2007). One noteworthy property of the Strait is the prevailing currents which may rise up to 8 knots speed. Other adverse meteorological conditions like fog, wind, rain and storm also increase the difficulty of navigation in the Strait. In dense fog conditions, vessel traffic may be partially or wholly suspended until meteorological conditions improve, which causes dangerous and unwanted pile-ups at the Strait entrances and puts further strains on the maritime traffic management, since it increases navigation problems (Özbaş 2005). The Vessel Traffic System (VTS) was established in 2004 in order to regulate and guide maritime traffic in the Strait, in accordance with international and

2.1. Vessel Classification The VTS has a specific vessel classification system based on vessel types, cargo characteristics and vessel lengths. In this study a somewhat simplified version of this classification (which is displayed in Figure 1) is used. The main reason why tankers and dangerous cargo vessels up to 100 meters and LPG-LNG up to 150 meters, tankers and dangerous cargo vessels between 100 and 150 meters and dry cargo carrying vessels between 150 and 300 meters are placed in the same class is that according to the VTS regulations, they have to satisfy the same conditions in entering and navigating the Strait. This way of classification simplifies the understanding of vessel entrance and sailing conditions.

439

2.3. The Istanbul Strait Traffic Rules and Regulations Vessels enter the Strait either from the north, (traveling south and thus are called as southbound vessels) or from the south (traveling north and thus are called northbound vessels) entrances. Some R&R related to vessel transit management that are also reflected in the simulation model are as follows: There should be at least a 10-minute interval between two consecutive ready to enter vessels from one direction. Class A and T6 vessels pass through the Strait only during daytime. No vessels are allowed to meet with Class A vessels. Class B, C and E vessels should not meet each other during bi-directional nighttime flow. There should be at least 75 minutes between two consecutive southbound Class A vessels and at least 90 minutes between two consecutive northbound Class A vessels. Passenger vessels are allowed to the Strait regardless of their direction of flow when pursuit distance, meteorological and pilot and tugboat request conditions are satisfied. Southbound stopover vessels have priority over northbound stopover vessels, which have priority over any non-stopover vessels.

Figure 1: Vessel Classification 2.2. The Arrival Process The Arena Input Analyzer which is a very efficient tool for distribution fitting to data is deployed in fitting interarrival time distributions. Via the Input Analyzer’s Fit menu, all probable distributions fitted to the actual data are revealed and “fit all” property estimates the distribution with the minimum square error. After fitting a distribution, a histogram and the probability density function (pdf) superimposed on the histogram summarize the characteristics of the fit (Law and Kelton 2007). To illustrate, the best fitted interarrival time distribution of northbound Class E vessels is found as the Gamma distribution with shape parameter α being 648 and scale parameter β being 0.974. In the summary report of Arena Input Analyzer (as displayed in Figure 2), the shape of the probability density function overlaps with the histogram and just looking at this figure, one gets the feeling that the selected function represents the actual interarrival time data quite well.

2.4. Vessel Sequencing Observations of the 2009 transit data and discussions with the VTS authorities have indicated that the implementation of the regulations regarding pursuit distances between two consecutive vessels of various classes can be parameterized into a set of easily followed rules. Let θ be the minimum pursuit distance between two consecutive vessels of class D, E, P traveling northbound and let μ be the minimum pursuit distance between two consecutive vessels of class D, E, P traveling southbound. According to the R&R, the minimum pursuit distance between a northbound (southbound) class D, E or P vessel and a class A, B or C vessel sailing in the same direction is also θ (μ). The minimum pursuit distance between two consecutive class C vessels traveling northbound (southbound) is 2*θ (2*μ) and the minimum pursuit distance between a northbound (southbound) class C and a class A or B vessel sailing in the same direction is also 2*θ (2*μ). The minimum pursuit distance between two consecutive A and B vessels traveling northbound (southbound) is respectively 6*θ (6* μ) and 4*θ (4* μ). 2.5. Daytime Vessel Scheduling As mentioned before, traffic flows from one direction at a time during daytime. The maximum duration of daytime and start time of the daytime traffic differ according to seasons. The first direction of vessel flow

Figure 1: Histogram of northbound Class E interarrivals

440

into the Strait at daytime is determined based on the total number of vessels in queues and their waiting time regarding vessel priorities (two hours before the starting time). The formula used for in the determination of starting direction is as follows:

Sd

a*

b*

Opposite direction traffic time window length is calculated as: NQ ( A)tdS' (3) Wd ' N NQ ( A)tdS N NQ ( A)tdS'

Ca * NQ ( A) ( d ) Cc C * NQ (C ) ( d ) Cd C * NQ ( D) ( d ) Ce C * NQ ( E ) ( d ) (d d ') (d d ') (d d ') NQ ( A) NQ NQ NQ N (C ) N ( D) N (E)(d d ')

The number of Class A vessels planned to enter the Strait during the starting direction vessel traffic flow is: Wd (4) N (d )

Ca *WT ( A) ( d ) C Cc *WT (C ) ( d ) Cd C *WT ( D) ( d ) Ce C *WT ( E ) ( d ) WT ( A) ( d

d ')

W WT (C ) ( d

d')

WT W ( D) ( d

d ')

WT W (E)(d

d ')

p

(1)

6*

where:

where: Sd: score value of the active direction d Sd’: score value in the opposite (passive) direction d’ α: multiplicative constant for number of vessels in queues b : multiplicative constant for waiting time of vessels in queues Ca: coefficient for A type vessels Cc: coefficient for C type vessels Cd: coefficient for D type vessels Ce: coefficient for E type vessels (d ) : number of i type vessels in queue in active NQ (i ) t

if d is northbound

(d )

The number of Class A vessels planned to enter the Strait during the opposite direction vessel traffic flow is: (6) Wd ' N p (d ' )

6 * (d ' )

Both N p (d ) and N p (d ' ) are rounded down to nearest integer numbers.

direction d at time t=ts ( d ' ) : number of i type vessels in queue in NQ (i )t s

Waiting time of vessels is adjusted depending on whether they are stopover vessels or not. The adjusted waiting time of vessel j is defined by:

passive direction d’ at time t=ts (d ) : total waiting time of j type vessels in active WT ( j )t s

direction d at time t=ts ( d ' ) : total waiting time of j type vessels in passive WT ( j )t

W ja

s

direction d’ at time t=ts

c *WTtsd ( j )

(7)

where:

This formula is applied for both directions and the direction with higher score is declared as the starting direction of the daytime traffic schedule. Two significant factors influencing the determination of the first direction of daytime flow are the number of vessels in queues and vessel waiting times and they are in different level of significance. (The associated weights α and are nominated as 0.25 and 0.75 respectively).

1.5 c

if j is a stopover southbound vessel

1.25 if j is a stopover northbound vessel 1

otherwise

(8)

Since passenger vessels have the highest priority in vessel sequencing, the model first searches the Class P queue in the determined direction. If there exist any P vessels in the determined direction and if the visibility conditions and pilot and tugboat demand are satisfied, the one having the maximum elapsed waiting time is allowed to the Strait and the time is incremented as θ (μ) minutes. Meanwhile, if there exist any P vessels on the other side, the one with the maximum elapsed waiting time is allowed to the Strait as well (even though a uni-directional time window is in action). If there is no P vessel in the determined direction, the model searches the Class A queue. If there is any A type vessel in the determined direction, then the pursuit distance requirements, meteorological situations and pilot and tugboat availabilities are checked. When all conditions are fulfilled, the class A vessel having the

Class A and T6 vessels are the most critical vessels in terms of the risks they generate. Therefore, in order to set out the framework for daytime schedule, (after attaining the first direction of daytime traffic), number of Class A vessels transiting from both directions are estimated. In this respect, maximum daytime duration is divided into two, proportion to the number of Class A vessels in northbound and southbound queues. Starting direction traffic time window length is calculated as: NQ ( A)td (2) W S

N NQ ( A)tdS

(5)

if d is southbound

The parameters in the denominator changes with regard to starting direction decision.

s

d

(d )

NQ N ( A)tdS'

441

maximum elapsed waiting time enters the Strait, otherwise model examines the Class C, E and D vessel queues respectively and allows the one having maximum elapsed waiting time regarding their minimum pursuit distances among class types. As soon as a vessel enters the Strait, again time is incremented as the minimum pursuit distance interval (as θ or μ minutes) and the other distance rules among vessel types are also checked until the last planned A vessel in the active direction enters the Strait.

' WT ( A) dSLACK

X

ts

Wd ' ))

direction d and d’ at time

NQt ( A)

M MAX(0, ( NQ( A) t

iii) The indicator

X*

Z

(12)

Z

is determined as follows:

Y *b

(13)

iv) The exact procedure of allocating the slack time to additional northbound and / or southbound class A vessels is as follows: a. If Z is greater than or equal to 1, it is deduced that the additional class A vessel (planned to pass in the slack time) should be a d-directional vessel and then the equations (11) and (12) are updated. Number of d-directional planned A vessels in slack time ( N (d ) SLACK ) is p incremented by one.

(8)

N p (d ' )))

calculated as:

ii) If the amount of slack time is larger than or equal to time length that allows a southbound A vessel transit (6* ), the slack time algorithm tries to make use of this time by scheduling one more northbound or southbound class A vessel.

(i) Number of Class A vessels in the opposite direction at time t= t is checked. One important detail at this point is ignoring the number of previously planned vessels in the opposite direction (Np (d’)), since they are already scheduled to pass in the original time window determined at plan time. Namely, the new arrivals (since plan time) of class A vessels in opposite direction are: d'

t is

WT ( A) ( d ) ' WT W ( A) dSLACK

Y

where t s is the start time of the first direction vessel traffic flow. The steps for slack time schedule at time t= t are as follows:

d' SLACK

(11)

i) The ratio for waiting time of unscheduled vessels in

The length of slack time is: (t (t

NQ( A) ( d ) ' NQ N ( A) dSLACK

Since t represents a time point at which all scheduled vessels in the active direction have already moved into the Strait, the numerator must only contain the new arrival class A vessels since plan time.

For slack time traffic plan, the number of Class A vessels planned to enter the Strait during the starting and opposite direction uni-directional traffic flow time windows is computed by dividing this apportioned times by the minimum pursuit distance between two consecutive Class A vessel transiting from starting and opposite directions time windows.

MAX M (0, DT

(10)

WT W ( A) tds '

(iii) The ratio for number of unscheduled class A vessels in both directions is estimated as:

Since the original daily schedule is made in the morning (two hours before traffic start time), the uni-directional time windows of that schedule are designated to service just the available vessels (especially A vessels) at that time. So, close to the end of the time window of the starting direction, say at time t= t , the model reviews the number of Class A vessels in queues and revises the original schedule to extend the uni-directional time windows as long as the maximum daytime duration permits. This extended time interval is named as the slack time.

ST

WT W ( A) ( d ')

N (d ) SLACK p

N (d ) SLACK p

(14)

1

and the slack time length is updated as: ST

b.

(9)

(ii) The additional waiting time of new arrival (since plan time) class A vessels in direction d’ at time t is computed. This can be done by removing the realized waiting time of planned A vessels from total waiting time of Class A in direction d’, that is:

S ST

6*

(15)

(d (d )

If Z is less than 1, it is deduced that the additional class A vessel (planned to pass in the slack time) should be a d’-directional vessel and then the equations (13) and (14) are updated. Number of d’-directional planned A vessels in slack time ( N (d ' ) ) is SLACK p

incremented by one and the slack time length is updated same as equation (15).

442

N (d ' ) SLACK p

N (d ' ) SLACK p

deciding windows length after sequencing class B

(16)

1

vessels) at nighttime plan (t= t n ) are updated in starting and opposite directions respectively as follows:

(viii) Returning to step (iii), the algorithm proceeds until the end of ST.

D

Daytime Start

NQtdn ' ( B)

MAX(0, ( NQtdn ' (C) ( NQtdn ' ( B) 1))) / 2)

Then, the tentative time window length in the nighttime active direction is calculated as follows: NW p (d )

NT N *

NQtdn ( Bup )

(18)

NQtdn ( Bup ) N NQtdn ' ( Bup )

The tentative time window length in the nighttime passive direction is calculated as follows: NQtd ' ( Bup ) (19) NW p (d ' )

N NT *

n

N NQtdn ' ( Bup )

NQtdn ( Bup )

where NT is the total nighttime duration, which is the time gap between the following day’s daytime traffic plan start time and the end of the present day’s daytime windows. Accordingly, the number of Class B vessels planned to enter the Strait in the active direction flow is: N pB ( d )

m min( NQtdn ( B ),

NW p ( d ) 4 * (d )

)

(20)

The number of Class B vessels planned to enter the Strait in the passive direction flow is: C

D

D

A

E

D

C

D

E

A N pB ( d ' )

Slack time flow Start

m min( NQtdn ' ( B ),

NW p ( d ' ) 4 * (d ' )

)

(21)

Presuming Class B the most critical group in the nighttime schedule, the length of the northbound and the southbound time windows are outlined by Class B vessels (similar to the role of Class A vessels in daytime scheduling). However, the relatively high population of the abundance of Class C vessels (around 9000 Class C vessels in a year) necessitates the consideration of this class while designing the nighttime traffic plan. Considering that minimum pursuit distance between two Class B vessels is 4*θ (4*μ) whereas minimum pursuit distance between a Class B vessel and a Class C vessel is 2*θ (2*μ), the duration of nighttime restricted traffic flow time is determined by the number of planned Class B vessels (multiplied by 2*θ or 2*μ, according to the active direction), and the number of remaining class C vessels (multiplied by θ or μ, according to the active direction).

Southbound Northbound A

d'

NQtn ( Bup )

Northbound Southbound D

NQtdn ( B) MAX(0, ( NQtdn (C) ( NQtdn ( B) 1))) / 2)

(17)

By means of this reschedule procedure, more vessels from both directions are scheduled and admitted to transit until the end of the slack time. At the end of the (extended) starting direction time window (i.e. with the entrance to the Strait of the last scheduled class A vessel from that direction), the traffic is closed from both directions until the last vessel leaves the Strait. Since it takes approximately 30 minutes for a class A at Filburnu (in northbound traffic flow case) or at Boğaziçi Bridge (in southbound traffic case) to completely exit the Strait, the time gap between the last northbound or southbound Class A vessel and the following vessel from the opposite direction should be 6* θ+ 30 or 6* μ+30 minutes, respectively. At the end of the starting direction time window (i.e. with the entrance to the Strait of the last scheduled class A vessel from that direction), the traffic is closed from both directions until the last vessel leaves the Strait. The start and execution of the vessel traffic flow in the opposite direction traffic is the same as the first direction flow. Vessels are allowed into the Strait until reaching the number of planned A vessels in this direction. If slack time admits any more A vessels in this direction, they also enter the Strait until the start of the nighttime vessel traffic. A typical example for daytime vessel schedule is displayed in Figure 3.

A

d

NQtn ( Bup )

P C D D D A D D C D E A Figure 3: Daytime Schedule

2.6. Nighttime Vessel Scheduling When daytime traffic ends, the active traffic flow direction remains as the first (active) direction of nighttime traffic. Additionally, unlike daytime unidirectional traffic, at nighttime, there exist two restricted vessel flows (according to the R&R, Class D vessels may enter from the opposite direction when there are such vessels available and meteorological conditions allow since no Class A vessels are allowed from either direction during nighttime). Number of Class B vessels and the number of all Class C vessels (the ones which will be used for

443

The total number of Class C vessels planned to enter the Strait after sequencing class B vessels in the active direction flow is: N p( C ) ( d )

max( m 0,

NQtdn (C )

( N pB ( d ) 2

1)

)

2.8. Pilot and Tugboat Services According to the R&R, having a pilot captain on board during the Strait passage is compulsory for vessels longer than 250 meters and optional (though strongly recommended) for other vessels. All vessels express their pilot captain and tugboat needs in their SP-1 and SP-2 reports. There are 20 pilots and 6 tugboats available(as in the real situation). In the simulation model pilots and tugboats are treated as resources which are seized by vessels at the embarking area in the Strait and released while leaving. In order to meet pilot and tugboat needs, every hour the model searches the number of available pilots (including transferring pilots) in the active direction and requests pilots from the opposite side when it is less than 6. The model also searches the number of available tugboats in the active direction and requests tugboats from the opposite side when it is less than 3. During the nighttime time windows, number of pilots at both sides is equalized to 6 and tugboats to 3 to meet the pilot and tugboat demand. Once a piloted vessel’s passage in a certain direction is completed, the pilot is released from its current duty and included in the set of available resources for the opposite direction.

(22)

The total number of Class C vessels planned to enter the Strait after sequencing class B vessels in the passive direction flow is:

N p( C ) ( d ' )

max( m 0,

NQtdn ' (C )

( N pB ( d ' ) 1) 2

)

(23)

Both equations (22) and (23) are rounded down to nearest integer numbers. The resulting total nighttime vessel traffic duration in the active direction is:

NW (d ) m min( NWp (d ), N pB (d ) * 4 * (d ) N Cp (d ) * 4 * (d ))

(24)

The resulting total nighttime southbound vessel traffic duration in the passive direction is:

2.9. Visibility Conditions According to the R&R, when visibility is less than one nautical mile in the Strait (called FogType1, only oneway traffic is permitted and when visibility in the Strait is less than 0.5 mile (called as FogType2), vessel traffic is suspended in both directions. The visibility module in the simulation model reads the fog information from the visibility data of (Almaz 2006) externally. Before a vessel is allowed to enter the Strait from the active direction during daytime, visibility condition is checked; if there is a FogType2 event, the vessel waits until it disappears. FogType1 does not affect daytime flows very much (since almost all vessel activity with the exception of class P vessels is uni-directional anyway); only the class P vessels coming from the opposite (passive) direction are stopped. When a FogType1 occurs at nighttime, however, two-way traffic is suspended.

NW (d ' ) min( m NWp (d ' ), N pB (d ' ) * 4 * (d ' ) N Cp (d ' ) * 4 * (d ' )) (25) Once the scheduled transit of Class B and C vessels is completed, if there is remaining nighttime, Class D and E vessels continue entering the Strait from both directions (with Class E still having higher priority) according to the minimum pursuit distances (θ or μ) rules. 2.7. The Traffic Lanes and Overtaking In the model, vessels follow two main lanes, (the northbound or the southbound lanes) and the overtaking lane, if permitted, while transiting the Strait. The whole Strait is divided into 22 slices with stations. Slices are at eight cables (0.8 nautical miles ≈ 1.482 km.) intervals and in order to sustain a predetermined pursuit distance between vessels each slice is also composed of 2 cables long substations. Since stopping in the Strait for any reason is not allowed, vessels continuously move from one station to another during their stay in the Strait. Overtaking is allowed in the Strait except at the narrowest part, according to these conditions:

2.10. Current Conditions The most dominant current type on the Strait is the southbound surface current caused by level difference between the Black Sea and the Mediterranean Sea. The current module of the simulation model is integrated into the model from the previous study (Almaz 2006). In the study, the most effective southbound current is taken into account and a moving average function is built to estimate a daily base current value. Then, the current level at different regions of the Strait are assigned as predetermined percentages of the base value, based on historical current data. In order to comply with the R&R, when current speed exceeds 4 knots, class A, B, C and E vessels having a speed less than 10 knots are not allowed into the Strait. Moreover, all vessels in these classes have to wait in their queues

When a vessel is in the overtaking lane, there should be no other vessel in this lane in the opposite direction at least up to the next station. There should be at least the pursuit distance between two closest vessel in the overtaking lane traveling the same direction. After overtaking is completed, vessels move back to the main lanes.

444

Table 2: Comparison of average waiting time of vessels

(until current conditions stabilize) when current speed exceeds 6 knots. 3.

OUTPUTS OF THE MODEL All Vessels

This model is run for the 13 months time period (between 1 December 2008 and 1 January 2010). The first month is considered as the warm up period. Some performance measures determined for the analysis are: R1:The average waiting time of vessels (aggregate and vessel type based); R2:Total number of vessels passed; R3:Average number of vessels in queues; R4:The entire Strait vessel density; R5: Pilot utilization;

Waiting Times (in minutes) The Simulation Model 2009 Half Data Average Width 842 814.4 123.13

Relative Error (%) -3.28

5. SCENARIO ANALYSIS AND RESULTS Four factors are selected for the scenario analysis of the simulation model: A: minimum pursuit distance (in time units) between vessels B: vessel profile C: pilot policy D: arrival rate The levels of identified factors for scenario analysis are displayed in Table 3.

4. VERIFICATION AND VALIDATION Due to the fact that the simulation model in this study consists of many submodels integrated into the main traffic model running concurrently, it is difficult to monitor the system. However; with the trace module of Arena, arrival of each vessel, attributes assigned to it, its movement to the anchorage area or to the appropriate queues and its admittance to the Strait are followed clearly, while simultaneously watching entities related to meteorological events affecting the system. Moreover, animation reveals all events in the whole system; therefore, logic errors can be captured easily. Variable indicator of the Arena is also a frequently utilized tool in this study. The change in values of performance measures can directly be traced by variable indicators. Extreme condition verification is first performed by increasing vessels arrival rates by 20% in a three month simulation run. When compared to the base scenario, average vessel waiting time shows more than fifteenfold increase (from 541 minutes to 9272 minutes), average number of vessels in queues increase from 52.6 to 1154.4, number of vessels passed increases to 14756 from 12845 and pilot utilization increases from 0.23 to 0.25. Another extreme conditions effect is reducing the total number of pilots in the model to 12 instead of 20. The model is run for one year with 25 replications and as expected, the pilot utilization, average, maximum waiting time of vessels and number of vessels in queues increased and total number of vessels passed the Strait decreased. The most conclusive of the validation tests in this study are the output comparisons with the real 2009 data. The results of selected performance measures are sufficiently close to the data 2009 to support the claim that the model mimics the actual system reasonably well. As an example, average waiting times of all vessels in model and in actual data are compared. The results are quite similar to each other, as displayed in Table 2.

Table 3: Main factors and their levels in scenario design Factor A

Name pursuit distance

B C D

vessel profile Pilot availability arrival rate

Low 13N11.5S base 16 base

Average 13.5N12S base 20 5% more

High 14N12.5S >=150 m 24 10% more

The first factor A with three levels is the minimum pursuit interval between two consecutive vessels (13N for the low setting means 13 minutes interval for northbound vessels and 11.5S means 11.5 minutes interval for southbound vessels). Regarding the vessel profile factor B, the low setting corresponds to the base scenario in which vessels demand pilots according to the pilot request frequency distribution of vessel classes generated based on the 2009 data. In the high setting, in addition to this random pilot demand, all vessels longer than 150 meters are routinely assigned a pilot while passing the Strait. In pilot availability factor C, the number of available pilots is set at 16 for the low level and 20 for the average level (as is the case in the current system) and 24 for the highest level. According to the last factor, regarding the arrival rate of vessels D, the low setting (which is the setting assumed in the base scenario) is taken as the rates estimated in the interarrival distribution for each subclass based on the 2009 data. In the average level, arrival rate of vessels is increased by 5 per cent (compared to the rates estimated based on the 2009 data) and in the high level, vessel arrival rates are increased by 10 per cent. Accordingly, a total of 54 different scenarios (including the base scenario), are projected and run with 25 replications for a full factorial design. The outputs of these scenarios are gathered from Arena reports, the significant factors and their interactions are investigated through the ANOVA tables in the Design Expert 8.0 software. The percent contribution of each factor on performance measures are displayed in Table 4.

445

for all vessels longer than 150 meters (comparison of scenarios 10 and 12) decrease the waiting time by 4.62 per cent and number of vessels in queues by 5.22 per cent. Furthermore, increasing 20 available pilots to 24 in the system while assigning pilot for all vessels longer than 150 meters under five per cent higher arrival rate (comparison of scenarios 49 and 52) have almost same performance measure results. In another scenario analysis, 4 factors influencing the response variables under high arrival rate conditions (number of arrived vessels increased by 10 per cent) are analyzed. The levels of factor A are 13.5 minutes for northbound and 12 minutes for southbound in low setting and 14 minutes for northbound and 12.5 minutes for southbound in high setting. Regarding the visibility factor (D), the low setting describes the base scenario in which vessels encounter fog events according to the visibility submodel, whereas in the high setting, the fog pattern of the worst case (i.e. the autumn fog realizations which have the longest fog durations) is chosen as the visibility data for the whole year.

Table 4: Percent contributions of main factors Average waiting time Total vessels passed Average transit time Pilot utilization Vessel density

A 38 0.3 0.3 0.3

B

C

0.1

0.1

3.1 0.1

93 0.1

D 59 98 89 3.5 99

AD 2.4 0.1

0.1

In order to track the effects of factors easily, single factor level change in scenarios is investigated through the comparison of scenarios 19, 3, 7 and 16 with the base scenario 1 as can be seen in Table 5. Table 5. Scenarios with various factor level changes Scenarios

R1

R2

R3

R4

R5

1

814

51,178

79.9

9.45

0.24

19 3 7 4 10 12

608 722 754 2289 2275 2170

51,206 51,204 51,200 56,628 56,624 56,677

59.2 70.3 73.4 251 250 237

9.46 9.46 9.45 10.5 10.5 10.5

0.24 0.31 0.25 0.26 0.27 0.22

25 49 52

663 614 622

51,193 53,880 53,882

64.6 62.8 63.7

9.45 9.95 9.95

0.25 0.22 0.25

In order to track the effects of factors easily as displayed in Table 6, level change in scenarios is investigated compared to the base scenario 1. Decreasing pursuit distance to 13.5 minutes for north entrances and to 12 minutes for south entrances (scenario 7) primarily decrease the waiting time by 41.5 per cent, decrease the number of vessels in queues by 41.6 per cent, while keeping the total number of vessels passed almost the same. Setting low visibility conditions (scenario 13) increases average waiting time by 88.8 per cent yet does not significantly change the total number of vessels passed. The effect of two and three factor level changes over responses may also be investigated in this table. For example, although reducing pursuit distances to 13.5 minutes for northbound passages and 12 minutes for southbound passages and deploying 24 pilots instead of 20, the average waiting time increases by 82 per cent under low visibility conditions (comparison of scenarios 7 and 20) and number of vessels in queues increase by 91 per cent.

Decreasing pursuit distance to 13.5 minutes for south entrances and to 12 minutes for north entrances (scenario 19), primarily decrease the waiting time (by 25 per cent), decrease the number of vessels in queues by 26.25 per cent, while keeping the total number of vessels passed and vessel density almost the same. Decreasing the number of available pilots from 20 to 16 (scenario 3) increases pilot utilization by 29.2 per cent and decreases waiting time by 11.30 per cent (the reason why the average waiting time decreases is due to decrease in waiting time of Class D vessels, which enter the Strait more frequently while other vessel types remain waiting because of pilot unavailability). Assigning pilots for all vessels longer than 150 meters (scenario 7) increases pilot utilization by 4.1 per cent. Increasing vessel arrival rate by ten per cent (scenario 4) increases total number of vessels passed by 10.64 per cent, average waiting time by 181 per cent, number of vessels in queues by 212 per cent, pilot utilization by 29.2 per cent and vessel density by 10.8 per cent. The effect of two, three and four factor level changes over responses is also investigated. For instance, decreasing pursuit distance to 13.5 minutes for northbound and to 12 minutes for southbound vessels while assigning pilot for all vessels longer than 150 meters (scenario 25) decrease the waiting time by 18.6 per cent when compared to the base scenario; however, waiting time is increased by 9.9 per cent when compared to the single factor level change case involving 13.5 minutes pursuit distance for south entrances and 12 minutes for north entrances (scenario 19). Table 5 also displays that increasing vessel arrival rate by ten per cent and 20 available pilots to 24 in the system while assigning pilot

Table 6. Scenarios with various factor level changes under high arrival rate conditions Scenarios

R1

R2

R3

R4

R5

1

2289 1367

56629 56850

250 146

10.5 10.5

0.3 0.3

4320 4087 2482 2354

56324 56336 56531 56615

474 448 279 262

10.4 10.4 10.5 10.5

0.3 0.4 0.2 0.2

7 13 15 20 23

In the full factorial analysis of the related scenarios, the 24 different scenarios are experimented through 25 replications (i.e. the scenario analysis is composed of 600 distinct observations). In the scenario analysis, the

446

most effective factor on performance measures is observed as visibility conditions. As fog in the Strait becomes stronger, average waiting time of vessels and transit time increase. Moreover, low visibility conditions decrease total number of vessels passed from directions, pilot utilization and vessel density in the Strait.

Özbaş, B. 2005, Simulation of Maritime Transit Traffic in the Istanbul Channel, M.S. Thesis, Department of Industrial Engineering, Boğaziçi University, Istanbul. Ulusçu, S. Ö., B. Özbaş, T. Altıok and İ. Or. 2009. “Risk Analysis of the Vessel Traffic in the Strait of Istanbul”, Risk Analysis, Vol. 20, No. 10, pp. 1454-1472. VTS Users Guide, Turkish Straits Vessel Traffic Service. 2004. General Management of Coastal Safety and Salvage Administrations, 3rd edition, Istanbul.

6. CONCLUSION AND FURTHER RESEARCH In this study, a simulation model is developed for representing the vessel traffic behavior in the Strait. In this simulation model, maritime rules and regulations about vessel admittance, pursuit distances among vessels, priority levels of distinct vessel types and pilot requirements are all considered. Moreover, submodels representing meteorological conditions such as fog, current and storm are integrated to the model. For validation purposes, the simulation outputs are compared with the actual 2009 data and quite satisfactory results are obtained. In order to analyze the effects of various factors such as vessel arrival rate, vessel profile, pilot availability and minimum pursuit distances between vessels, on performance measures, 54 scenarios are performed with the full factorial design. The most significant factor for all selected variables is observed as the vessel arrival rate. The minimum pursuit distance between vessels is also significant for most performance measures. The interaction of arrival rate and pursuit distance is effective on the most responses, as well. Pilot availability is principally important for pilot utilization. Another scenario analysis is conducted when vessel arrival rate is increased by 10 per cent and the visibility factor is added. Results associated with the considered 24 scenarios show that visibility is the most critical factor for performance measures and its interaction with minimum pursuit distance at different levels is also significant for performance measures such as average waiting time of vessels, number of vessels passed and pilot utilization. This study is planned to be used for risk analysis of the Strait. Incorporating probable vessel accidents and the consequences to the model can have a very beneficial effect for revising the policies and minimizing risk.

AUTHOR BIOGRAPHIES İLHAN OR was born in Istanbul, 1951. He received his BS (1973), MS (1974) and Ph.D. degrees (1976) from Northwestern University, Evanston, Illinois, USA. He has been a faculty member at Department of Industrial Engineering of Bogazici University, Istanbul, Turkey since 1976. He was a visiting faculty member at Syracuse University (1982-1983) and University of Maryland (1983). He served on the “Naval Research Logistics Quarterly” Journal’s Editorial Board between 1993 and 2003.Research areas are environmental and risk management, energy policy and planning, production and maintenance planning. His e-mail address is: [email protected] and personal web page is: http://www.ie.boun.edu.tr/~or/ ŞİRİN ÖZLEM was born in Bursa, 1985. She received her BS (2008) from Uludag University, Bursa,Turkey and MS (2011) from Bogazici University, Istanbul, Turkey. She is currently a doctoral student at the Department of Industrial Engineering of Boğaziçi University, Turkey. Her e-mail address is: [email protected] BİRNUR ÖZBAŞ was born in Istanbul, 1980. She received her BS (2003) in Systems Engineering from Yeditepe University, Istanbul, Turkey MS (2005) and PhD (2010) in Industrial Engineering from Bogazici University, Istanbul, Turkey. She is currently a post doctoral associate at Center for Advance Infrastructure and Transportation (CAIT) / Laboratory for Port Security (LPS), Rutgers, The State University of New Jersey,U.S.A. Her e-mail address is: [email protected]

REFERENCES Almaz, A. Ö., 2006, Investigation of the Maritime Transit Traffic in the Istanbul Channel through Simulation Modeling and Scenario Analysis, M.S. Thesis, Department of Industrial Engineering, Boğaziçi University, Istanbul. Gönültaş, E., Analysis of the Extreme Weather Conditions, Vessel Arrival Processes and Prioritization in the Strait of Istanbul through Simulation Modeling, M.S. Thesis, Department of Industrial Engineering, Boğaziçi University, Istanbul, 2007. Law, A. M. and W. D. Kelton. 2007 Simulation Modeling and Analysis, McGraw-Hill Press, Singapore.

447

NEW GENETIC PROGRAMMING HYPOTHESIS SEARCH STRATEGIES FOR IMPROVING THE INTERPRETABILITY IN MEDICAL DATA MINING APPLICATIONS Michael Affenzeller, Christian Fischer, Gabriel Kronberger, Stephan M. Winkler, Stefan Wagner Upper Austria University of Applied Sciences School for Informatics, Communications, and Media Heuristic and Evolutionary Algorithms Laboratory Softwarepark 11, 4232 Hagenberg, Austria [email protected], [email protected], [email protected], [email protected], [email protected]

One of the reasons for genetic bloat is identified in the tendency of GP to favor more complex hypothesis structures for explaining equivalent correlations (Luke and Panait, 2006). The new proposed offspring selection variant aims to counteract this phenomenon by including additional offspring selection criteria: Instead of only considering the error measure, the enhanced offspring selection (OS) criteria also consider the complexity as well as the number of variables of the candidate hypothesis in order to decide, whether or not a new candidate hypothesis is accepted for the next generation. By this means the hypothesis search should be lead not only to models with more predictive power, but also to more compact and more unique models which are easier to interpret. Especially the latter aspects are considered as important in the field of medical data mining where the domain expert should be able to analyze not only the statistical properties of the prediction models but also their usefulness in the medical context. The effects of the new introduced extended offspring selection formulation for data based modeling are discussed for medical benchmark datasets from the UCI machine learning repository1.

ABSTRACT In this paper we describe a new variant of offspring selection applied to medical diagnosis modeling which is designed to guide the hypothesis search of genetic programming towards more compact and more easy to interpret prediction models. This new modeling approach aims to combat the bloat phenomenon of genetic programming and is evaluated on the basis of medical benchmark datasets. The classification accuracies of the achieved results are compared to those of published results known from the literature. Regarding compactness the models are compared to genetic programming prediction models achieved without the new offspring selection variant. Keywords: Medical data mining, Genetic programming, Offspring selection. 1.

INTRODUCTION Genetic Programming (GP) plays an outstanding role among the various data-mining techniques from the field of machine learning and computational intelligence. Due to its model representation, GP is able to produce human interpretable models without taking any assumptions about the nature of the relationship. Also GP-based data analysis has quite good generalization properties. Furthermore, GP is able to simultaneously evolve the structure and the parameters of a model with implicit feature selection. The combination of these aspects makes GP a very powerful and also robust method for various data analysis tasks.

The rest of the paper is organized as follows: Section 2 describes standard offspring selection with its parameters, its main characteristics, and how it can be integrated into genetic programming. Section 3 discusses specific extensions of offspring selections designed for data based modeling which aim to guide hypothesis search to simpler and easier to interpret models. In section 4 the characteristics of the extended offspring selection variant are discussed exemplarily for medical benchmark data sets. Finally, section 5 summarizes the achieved results and points out future perspectives for future research.

Nevertheless, there are still some aspects in the practical application of GP-based data analysis which leave room for improvement: GP-based data analysis suffers from the fact that – even if the models are interpretable – the results are often quite complex and far from being unique. Often the models are still quite complex because of the tendency of GP to bloat and also because of introns which is counterproductive in terms of interpretability as well.

1

448

http://archive.ics.uci.edu/ml/

the end of a single generation is defined by the quotient of individuals that had to be created until the success ratio was reached and the number of individuals in the population:

2. OFFSPRING SELECTION The basic principles of offspring selection have been described in (Affenzeller and Wagner 2005). In the meanwhile, offspring selection has been discussed for several benchmark problems from the field of combinatorial optimization, function optimization and data based modeling. The following description of standard offspring selection is taken from (Wagner et al 2010) where the aspect of mutation in offspring selection has been discussed in further detail. In general, offspring selection consists of the following steps: At first parents are selected for reproduction either randomly or in any other well-known way of genetic algorithms (e.g., fitness proportional selection, linear rank selection, tournament selection). After crossover and optionally mutation have been applied to create a new child solution, another selection step is introduced which considers the success of the applied reproduction procedure. The goal of this second selection step is to continue the search process with offspring which surpass their parents’ quality. Therefore, a new parameter called success ratio (SuccRatio) is introduced. The success ratio defines the relative amount of members in the next population that have to be generated by successful mating (crossover, mutation). Additionally, it has to be defined when a solution is considered to be successful: Is a child solution better than its parents, if it surpasses the fitness of the weaker, the better, or some kind of mean value of both? For this purpose a parameter called comparison factor (cf) is used to define the success criterion for each created solution as a weighted average of the quality of the worse and the better parent (i.e., if the comparison factor is 0, successful solutions at least have to be better than the worse parent, and if it is 1 they have to outperform the better parent). For steering the comparison factor, the authors decided to introduce a cooling strategy which is similar to simulated annealing. Following the basic principle of simulated annealing, an offspring only has to surpass the fitness value of the worse parent in order to be successful at the beginning of the search process (cf is initialized with 0 or a rather small value). While evolution proceeds solutions have to be better than a fitness value continuously increasing between the fitness of the weaker and the better parent (cf is increased in each generation until it reaches 1 or a rather high value). As in the case of simulated annealing, this strategy leads to a broader search at the beginning, whereas at the end the search process becomes more and more directed. After the amount of successful solutions in the next generation has reached the success ratio, the remaining solutions for the next generation (i.e., (1-SuccRatio)∙|POP|) are taken from the pool of solutions which were also created by crossover and mutation but did not necessarily reach the success criterion. The actual selection pressure ActSelPress at

ActSelPres s

POP ˜ SuccRatio  POOL POP

(1)

Figure 1: Flowchart of Offspring Selection Figure 1 shows these basic steps of offspring selection and how they are embedded into a classical genetic algorithm. Furthermore, an upper limit for the selection pressure (MaxSelPress) can be defined as another parameter which states the maximum number of children (as a multiple of the population size) that might be created in order to fulfill the success ratio. With this additional parameter offspring selection also provides a precise detector for premature convergence: If the algorithm cannot create a sufficient number of successful solutions (SuccRatio∙|POP|) even after MaxSelPress∙|POP| solutions have been created, convergence has occurred and the algorithm can be stopped. If OS is applied with the parameters cf = 1 and SuccRatio = 1, it is commonly referred to as strict OS. Strict OS has the property that children with worse quality compared to its better parent are automatically discarded and therefore the overall quality of the population steadily increases. 3.

NEW OFFSPRING SELECTION FOR DATA ANALYSIS The standard variant of offspring selection as discussed in Section 2 implements the offspring selection criterion purely on the basis of solution quality. For data based modeling the offspring selection criterion is usually based on the mean squared error (MSE) for classification problems and on the coefficient of correlation R2 or MSE for regression problems. This means that an offspring solution candidate is considered successful if the MSE or R2 fitness measure of the candidate offspring is better than the respective fitness measure of the parent solutions. This means that only the quality of the models is considered and not the

449

simplicity of interpretability of the involved solution candidates. The main idea of the here discussed offspring selection extension is that not only the quality of the candidate models should be considered but also its compactness in order to combat the bloat. From theoretical bloat analyses (Luke and Panait, 2006) it is known that genetic programming based hypothesis search tends to find rather more complex models in order to achieve the same model quality. Therefore, it seems reasonable to include also model complexity measures into the offspring selection criterion. In that sense, an offspring candidate model is considered successful if it surpasses not only the model quality of its own parents but is also not more complex than its parent models. As model complexity measures we have introduced the number of nodes as well as the number of used input features (variables) of the involved structure trees. In that sense an offspring solution is considered successful not only if it surpasses the model quality of its own parents; additionally, the offspring model must not be more complex than its parent model. Similar to the standard case of offspring selection we have to decide if the criterion compares the resulting offspring to the better, the weaker, or to some intermediate value. In order to handle this aspect we introduce new model complexity comparison factors: Let qb and qw be the model qualities of the better and the weaker model, lb and lw the length of the shorter (better) and the more complex (worse) model. As a model complexity measure we here use the number of nodes of the two parent structure trees. For the number of variables or the two parent models let vb be the model using less variables (better) and vw the model using more variables (worse). Similar to the standard case of offspring selection comparison factors cfq , cfl , and cfv  >0, 1@ define the certain thresholds which distinguish a successful offspring from an unsuccessful offspring based on the characteristic features of the parents. But in contrast to original offspring selection a candidate offspring has to fulfill three criteria instead of one in order to be accepted; it does not only have to be better but also less complex and use less variables. Similar to the standard case a comparison factor of 0 means that it is sufficient to surpass the certain characteristics of the worse parent whereas a comparison factor of 1 means that the candidate offspring has to be better than the better of the two parents. Obviously it becomes harder to evolve successful offspring solution candidates which results in higher selection pressures on the one hand; on the other hand due to the preference to simpler and more compact models genetic diversity can hardly emerge. Therefore, the additional offspring selection criteria concerning the model complexities and the number of variables should better not be activated from the start. First studies have shown that the new OS variant works a lot better if the additional criteria are activated not until genetic diversity can emerge which usually happens after about

one or two dozen of iterations. Algorithmically we have considered this aspect by introducing further parameters which specify two time windows twl and twv which specify when the additional length and number of variables criterion should be active. Summarizing the above mentioned aspects, the here discussed first version of a new offspring selection criterion dedicated to the reduction of bloat can be stated as follows (in the minimization variant for MSE as quality): ࢏࢙ࡿ࢛ࢉࢉࢋ࢙࢙ࢌ࢛࢒ሺࢉ࢕ǡ ࢖૚ǡ ࢖૛ǡ ࢍࢋ࢔ሻ֞ ൣࢗሺࢉ࢕ሻ ൏ ࢗ࢝ ൅ ࢉࢌࢗ ሺࢗ࢝ െ ࢗ࢈ ሻ൧ࢇ࢔ࢊ ൣ൫࢒ሺࢉ࢕ሻ ൏ ࢒࢝ ൅ ࢉࢌ࢒ ሺ࢒࢝ െ ࢒࢈ ሻ൯࢕࢘ሺࢍࢋ࢔ ‫ ࢒࢚࢝ ב‬ሻ൧ࢇ࢔ࢊ ൣ൫࢜ሺࢉ࢕ሻ ൏ ࢜࢝ ൅ ࢉࢌ࢜ ሺ࢜࢝ െ ࢜࢈ ሻ൯࢕࢘ሺࢍࢋ࢔ ‫ ࢚࢜࢝ ב‬ሻ൧ This means that in order to be considered as successful, a candidate offspring (co) has to be better than some intermediate fitness value of its own parents (defined by cfq) in any case. Additionally, in some predefined time window twl an offspring does not only have to be better but also at least as compact than some intermediate compactness value of its own parents and in the same sense there is a time window twl where the candidate offspring have to use not more variables than some intermediate value of variables used by its parent models. Therefore, also the actual generation gen has to be considered in order to decide if one of the two time windows is active at the moment. The empirical discussion of the next section compares the achieved results on the basis of standard classification benchmark datasets for generating prediction models for breast cancer, thyroid, and melanoma. 4.

RESULTS The configurations used for the test runs in table 2, 4 and 6 with Melanoma, Thyroid and Wisconsin datasets are shown in table 1. If not otherwise stated the time windows include all generations. The maximum solution length is 100, maximum solution height is 12. Up to 1000 Generations are created with a maximum permitted selection pressure of 555. # 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

450

Configuration cfq=0,cfl=0 cfq=1,cfl=0 cfq=0..1,twq=1..100,cfl=0 cfq=0,cfl=-100..1,twl=20..100 cfq=1,cfl=-100..1,twl=20..100 cfq=0..1,cfl=-100..1,twl=20..100 cfq=0,cfl=1,twl=20..100 cfq=1,cfl=1,twl=20..100 cfq=0..1,cfl=1,twl=20..100 cfq=0,cfh=0,twh=20..100 cfq=1,cfh=0,twh=20..100 cfq=0..1,twq=1..100,cfh=0,twh=20..100 cfq=0,cfh=-100..1,twh=20..100 cfq=1,cfh=-100..1,twh=20..100 cfq=0..1,cfh=-100..1,twh=20..100

16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

cfq=0,cfh=1,twh=20..100 cfq=1,cfh=1,twh=20..100 cfq=0..1,cfh=1,twh=20..100 cfq=0,cfv=0,twv=20..100 cfq=1,cfv=0,twv20..100 cfq=0..1,twq=1..100,cfv=0,twv=20..100 cfq=0,cfv=-100..1,twv=20..100 cfq=1,cfv=-100..1,twv=20..100 cfq=0..1,cfv=-100..1,twv=20..100 cfq=0,cfv=1,twv=20..100 cfq=1,cfv=1,twv=20..100 cfq=0..1,cfv=1,twv=20..100 cfq=0,cfl=cfh=cfv=-100..1,twl=twh=twv=20..100 cfq=1,cfl=cfh=cfv=-100..1,twl=twh=twv=20..100 cfq=0..1,cfl=cfh=cfv=-100..1,twl=twh=twv=20..100

Table 1: Configurations for all datasets

03:04 03:24 02:30 39:03 02:48 08:37 01:22 03:53 01:12 01:04 02:46 01:55 09:25 09:15 05:33 01:41 11:14 01:10 02:19 02:47 03:12 15:39 03:00 12:02 01:28 03:32 02:12 10:41 03:07 07:07

Acc.(Tr.)

Acc.(Te.)

Height

Length

Nr.OfVar

MSE(Tr.)

MSE(Te.)

Exec.Time

Table 3: Regular OSGA results Melanoma 4.2. Results Thyroid Dataset The results of the performed test runs with the Thyroid dataset are shown in table 4. The regular OSGA results for Thyroid are shown in table 5.

Table 2: Results with Melanoma dataset

451

Exec.Time

0.074 0.166 0.075 0.068 0.083 0.077 0.075 0.069 0.096 0.077 0.249 0.081 0.066 0.071 0.072 0.077 0.153 0.087 0.073 0.078 0.075 0.074 0.071 0.070 0.074 0.095 0.082 0.070 0.125 0.064

02:49 03:46 04:30

MSE(Te.)

0.091 0.069 0.070 0.062 0.078 0.068 0.080 0.063 0.075 0.073 0.072 0.081 0.060 0.065 0.065 0.075 0.078 0.123 0.071 0.071 0.075 0.054 0.072 0.091 0.077 0.111 0.086 0.064 0.104 0.077

0.100 0.072 0.078

MSE(Tr.)

5.6 6.6 5.8 7.2 5.4 8.8 5.8 7.2 6.6 4.0 8.6 5.4 8.0 7.2 6.4 7.4 5.8 6.2 6.8 6.6 7.2 10.8 4.8 10.0 8.8 6.0 6.8 5.8 6.0 8.8

0.086 0.075 0.071

Nr.OfVar

Exec.Time

31.6 37.8 36.6 49.8 32.2 68.0 30.4 40.8 32.8 18.2 56.8 32.0 66.4 42.0 50.8 46.2 36.8 34.6 49.8 41.0 45.4 85.0 29.6 67.8 45.8 33.6 41.2 43.2 31.2 65.8

3.2 7.0 11.6

Length

MSE(Te.)

9.8 8.2 9.4 9.8 9.0 11.8 8.6 11.0 9.6 7.8 10.6 9.4 12.0 11.4 10.4 11.2 11.2 8.4 11.8 10.6 11.4 12.8 9.4 11.8 11.0 8.6 9.8 10.8 10.2 11.6

15.6 35.6 71.0

Height

MSE(Tr.)

0.925 0.934 0.907 0.929 0.913 0.926 0.924 0.929 0.914 0.920 0.921 0.921 0.927 0.911 0.914 0.914 0.921 0.915 0.921 0.907 0.918 0.927 0.915 0.918 0.917 0.902 0.911 0.917 0.915 0.922

6.8 9.0 13.2

Acc.(Te.)

Nr.OfVar

0.924 0.934 0.923 0.935 0.916 0.931 0.927 0.929 0.927 0.924 0.920 0.917 0.937 0.931 0.925 0.921 0.919 0.918 0.927 0.926 0.930 0.944 0.921 0.925 0.920 0.924 0.911 0.930 0.923 0.932

0.917 0.919 0.915

Acc.(Tr.)

Length

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

0.919 0.931 0.927

#

Height

# a b c

Acc.(Te.)

4.1. Results Melanoma Dataset The results of the performed test runs with the Melanoma dataset are shown in table 2. The regular OSGA results for Melanoma are shown in table 3. Acc.(Tr.)

Figure 2: Melanoma Results: Configuration vs. Quality; little–high complexity (blue–red and bubble size)

#

For comparison purposes the same datasets were used in regular Offspring Selection Genetic Algorithms (OSGA) as seen in table 3, 5 and 7. A strict configuration with a comparison factor of 1 is used. Maximum allowed length is 50, 100 and 200; maximum allowed height is 7, 12 and 17 for configurations a, b, and c respectively.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

0.970 0.983 0.989 0.991 0.983 0.990 0.941 0.984 0.950 0.943 0.983 0.952 0.992 0.988 0.993 0.938 0.982 0.935 0.942 0.987 0.953 0.993 0.989 0.993 0.943 0.979

0.962 0.976 0.987 0.988 0.980 0.986 0.935 0.983 0.945 0.943 0.975 0.947 0.990 0.987 0.991 0.938 0.980 0.936 0.939 0.984 0.952 0.987 0.987 0.989 0.940 0.979

6.6 10.6 7.8 12.6 12.4 12.6 9.0 11.4 8.8 7.4 10.4 9.0 12.2 11.4 11.8 10.2 11.4 10.6 9.2 11.8 10.4 12.8 12.8 12.6 8.6 11.2

14.2 49.8 22.8 85.0 67.2 76.4 30.2 61.8 23.2 29.6 61.2 32.8 88.0 82.2 76.4 37.4 60.0 48.4 34.2 67.6 44.8 91.6 73.0 92.0 24.0 60.2

2.2 6.8 3.2 6.6 7.6 6.0 3.4 6.2 3.0 3.8 6.8 4.8 6.2 7.4 7.4 4.0 6.4 6.2 1.4 6.6 2.2 6.2 7.4 7.0 3.0 7.0

90.87 57.37 61.17 32.81 69.52 36.74 220.92 63.17 183.75 197.61 65.27 150.90 36.00 46.20 34.88 199.67 63.38 216.02 161.51 53.26 136.96 32.45 48.39 41.28 230.51 73.97

220.17 78.24 74.60 39.22 87.90 52.99 207.82 63.53 194.81 192.17 94.61 170.13 59.93 60.26 45.13 186.36 76.04 236.95 160.31 65.30 171.29 50.64 54.27 46.95 235.51 96.89

04:08 08:55 06:48 58:05 13:18 24:02 02:53 05:00 02:17 02:21 05:46 03:19 17:11 33:37 22:04 01:50 03:14 01:34 01:06 13:02 02:10 18:20 45:34 21:06 01:10 03:13

27 28 29 30

0.938 0.993 0.986 0.985

0.935 0.991 0.982 0.983

10.0 11.8 12.8 12.2

40.4 86.0 88.0 75.6

2.0 5.8 7.6 7.6

189.46 39.09 51.26 56.67

189.00 223.87 66.97 90.64

01:43 37:29 09:51 10:13

21 22 23 24 25 26 27 28 29 30

Table 4: Results with Thyroid dataset

0.948 0.965 0.966 0.962 0.940 0.963 0.935 0.966 0.967 0.965

0.938 0.950 0.947 0.950 0.930 0.958 0.921 0.955 0.960 0.960

10.4 12.4 12.2 12.6 9.2 12.2 9.6 11.4 11.8 10.4

30.4 55.4 69.0 79.0 30.8 65.0 30.2 52.8 51.0 53.4

3.2 5.6 6.4 6.0 2.8 6.4 2.8 5.4 6.6 5.0

0.241 0.158 0.144 0.149 0.229 0.175 0.266 0.141 0.148 0.155

0.239 0.217 0.200 0.155 0.255 0.193 0.240 0.213 0.170 0.160

01:28 09:11 03:18 11:06 00:57 03:27 00:53 16:04 02:59 11:01

Table 6: Results with Wisconsin dataset

Exec.Time

70.50 98.74 67.22

08:33 09:36 09:45

Figure 4: Wisconsin Results: Configuration vs. Quality; little–high complexity (blue–red and bubble size)

Table 5: Regular OSGA results Thyroid # a b c

#

Acc.(Tr.)

Acc.(Te.)

Height

Length

Nr.OfVar

MSE(Tr.)

MSE(Te.)

Exec.Time

4.3. Results Wisconsin Dataset The results of the performed test runs with the Wisconsin dataset are shown in table 6. The regular OSGA results for Wisconsin are shown in table 7.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0.956 0.959 0.960 0.960 0.967 0.961 0.944 0.968 0.950 0.947 0.967 0.942 0.963 0.966 0.960 0.947 0.959 0.946 0.944 0.964

0.955 0.947 0.955 0.943 0.959 0.934 0.925 0.959 0.952 0.946 0.955 0.934 0.953 0.949 0.937 0.937 0.946 0.925 0.937 0.959

9.2 12.4 11.4 12.6 12.0 12.2 8.4 11.2 9.4 7.6 11.8 9.0 12.0 12.0 11.8 8.8 12.4 7.4 10.8 12.0

26.4 59.0 39.8 71.4 63.0 67.6 23.6 50.0 23.2 20.4 62.0 31.4 71.6 60.0 58.4 34.8 59.8 15.6 36.4 55.8

4.2 6.0 4.4 4.4 6.8 5.6 3.4 6.4 3.2 3.2 5.8 3.8 6.6 6.2 5.6 4.4 6.6 2.8 3.2 6.8

0.188 0.168 0.190 0.149 0.151 0.162 0.225 0.159 0.239 0.205 0.153 0.208 0.155 0.146 0.158 0.220 0.183 0.256 0.206 0.161

0.204 0.236 0.174 0.211 0.185 0.197 0.238 0.166 0.209 0.232 0.180 0.241 0.189 0.172 0.214 0.239 0.198 0.285 0.202 0.191

02:25 03:03 03:28 14:09 13:39 11:21 01:10 04:17 02:20 01:33 03:17 02:12 14:08 04:41 08:53 01:26 02:49 00:59 01:20 02:41

Exec.Time

MSE(Te.)

50.67 311.36 53.14

MSE(Te.)

MSE(Tr.)

4.8 7.6 9.0

MSE(Tr.)

Nr.OfVar

36.2 68.0 115.4

Nr.OfVar

Length

7.8 12.0 17.4

Length

Height

0.977 0.972 0.986

Height

Acc.(Te.)

0.982 0.977 0.990

Acc.(Te.)

Acc.(Tr.)

a b c

Acc.(Tr.)

#

Figure 3: Thyroid Results: Configuration vs. Quality; little–high complexity (blue–red and bubble size)

0.976 0.979 0.981

0.966 0.966 0.960

8.0 13.0 17.8

36.6 92.8 181.8

5.6 6.6 8.4

0.109 0.087 0.085

0.157 0.139 0.151

53:56 138:34 64:17

Table 7: Regular OSGA results Wisconsin 5.

CONCLUSION In this paper we have considered the aspects of model interpretability and uniqueness in genetic programming based medical data mining. Due to introns and the bloat phenomenon GP models tend to produce more complex than necessary (Luke and Panait, 2006). In contrast to the so called bloat free GP (Silva, 2011) which allows only those models which do not exceed a certain model complexity we have adapted the concept of offspring selection in a way that the hypothesis search process should he guided towards simple and good prediction models. For this purpose the offspring selection criterion has been extended in a way that it considers not only the model quality in order to decide whether or not a candidate hypothesis should be accepted; in addition also the complexity in terms of number of nodes and the interpretability in terms of number of used variables are considered for the offspring selection criterion. The effects of this approach have been analyzed for some well-known

452

benchmark problems from the field of medical data mining. The results show that the new offspring selection criterion is quite sensitive in terms of causing premature convergence due to the loss of genetic diversity caused by the complexity limiting aspects in the OS-criterion. Therefore, it remains as a topic for future research to further develop this new way of hypothesis search in a way that a sufficient amount of genetic diversity is maintained in the GP population. One possible way of achieving such kind of behavior might an automated switch on/off of the additional criteria depending on the average model complexity or the diversity in the actual population.

CHRISTIAN FISCHER received his BSc in software engineering in 2009 from the Upper Austria University of Applied Sciences, Campus Hagenberg. He is currently pursuing studies for his master’s degree. In the course of his studies he is involved in the project team for the prediction of blood demands in a hospital in cooperation with the Josef Ressel Centre Heureka! and the General Hospital Linz. GABRIEL KRONBERGER authored and co-authored numerous papers in the area of evolutionary algorithms, genetic programming, machine learning and data mining. Currently he is a research associate at the Research Center Hagenberg of the Upper Austria University of Applied Sciences working on data-based modeling algorithms for complex systems within the Josef-Ressel Centre for Heuristic Optimization Heureka!.

ACKNOWLEDGMENTS The work described in this paper was done within the Josef Ressel Centre for Heuristic Optimization Heureka! (http://heureka.heuristiclab.com/) sponsored by the Austrian Research Promotion Agency (FFG). REFERENCES Affenzeller, M. and Wagner, S., 2005. Offspring selection: A new self-adaptive selection scheme for genetic algorithms. Adaptive and Natural Computing Algorithms, Springer Computer Science, pp. 218-221. Affenzeller, M., Winkler, S., Wagner, S., A. Beham, 2009. Genetic Algorithms and Genetic Programming - Modern Concepts and Practical Applications. Chapman & Hall/CRC. ISBN 9781584886297. 2009. Luke S. and Panait L., 2006. A Comparison of Bloat Control Methods. Journal of Evolutionary Computation, Vol. 14, No.3 pp. 48-48. Silva S., 2011 Reassembling Operator Equalisation - A Secret Revealed. Proceedings of GECCO 2011, pp. 1395-1403. Wagner, S., Affenzeller M., Beham A., Kronberger G., and Winkler S.M., 2010. Mutation Effects in Genetic Algorithms with Offspring Selection Applied to Combinatorial Optimization Problems. Proceedings of EMSS 2010, pp. 48-48.

STEPHAN M. WINKLER received his MSc in computer science in 2004 and his PhD in engineering sciences in 2008, both from Johannes Kepler University (JKU) Linz, Austria. His research interests include genetic programming, nonlinear model identification and machine learning. Since 2009, Dr. Winkler is professor at the Department for Medical and Bioinformatics at the Upper Austria University of Applied Sciences, Campus Hagenberg. STEFAN WAGNER received his MSc in computer science in 2004 and his PhD in engineering sciences in 2009, both from Johannes Kepler University (JKU) Linz, Austria; he is professor at the Upper Austrian University of Applied Sciences (Campus Hagenberg). Dr. Wagner’s research interests include evolutionary computation and heuristic optimization, theory and application of genetic algorithms, machine learning and software development.

AUTHORS BIOGRAPHIES MICHAEL AFFENZELLER has published several papers, journal articles and books dealing with theoretical and practical aspects of evolutionary computation, genetic algorithms, and meta-heuristics in general. In 2001 he received his PhD in engineering sciences and in 2004 he received his habilitation in applied systems engineering, both from the Johannes Kepler University of Linz, Austria. Michael Affenzeller is professor at the Upper Austria University of Applied Sciences, Campus Hagenberg, and head of the Josef Ressel Center Heureka! at Hagenberg.

453

ON THE USE OF ESTIMATED TUMOR MARKER CLASSIFICATIONS IN TUMOR DIAGNOSIS PREDICTION - A CASE STUDY FOR BREAST CANCER Stephan M. Winkler (a), Michael Affenzeller (b), Gabriel Kronberger (c), Michael Kommenda (d), Stefan Wagner (e), Witold Jacak (f), Herbert Stekel (g) (a – f)

Upper Austria University of Applied Sciences School for Informatics, Communications, and Media Heuristic and Evolutionary Algorithms Laboratory Softwarepark 11, 4232 Hagenberg, Austria (g)

General Hospital Linz Central Laboratory Krankenhausstraße 9, 4021 Linz, Austria (a)

[email protected], (b) [email protected], (c) [email protected], (d) [email protected], (e)[email protected], (f) [email protected], (g) [email protected]

order to identify mathematical models for cancer diagnoses. We have used a medical database compiled at the central laboratory of AKH in the years 2005 – 2008: 28 routinely measured blood values of thousands of patients are available as well as several tumor markers (TMs, substances found in humans that can be used as indicators for certain types of cancer). Not all values are measured for all patients, especially tumor marker values are determined and documented only if there are indications for the presence of cancer. The results of empirical research work done on the data based identification of estimation models for cancer diagnoses are presented in this paper: Based on patients' data records including standard blood parameters, tumor markers, and information about the diagnosis of tumors we have trained mathematical models for estimating tumor markers and cancer diagnoses. In previous work (Winkler et al. 2010; Winkler et al. 2011) we have identified classification models for tumor markers that can be used for estimating tumor marker values on the basis of standard blood parameters. These tumor marker models (also referred to as virtual markers) are now used in combination with standard blood parameters for learning classifiers that can be used for predicting tumor diagnoses. Our goal is to show to which extent virtual tumor markers can replace tumor marker measurements in the prediction of cancer diagnoses. These research goals and the respective modeling tasks are graphically summarized in Figure 1.

ABSTRACT In this paper we describe the use of tumor marker estimation models in the prediction of tumor diagnoses. In previous work we have identified classification models for tumor markers that can be used for estimating tumor marker values on the basis of standard blood parameters. These virtual tumor markers are now used in combination with standard blood parameters for learning classifiers that are used for predicting tumor diagnoses. Several data based modeling approaches implemented in HeuristicLab have been applied for identifying estimators for selected tumor markers and cancer diagnoses: Linear regression, k-nearest neighbor learning, artificial neural networks, and support vector machines (all optimized using evolutionary algorithms) as well as genetic programming. In the results section we summarize classification accuracies for breast cancer; we compare classification results achieved by models that use measured marker values as well as models that use virtual tumor markers. Keywords: Evolutionary Algorithms, Medical Data Analysis, Tumor Marker Modeling, Data Mining,

1. INTRODUCTION, RESEARCH GOALS In this paper we present research results achieved within the Josef Ressel Centre for Heuristic Optimization Heureka!: Data of thousands of patients of the General Hospital (AKH) Linz, Austria, have been analyzed in

454

Training (3)

Standard blood parameter data

Training (2.1)

Virtual tumor markers

Tumor marker data

Training (1)

Tumor diagnosis models (3)

Tumor diagnosis models (1)

Training (2.2)

Tumor diagnosis models (2)

Tumor diagnoses

Figure 1: Modeling tasks addressed in this research work: Tumor markers are modeled using standard blood parameters and tumor markers (training scenario 1); tumor diagnosis models are trained using standard blood values and on the one hand tumor marker data and on the other hand using estimated tumor markers (scenario 2); alternatively we also train diagnosis estimation models only using standard blood parameters and diagnosis information (scenario 3). x

x

x

programming (as described in Section 2.2). In (Winkler et al. 2011), for example, these methods have also been described in detail.

First, models are trained for estimating tumor diagnoses using the full set of available data: Standard blood parameters, tumor marker data and diagnoses information are used. Second, machine learning is applied for learning estimation models for tumor markers. Concretely, we identify classification models that predict tumor marker values for given standard blood parameters either as “normal” or as “elevated”. Subsequently models are trained for estimating tumor diagnoses using standard blood parameters and diagnoses data, but instead of tumor marker values the estimated tumor marker classifications (calculated using models learned in the first modeling step) are used. Third, we also train diagnosis estimation models only using standard blood parameters and diagnosis information.

2.1. Hybrid Modeling Using Machine Learning Algorithms and Evolutionary Algorithms for Parameter Optimization and Features Selection Feature selection is often considered an essential step in data based modeling; it is used to reduce the dimensionality of the datasets and often conducts to better analyses. Given a set of n features F = {f1, f2, …, fn}, our goal here is to find a subset of F, F', that is on the one hand as small as possible and on the other hand allows modeling methods to identify models that estimate given target values as well as possible. Additionally, each data based modeling method (except plain linear regression) has several parameters that have to be set before starting the modeling process. The fitness of feature selection F' and training parameters with respect to the chosen modeling method is calculated in the following way: We use a machine learning algorithm m (with parameters p) for estimating predicted target values est(F',m,p) and compare those to the original target values orig; the coefficient of determination R² function is used for calculating the quality of the estimated values. Additionally, we also calculate the ratio of selected features |F'|/|F|. Finally, using a weighting factor D, we calculate the fitness of the set of features F' using m and p as fitness(F’, m, p) = D·|F'|/|F| + + (1-D)·(1-R2(est(F',m,p),orig)). (1) In (Alba et al. 2007), for example, the use of evolutionary algorithms for feature selection optimization is discussed in detail in the context of gene selection in cancer classification. We have now used

In the following section (Section 2) we summarize the machine learning methods that were applied in this research project, in Section 3 we give an overview of the data base that was used, and in Section 4 we document the modeling results that could be achieved. This paper is concluded in Section 5. 2. MACHINE LEARNING METHODS APPLIED In this section we describe the modeling methods applied for identifying estimation models for tumor markers and cancer diagnoses: On the one hand we apply hybrid modeling using machine learning algorithms and evolutionary algorithms for parameter optimization and feature selection (as described in Section 2.1), on the other hand we use genetic

455

evolutionary algorithms for finding optimal feature sets as well as optimal modeling parameters for models for tumor diagnosis; this approach is schematically shown in Figure 2: A solution candidate is here represented as [s1…np1…q] where si is a bit denoting whether feature Fi is selected or not and pj is the value for parameter j of the chosen modeling method m. This rather simple definition of solution candidates enables the use of standard concepts for genetic operators for crossover and mutation of bit vectors and real valued vectors: We use uniform, single point, and 2-point crossover operators for binary vectors and bit flip mutation that flips each of the given bits with a given probability. Explanations of these operators can for example be found in (Holland 1975) and (Eiben 2003). We have used strict offspring selection (Affenzeller et al. 2009): Individuals are accepted to become members of the next generation if they are evaluated better than both parents. In (Winkler et al. 2011) we have documented classification accuracies for tumor diagnoses using this approach for optimizing feature set and modeling parameters. The following machine learning algorithms have been applied for identifying estimators for selected tumor markers and cancer diagnoses: Linear regression, k-nearest neighbor learning, artificial neural networks, and support vector machines.

We have used the following parameter settings for our GP test series: The mutation rate was set to 20%, gender specific parents selection (Wagner 2005) (combining random and roulette selection) was applied as well as strict offspring selection (Affenzeller et al. 2009) (OS, with success ratio as well as comparison factor set to 1.0). The functions set described in (Winkler 2008) (including arithmetic as well as logical ones) was used for building composite function expressions. In addition to splitting the given data into training and test data, the GP based training algorithm used in our research project has been designed in such a way that a part of the given training data is not used for training models and serves as validation set; in the end, when it comes to returning classifiers, the algorithm returns those models that perform best on validation data.

2.2. Genetic Programming As an alternative to the approach described in the previous sections we have also applied a classification algorithm based on genetic programming (GP, Koza (1992)) using a structure identification framework described in Winkler (2008) and Affenzeller et al. (2009).

Figure 2: Genetic programming including offspring selection.

1 0

0

1

1

0 0

1

4 0.2 7

0.551

1 0

1

1

0

0 0

1

5 0.3 4

0.482

0 0

0

1

1

0 1

1

7 0.8 6

0.693

Offspring selection

Parents selection, crossover, mutation 1 0

1

1

1

0 1

0

6 0.5 8

Full medical data set (blood parameters, tumor marker target values)

Evaluation, i.e., modeling: Data subset (selected blood parameters, tumor marker values, diagnosis data)

lin. reg., kNN, ANN, SVM, … (k-fold cross validation)

Figure 3: A hybrid evolutionary algorithm for feature selection and parameter optimization in data based modeling. Machine learning algorithms are applied for evaluating feature sets.

456

3.

x

DATA BASIS

Third, all samples are removed that contain less than 15 valid values. x Finally, variables with less than 10% valid values are removed from the data base. This leads to a data set specific for breast cancer; this BC data set consists of 706 samples with 324 samples (45.89%) labeled with ‘0’ and 382 (54.11%) labeled with ‘1’. The following variables are stored in this so compiled data set BC: Age, sex, tumor diagnosis (0/1), ALT, AST, BSG1, BUN, C125 (TM), C153 (TM), , CBAA, CEA (TM), CEOA, CH37, CHOL, CLYA, CMOA, CNEA, CRP, FE, FER, GT37, HB, HDL, HKT, HS, KREA, LD37, MCV, PLT, RBC, TBIL, TF, and WBC. Three tumor markers (C125, C153, and CEA) are available in BC.

3.1. General Information The blood data measured at the AKH in the years 20052008 have been compiled in a database storing each set of measurements (belonging to one patient): Each sample in this database contains an unique ID number of the respective patient, the date of the measurement series, the ID number of the measurement, standard blood parameters, tumor marker values, and cancer diagnosis information. Patients’ personal data were at no time available for the authors except for the head of the laboratory. In total, information about 20,819 patients is stored in 48,580 samples. Please note that of course not all values are available in all samples; there are many missing values simply because not all blood values are measured during each examination. Further details about the data set and necessary data preprocessing steps can for example be found in Winkler et al. (2010) and Winkler et al. (2011), e.g. Standard blood parameters include for example the patients’ sex and age, information about the amount of cholesterol and iron found in the blood, the amount of hemoglobin, and the amount of red and white blood cells; in total, 29 routinely available patient parameters are available. Literature discussing tumor markers, their identification, their use, and the application of data mining methods for describing the relationship between markers and the diagnosis of certain cancer types can be found for example in Koepke (1992) (where an overview of clinical laboratory tests is given and different kinds of such test application scenarios as well as the reason of their production are described) and Yonemori (2006). Information about the following tumor markers is stored in the AKH database: AFP, CA 125, CA 15-3, CA 19-9, CEA, CYFRA, fPSA, NSE, PSA, S-100, SCC, and TPS. Finally, information about cancer diagnoses is also available in the AKH database: If a patient is diagnosed with any kind of cancer, then this is also stored in the database. Our goal in the research work described in this paper is to identify estimation models for the presence of the breast cancer, cancer class C50 according to the International Statistical Classification of Diseases and Related Health Problems 10th Revision (ICD-10).

4. MODELING TEST SERIES We have trained models for estimating breast cancer diagnoses using genetic programming as described in Section 2.2 with strict OS and population size 200; as rather small and compact models are preferred by the authors’ medical cooperation partners, the maximum size of the evolved models was set to 20, 35, and 50 nodes. For modeling and test results achieved using bigger model structures please see Winkler et al. (2011). For training virtual tumor markers we have used the hybrid modeling approach described in Section 2.1: 5-fold cross validation was applied, linear regression (linReg), k-nearest-neighbor (kNN) learning, artificial neural networks (ANNs), and support vector machines (SVMs) have been used as machine learning algorithms, and their feature selections and modeling parameters have been optimized by an evolutionary algorithm. Details about these machine learning approaches and their implementation can for example be found in Winkler et al. (2010) and Winkler et al. (2011). We have used the implementations in the open source framework HeuristicLab (Wagner (2009)) (http://dev.heuristiclab.com). 4.1. Modeling Strategies For training estimation models for breast cancer we have applied four different strategies: One using measured tumor markers, one using virtual tumor marker classifiers (combined with OR-conjunctions), one using virtual tumor marker classifiers (combined with majority voting), and finally one not using tumor markers at all.

3.2. Data Preprocessing Before starting the modeling algorithms for training classifiers we had to compile a data set specific for breast cancer diagnosis: x First, blood parameter measurements were joined with diagnosis results; only measurements and diagnoses with a time delta less than a month were considered. x Second, all samples containing measured values for breast cancer are extracted.

4.1.1. Strategy I: Using standard blood parameters and measured tumor markers Measured tumor markers were used as well as standard blood parameters, classification models for breast cancer diagnoses were trained using GP as described above. This corresponds to scenario 1 as described in Section 1.

457

4.1.2. Strategy II: Using standard blood parameters and virtual tumor marker classifiers We have used hybrid modeling as described in Section 2.1 for creating virtual tumor marker classifiers on the basis of the data available in the BC data set. The population size for the optimization algorithm (a GA with strict OS) was set to 20, the maximum selection pressure to 100, and D to 0.1; the test classifications of the so identified best virtual tumor marker classifiers were used as estimated tumor marker values. Classification models for breast cancer diagnoses were trained using GP as described above using virtual tumor markers as well as standard blood parameters. This corresponds to scenario 2 as described in Section 1. We applied each machine learning algorithm used here (namely linear regression, support vector machines, neural networks, and kNN classification) twice in each modeling process, leading to eight estimated binary classifications for each tumor marker; these classifications were combined into one binary classification variable for each tumor marker using either an OR conjunction or majority voting: Strategy II.a: Using OR: If any of the classifiers for a tumor marker return 1 for a sample, then this sample’s virtual tumor marker is 1; else it is set to 0. Strategy II.b: Using majority voting: If more than the half of the classifiers for a tumor marker return 1 for a sample, then this sample’s virtual tumor marker is 1; else it is set to 0.

measured tumor markers perfectly, as the classification accuracy of models using measured TMs reaches ~77.7%. Future work on this topic shall include the investigation of virtual TMs for other types of diseases; furthermore, we will also focus on the practical application of the here presented research results in the treatment of patients.

AverageClassificationAccuracy

0,8 0,78 0,76 0,74 0,72 0,7 0,68 0,66

InputData

4.1.3. Strategy III: Using only standard blood parameters In this strategy no tumor markers were used; instead, only standard blood parameters were available for training classification models for breast cancer diagnoses using GP as described above. This corresponds to scenario 3 as described in Section 1.

Figure 4: Average test accuracies achieved using the four modeling strategies described in Section 4.1 ACKNOWLEDGMENTS The work described in this paper was done within the Josef Ressel Centre for Heuristic Optimization Heureka! (http://heureka.heuristiclab.com/) sponsored by the Austrian Research Promotion Agency (FFG).

4.2. Test Results As already mentioned, each strategy was executed using five-fold cross validation; we here report on the average classification accuracies on test samples (P ± V) which are also shown in Figure 4: x Strategy I: 0.777 ± 0.104 x Strategy II.a: 0.713 ± 0.107 x Strategy II.b: 0.752 ± 0.042 x Strategy III: 0.699 ± 0.113

REFERENCES Affenzeller, M., Winkler, S., Wagner, S., A. Beham, 2009. Genetic Algorithms and Genetic Programming - Modern Concepts and Practical Applications. Chapman & Hall/CRC. ISBN 9781584886297. 2009. Alba, E., García-Nieto, J., Jourdan, L., Talbi, E.-G., 2005. Gene selection in cancer classification using PSO/SVM and GA/SVM hybrid algorithms. IEEE Congress on Evolutionary Computation 2007, pp. 284 – 290. Eiben, A.E. and Smith, J.E. 2003. Introduction to Evolutionary Computation. Natural Computing Series, Springer-Verlag Berlin Heidelberg. Holland, J.H., 1975. Adaption in Natural and Artifical Systems. University of Michigan Press. Koepke, J.A., 1992. Molecular marker test standardization. Cancer, 69, pp. 1578–1581.

5. CONCLUSIONS In the test results summarized in Section 4 we see that the virtual tumor markers have turned out to be able to improve classification accuracy for the modeling application described in this paper: Whereas classifiers not using tumor markers classify approximately 70% of the samples correctly, the use of virtual tumor markers (combined using majority voting) leads to an increase of the classification accuracy to ~75%. Still, virtual tumor markers have in this example not been able to replace

458

Rechenberg, I., 1973. Evolutionsstrategie. Friedrich Frommann Verlag. Schwefel, H.-P., 1994. Numerische Optimierung von Computer-Modellen mittels der Evolutionsstrategie. Basel: Birkhäuser Verlag. Wagner, S., 2009. Heuristic Optimization Software Systems - Modeling of Heuristic Optimization Algorithms in the HeuristicLab Software Environment. PhD Thesis, Institute for Formal Models and Verification, Johannes Kepler University Linz, Austria. Winkler, S., Affenzeller, M., Jacak, W., Stekel, H., 2010. Classification of Tumor Marker Values Using Heuristic Data Mining Methods. Proceedings of Genetic and Evolutionary Computation Conference 2010, Workshop on Medical Applications of Genetic and Evolutionary Computation, pp. 1915–1922. Winkler, S., Affenzeller, M., Jacak, W., Stekel, H., 2011. Identification of Cancer Diagnosis Estimation Models Using Evolutionary Algorithms – A Case Study for Breast Cancer, Melanoma, and Cancer in the Respiratory System. Proceedings of Genetic and Evolutionary Computation Conference 2011, Workshop on Medical Applications of Genetic and Evolutionary Computation. Winkler, S., 2009. Evolutionary System Identification Modern Concepts and Practical Applications. Schriften der Johannes Kepler Universität Linz, Reihe C: Technik und Naturwissenschaften. Universitätsverlag Rudolf Trauner. ISBN 978-385499-569-2. Yonemori, K., Ando, M., Taro, T. S., Katsumata, N., Matsumoto, K., Yamanaka, Y., Kouno, T., Shimizu, C., Fujiwara, Y., 2006. Tumor-marker analysis and verification of prognostic models in patients with cancer of unknown primary, receiving platinum-based combination chemotherapy. Journal of Cancer Research and Clinical Oncology, 132(10), pp. 635–642.

MICHAEL AFFENZELLER has published several papers, journal articles and books dealing with theoretical and practical aspects of evolutionary computation, genetic algorithms, and meta-heuristics in general. In 2001 he received his PhD in engineering sciences and in 2004 he received his habilitation in applied systems engineering, both from the Johannes Kepler University of Linz, Austria. Michael Affenzeller is professor at UAS, Campus Hagenberg, and head of the Josef Ressel Center Heureka! at Hagenberg. GABRIEL KRONBERGER received his PhD in engineering sciences in 2010 from JKU Linz, Austria, and is a research associate at the UAS Research Center Hagenberg. His research interests include genetic programming, machine learning, and data mining and knowledge discovery. MICHAEL KOMMENDA finished his studies in bioinformatics at Upper Austria University of Applied Sciences in 2007. Currently he is a research associate at the UAS Research Center Hagenberg working on data-based modeling algorithms for complex systems within Heureka!. STEFAN WAGNER received his PhD in engineering sciences in 2009 from JKU Linz, Austria; he is professor at the Upper Austrian University of Applied Sciences (Campus Hagenberg). Dr. Wagner’s research interests include evolutionary computation and heuristic optimization, theory and application of genetic algorithms, and software development. WITOLD JACAK received his PhD in electric engineering in 1977 from the Technical University Wroclaw, Poland, where he was appointed Professor for Intelligent Systems in 1990. Since 1994 Prof. Jacak is head of the Department for Software Engineering at the Upper Austrian University of Applied Sciences (Campus Hagenberg) where he currently also serves as Dean of the School of Informatics, Communications and Media.

AUTHORS BIOGRAPHIES STEPHAN M. WINKLER received his PhD in engineering sciences in 2008 from Johannes Kepler University (JKU) Linz, Austria. His research interests include genetic programming, nonlinear model identification and machine learning. Since 2009, Dr. Winkler is professor at the Department for Medical and Bioinformatics at the University of Applied Sciences (UAS) Upper Austria at Hagenberg Campus.

HERBERT STEKEL received his MD from the University of Vienna in 1985. Since 1997 Dr. Stekel is chief physician at the General Hospital Linz, Austria, where Dr. Stekel serves as head of the central laboratory.

459

AUTOMATIC SELECTION OF RELEVANT DATA DURING ULTRASONIC INSPECTION T. Merazi Meksen(a), M. Boudraa(a), B. Boudraa(a) (a)

University of Science & Technology H. Boumedienne BP 32, El Alia, Bab Ezzouar, Algiers, Algeria. [email protected]

decade, ultrasonic techniques have shown to be very promising for non destructive testing (Blitz 1996) and they are becoming an effective alternative to radio-graphic tests. X-ray widely used to detect and sizing discontinuities, presents the disadvantage to produce ionising radiation and needs to develop a film, which takes some times to provide the results. Operators are often required to acquire and interpret large volumes of complex inspection data. So, automated signal analysis systems are finding increasing applications in a variety of industries where the diagnostics is difficult. Ultrasonic data can be displayed as images and can add additional and significant dimension in NDT information and thus for exploiting in applications. Many advanced image processing algorithms have provided powerful techniques to extract from ultrasonic images the desired information on sizing and defect detection (Chen 2004; .Merazi, 2006; Jasiuniene 2007). But all these methods require considerable amount of computation, making them difficult for real-time operation. Many mechanized inspection techniques, sensors, and systems for automating defect detection and location have been developed (Cchatzakos 2007, Martin 2007, Berke 2000; Moles 2005; Shuxiang 2004). However, the location and sizing of a defect is an almost entirely manual process: An operator will mark on the scan, using a mouse, where the component echoes lie, and thus where defect lies. The apparatus will then perform the correction and give an indication of the defect size according to what has been indicated by the operator. Hardwares have been developed and integrated tools of image processing are implemented in order to completely automate the control. Most of these algorithms are computationally intensive, so it is desirable to implement them in high performance reconfigurable systems. Recently, Field Programmable Gate Array (FPGA) technology becomes a viable target for the implementation of algorithms for image processing. (Satoh 2009; Johnston 2004; Nelson 2000). E. Ashari developed a method for NDT image binarization by thresholding, implemented in FPGA (Ashari 2004). K.Appiah uses a single ship of FPGA in

ABSTRACT In recent years, research concerning the automatic interpretation of data from non destructive testing (NDT) is being focused with an aim of assessing embedded flaws quickly and accurately in a cost effective fashion. This is because data yielded by NDT techniques or procedures are usually in the form of signals or images which often do not present direct information of the condition of the structure. Signal processing has provided powerful techniques to extract from ultrasonic signals the desired information on material characterization, sizing and defect detection. The imagery available can add additional and significant dimension in NDT information and for exploiting information. The task of this work is to minimize the volume of data to process replacing ultrasonic images type TOFD by sparse matrix, as there is no reason to store and operate on a huge number of zeros. A combination of two types of neural networks, a perceptron and a Self Organizing Map of Kohonen is used to distinguish between a noise signal from a defect signal in one hand, and to select the sparse matrix elements which correspond to the locations of the defects in the other hand. This new approach to data storage will provide an advantage for the implementations on embedded systems. Keywords: workstation design, work measurement, ergonomics, decision support system. KeywordsAutomatic Ultrasonics; Neural Networks.

tesing;

Materials;

1. INTRODUCTION The use of non destructive testing (NDT) allows the analysis of internal properties of structures without causing damage to the material. Various methods have been developed to detect defects in structure and to evaluate eventually their locations, sizes and characteristics. Some of these methods are based on analysis of the transmission of different signals such as ultrasonics, acoustic emission, thermography, xradiography, eddy current (Cartz 1995). In the last

460

order to extract background models presented in an image and to reduce inspection time (Appiah 2005) and D. J Durlington uses a reconfigurable features FPGA performing a variety of operations in hardware, the control program being executed on a microprocessor (Durlington 1997). On ultrasonic images the zone of interest is often very small in comparison with the image dimensions. This make sense to use a special matrix type called sparse matrix, where only the non zero elements are stored. Not only this reduces the amount of data to store, but also operations of this type of matrix can take advantage of the a-priori knowledge of the positions of non-zero elements to accelerate the calculations (Pissanetzky; Duff 1987 ). The aim of this work is to minimize the data to store and to process in order to save memory and computational time. An original approach for data acquisition and representation, which consists on sparse matrix construction instead of an ultrasonic image type TOFD (Time Of Flight Diffraction) is described. It is based on the TOFD technique but avoids the image formation. The sparse matrix is built by combination of a perceptron and a self organizing map algorithm of Kohonen in order to select a defined number of samples from the signals. Section 2 and 3 in this paper describe respectively ultrasonic non destructive inspection and TOFD technique. In section 4, the two types of Neural Networks used in this work are developed, namely perceptron and Self Organizing Map of Kohonen. Experimental measurements and application of combination of the neural nets are described in section 5. Section 6 concerns the conclusion. Papers that don’t adhere to the guidelines provided in this template will be returned to authors for appropriate revision.

The pulser section of the pulser/receiver generates short electrical pulses which travel through the cabling to the transmitting transducer. The transducer converts thes electrical pulses into acoustic pulse at its acoustic output port, which can be or not be in contact with the material under control. In the latter case, a liquid (couplant) is used to facilitate the transmission of ultrasonic vibrations from the transducer to the test surface. This ultrasonic beam is also transmitted into the solid component being inspected and interacts with any flaw that is present. The flaw generates scattered wave pulses travelling in many directions, and some of these pulses reach the receiving transducer which converts them into electrical pulses. These electrical pulses travel again through cabling to the receiver section of the pulser/receiver, where they are amplified and displayed as a received A-scan voltage Vr(t) as a function of the time. Figure 2 shows an example:

Figure 2: Example of an ultrasonic signal 3. THE TOFD TECHNIQUE TOFD technique uses the travel time of a diffracted wave at the tip of a discontinuity (Silk1984). The TOFD research concludes that technique is portable, fast, reliable, accurate and inexpensive in the defect detection and sizing. Further, inspection can be semi or fully automated for the defect detection in metal structures. Two transducers, one as a transmitter and the second as a receiver are moved automatically step by step according to a straight line and the diffracted signals are recorded and displayed as images. Those images provide different texture patterns for the detected defects and automatic texture segmentation is investigated using different techniques to improve the detection and classification of defects. In his thesis, J. Sallard demonstrates that a generator of a hole, can be assimilated to a top of a crack (Sallard 1999). So, a test block containing a hole has been used in this work to test this method (figure 3).

2. ULTRASONIC INSPECTION The basic components of an ultrasonic inspection system are a pulser/receiver, the cabling, the transducers, and the acoustic/elastic wave propagation and scattering (Figure 1).

Figure 1: Signal acquisition system for ultrasonic inspection

461

4.1 The Perceptron The Perceptron is a binary classifier that maps its input x (a real-valued vector) to an output single binary value. If two sets are linearly separable, this classification can be used to decide whether a given vector belongs to one class or another (Rosenblatt 1962).

Displacement Probes

w1

Figure 3: Test block with an artificial defect

X1

Figure bellow shows the result obtained scanning a test block containing an artificial crack. Every row is constituted of a samples of a reached signal.

Perceptron

w2

a1

wN

a2

X2 XN

20

Figure 5: Functional description of a perceptron

pixels

40 60

The function of each neuron is to compute a weighted sum of all synapse inputs, add the sum to a predefined bias and pass the result through a nonlinear sigmoidal (threshold) function f whose output ranges between 0 and 1:

80 100 120

100

200 300 pixels

400

500

a = f (u) =

Figure 4:TOFD image showing a defect IV. ARTIFICIAL NEURAL NETWORKS

.

(1)

In this work, inputs (X1, X2,...XN) correspond to the N signal samples. The outputs a1 and a2, are respectively the defect-signal-classe and the noise-signal-classe.

The Artificial Neural Networks (ANN) parallel, distributive computational structure is reminiscent of the human neural system. In an ANN structure, many simple nonlinear processing elements, called neurons, are interconnected via weighted synapses to form a network inputs. The functionality is mostly dependent on the values of the weights which can be updated over time, causing the neural network to adapt and possibly “learn”. The learning process is of different types: supervised learning, unsupervised learning, selforganized learning. In a supervised approach, the network is fed with necessary input and the appropriate output for the specified inputs is given the output is achieved together with a global error function. The computed output is compared to the desired output to evaluate the performance of the neural network. The computed error function is then used to update the weights with an aim of achieving output that is close to the desired output. In contrast to the supervised learning, unsupervised learning or self organizing learning does not require any assistance of desired outputs or an external teacher. Instead, during the training session, the neural network receives a number of different patterns and discovers significant features in these patterns and learns how to classify input data into appropriate categories.

4.2 Kohonen Self-Organizing Maps Kohonen Self Organizing Maps (SOM) is a widely used ANN model based on the idea of self organized or unsupervised learning (Kohonen 1988). The SOM network is a data visualization technique, which reduces the dimensions of data through a variation of neural computing networks. It is a non parametric approach that makes no assumptions about the underlying population distribution and in independent of prior information. The problem that data visualization attempts to solve is that humans simply cannot visualize high dimensional data. So, SOM goes about reducing dimensions by producing maps of usually one or two dimensions which plot the similarities of the data by grouping similar data items together. Thus, SOM’s accomplish two things : they reduce dimensions, and display similarities. In this work, this property allows to select a fixed number of relevant data to store and to process. Figure 6 demonstrates basic structure of self-organizing Kohonen map:

462

Kohonen layer

Input layer

P1 P2

Pn

I1

K1

I2

K2

In

KM

Figure 6: Self Organizing Kohonen Map network

Figure 7: Example of ultrasonic reached signal. The horizontal coordinate of the maximum (indicated by the cross) corresponds to the time of flight of the reached ultrasonic wave

The network has input and output layers of neurons that are fully interconnected among themselves. At each step of training phase an output layer’s neuron with weights that best match with input data (usually in a minimum Euclidian distance sens) is proclaimed as the winner. The weights of this neuron and its neighborhood neurons are then adjusted to be closer to the presented input data. The algorithm is described as follows: 1- Initialize W with uniform-random values. 2- For each input vector P, compute the distance d between the vector P and the M weight vectors based on the mean square root: d = (Pk – Wl)2

l = 1, 2, 3,…M

From every signal of defect-signal class, the coordinates pi and pj corresponding respectively to the time of flight and the position of the probe, according to the straight line of the displacement, are stored. At the end of this processus, a set of points P(pi, pj) are determined and their number equals the number of the signals in class a1. The Self Organization Map of Kohonen is applied with those points as inputs in order to reduce their number to a defined one depending on the desired sparse matrix dimension (30 in this work). Figure below shows the positions of the output elements resulting, obtained using the signals that form the TOFD image on figure 4.

(2)

3- Select the vector Wx such that Wx satisfies Equation (3) : l = 1, 2,3,…M (3) (Pk–WX)2=min(Pk-Wl)2

80

Wx(t+1)=Wx(t) + Į(Xk- Wx) (0 < Į < 1) 5- Go to step 2 until Wl § Pl

probe positions

4- Update Wx using Equation 4: (4)

.

In the early learning stage, Į is set about 0.8. As the learning progresses, Į gradually becomes closer to 0.

60 40 20 0

V. MEASUREMENTS AND RESULTS According to the principle of the TOFD technique, two transducers (2 Mhz) are moved step by step by 5 mm each

time, straight a line. At each position of the probes, the reached diffracted signal is first analyzed by a perceptron in order to determine if it is a “defect signal” or “a noise signal”. In the next step, every defect signal (classe a1) is processed and the sample corresponding to the maximum of the amplitude is detected. This sample corresponds to the time of flight of the reached ultrasonic signal (cross on figure bellow).

0

200 400 Time of Flight (samples)

600

Figure 8: Sparse Matrix obtained instead of the TOFD image of figure 4. Instead of 120x500 pixels of the TOFD image, this method selects 30 elements which are sufficient to describe the pattern presented in the image. This number is defined by the Self organizing Map algorithm outputs and is independent of the quantity of initial data.The economy of memory is important and the defect location and characterization will be faster when analyzing only the sparse matrix elements. V. CONCLUSIONS In this work, a method to detect and locate cracks by analyzing a sparse matrix built from TOFD signals has been described. A first layer of a neural net selects a

463

point from the reached signal if an echo signal is presented in the zone of interest. The co-ordinates of this point which correspond respectively to the probe position and the time of flight of the signal are stored and used as inputs for a self organizing map network. Outputs will represent a group of points corresponding to the defect presented in the structure. Results of the application of this technique have been promising in terms of speed, and robustness. This would make the proposed system suitable for implementation in situations requiring near real-time processing and interpretation of large volumes of data giving thus an important help in the decision making.

Martin C.J and Gonzalez Bueno R.. 2007. Ultrascope TOFD : Un sistema compacto para captura y procesamiento de imagenes TOFD . IV Conferencia Panamericana de END, PANNDT 2007, Buenos Aires, Aregentina, October 22-26, .Merazi Meksen T &. Boudraa M, (2006), Application of the Randomized Hough Transform on Ultrasonic images in Non Destructive Testing. WSEAS transactions on signal processing Vol.2, Issue 8, pp. 1053-1056 Aug. 06. Moles M. (2005), Portable Phase Array Application, 3rd, Middle East Nondestructive Testing Conference & Exhibition MENDT. Manama, Bahrein, 27-30 Nov. Nelson A.E., 2000. Implementation of Image Processing Algorithm on FPGA Hardware. Thesis in Electrical Engineering, Faculty of the graduate school of Vanderbilt. Pissanetzky, Sparse Matrix Technology, http://www.scicontrols.com Rosenblatt F (1962) Principles of neurodynamics. New York, Spartan. Sallard J. 1999. Etude d’une Méthode de Déconvolution Adaptée aux Images Ultrasonores. Thesis présented at the Institut National Polytechnique de Grenoble. France. Satoh K and Tad J., 2009. Three dimensional ultrasonic Imaging Operation using FPGA, IEICE Electronics Express. Vol. 2, N°2, pp 84-89. Shuxiang J. 2004, Developement of an Automated Ultrasonic Testing System , SPIE Proceeding, 3rd International Conference on Experimental Mechanics, Vol.5852, Singapore. Silk M.G., 1984. The Use of Diffraction-Based Time of Flight Measurements to Locate and Size Defects, British Journal of NDT, Vol 26, Pages 208-213, May 1984. Yella S. and. Dougherty M.S., 2006, Artificial Intelligence Techniques for the automatic Interpretation of Data from Non Destructive Testing Insight, Vol. 48, N°1, January 2006.

REFERENCES Appiah K and.Hunter A. 2005. A Single-Ship FPGA Implementation of Reaf time Adaptive Background Model. Proceeding of IEEE International Conference on Field Programmable Technology. FPT’05, pp. 95-102.Singapore, December,11-14. Ashari E.and Hornsey R., 2004 ‘ FPGA Implementation of Real Time Adaptive Image Thresholding’, Thesis presented to University of Waterloo, Ontario, Canada. Berke M and Kleinert W.D. 2000. Portable Work Station for Ultrasonic Weld Inspection ,15th World Conf. of Non Destructive Testing, WCNDT2000, Roma, Italy. Cartz, L. 1995. Nondestructive testing, Radiography ultrasonics, Liquid penetrant, Magnetic Particle, Eddy Current. ASM International. Cchatzakos P.and. Markopoulos Y. P. 2007. Towards Robotic Non Destructive Inspection of Industrial Pipelines . 4th International Conference on NDT, HSNDTint 2007, Chania-Crete, Greece Oct. 1114. Chen C.H. 2004. Advanced Image Processing Methods for Ultrasonic NDE Research . World Congress of Non Destructive Testing, Proc. WCNDT 2004 , Aug. 30-Sep. 3,2004, pp 39-43. Montreal, Canada. Darlington D.J. and Campbell, D.R.. 1997 Reconfigurable FPGAs for data Compression in Ultrasonic Non Destructive Testing, IEE Colloquium on DSP chips in Real Time Measurement and Control. 25 September, Leicester, UK. Duff I.S and Erisman A.M., 1987, Direct Method for Sparse Matrix, Clarendon Press, New York, USA. Jasiuniene J. (2007), Ultrasonic Imaging Techniques for Non Destructive Testing of Nuclear Reactors, cooled by liquid Metals: Review. ULTRAGRAS, Vol.62, N° 3, pp.39-43. C.J Johnston and K.T Gribbon, 2004. ‘Implementing Image Processing Algorithms on FPGAs ‘, Proceeding of 11 Electronic New Zeland Conference ENZCon’04, pp. 118-123, Palmerston North, New Zeland. Kohonen T. 1988, Self Organization and Associative Memory , Springer Verlag, Heidelberg..

464

CT BASED MODELS FOR MONITORING BONE CHANGES IN PARAPLEDIC PATIENTS UNDERGOING FUNCTIONAL ELECTRICAL STIMULATION Páll Jens Reynisson(a), Benedikt Helgason(b), Stephen J. Ferguson (b), Thordur Helgason (c),(d), Rúnar Unnþórsson(a), Páll Ingvarsson(e), Helmut Kern(f), Winfried Mayr(g), Ugo Carraro(h) and Paolo Gargiulo (c),(d)

a. School of Engineering and Natural Sciences, Faculty of Industrial Engineering, Mechanical Engineering and Computer Science, University of Iceland, b. Institute for Surgical Technology and Biomechanics, University of Bern, Switzerland c. Department of Development and Consultancy UTS, Landspitali-University Hospital, Reykjavik, Iceland. d. Department of Biomedical engineering, University of Reykjavik, Iceland. e. Department of Rehabilitation Medicine, Landspitali-University Hospital. f. Ludwig Boltzmann Institute of Electrostimulation and Physical Rehabilitation, Department of Physical Medicine, Wilhelminenspital. Vienna, Austria. g. Medical University of Vienna, Center of Biomedical Engineering and Physics h. Laboratory of Translational Myology of the University of Padova, Department of Biomedical Sciences, Padova, Italy. (a)

[email protected] , (b)[email protected], (b)[email protected] (c),(d) [email protected], (a)[email protected] (e)[email protected], (f)[email protected] (g) [email protected], (h)[email protected], , (c),(d)[email protected],

fracture in the paraplegic’s extremities as a consequence (Maïmoun et al 2005; Lazo et al 2001; Jiang et al 2006). In order to decrease this acceleration of tissue deterioration, electrotherapy such as Functional Electrical Stimulation (FES) has been proposed (Gargiulo, 2008; Gargiulo et al 2009, 2011; Kern 2002, 2005; Gallasch & Rafolt et al 2005; Mandl et al 2008). However, recent studies claim that the lack of osteogenic response in paralyzed extremities to electrically evoked exercise during sub-acute and rehabilitation/recovery phases, could not be fully explained, and may warrant further evaluation (Clark et al 2007). CT data can be used to monitor bone changes in paraplegic patients by quantification of morphological parameters. Additionally CT based finite element models can provide useful information on the structural changes that result from the changes in bone morphology when direct mechanical testing is not possible. The absence of mechanical stimulation in the lower extremities of paraplegic patients makes the patella bone, ideal for monitoring such changes resulting directly from external stimulation. The aim of the present study was thus to perform a CT based evaluation of the osteogenic response of the patella of paraplegic patients undergoing FES.

ABSTRACT Spinal Cord Injured (SCI) paraplegics suffer from pathological changes in the lower extremities such as muscle degeneration, hormonal alterations and bone resorption. The aim of the present study was, with the help of finite element and image analysis, to evaluate the osteogenic response of the patella of paraplegic patients undergoing Functional Electrical Stimulation (FES). For this purpose, a patient that began a daily home based FES treatment 5 years after paralysation was monitored. Bone mechanical parameters were compared at the beginning of FES and after 3 years of treatment. According to our results, it is possible to conclude that application of long term FES treatment on denervated, degenerated muscles can be beneficial for bone growth in bones attached to the stimulated muscles. Keywords: Finite Element Modelling, Functional Electrical Stimulation, Flaccid Paraplegia, Bone Strains, Patella. 1. INTRODUCTION Spinal Cord Injured (SCI) paraplegics suffer from pathological changes in the lower extremities, such as muscle degeneration, hormonal alterations and bone resorption (Maïmoun et al 2006). Loss of bone mineral density (BMD), which is one of the symptoms of osteoporosis (Marcus et al 2008), results in bones becoming more fragile, with an increased risk of

2. MATHERIAL AND METHODS A pre-treatment CT dataset of the lower extremities for a paraplegic patient (a 32 year old male suffering

465

N after 3 year of FES treatment. The volumetric strain histograms derived from the FE simulations are illustrated in Fig. 1, with a Frost interval of [200, 2000] micro strains (Frost HM, 1987) and, Rubin and Lanyon threshold of 1000 micro strains, (Rubin & Lanyon, 1987) indicated but these authors reported these thresholds being relevant for maintaining bone mass. 71% of the total bone volume was found to be strained beyond 200 micro strains and 5% beyond 1000 micro strains at the start of FES treatment. The corresponding results after three years of FES were 71% and 19%. The calculated patella weight (mg), bone volume (mm3) and average Young's modulus (MPa) per element are presented in Tab. 1. Fig. 2 and 3 illustrate a comparison between the Young’s modulus distribution for the healthy subject and the patient before and after FES treatment.

complete flaccid ThXI syndrome with paralysis and areflexia in the legs and medium atrophy) was used to create a FE model of the patient’s right patella. A QUASAR phosphate phantom (Modus Medical Devices Inc., London, Ontario, Canada) was used to calibrate the images. The construction of the FE model was carried out in several steps briefly explained as follows: I) Creation of a 3D triangular surface mesh (STL) through semi-automatic segmentation in MIMICS (Materialize Interactive Medical Image Control System, Leuven, Belgium) using a Hounsfield (HU) threshold of 200 to determine the boundary between bone and soft tissue. II) Creation of a solid model (IGES) in SOLIDworks (Dassault Systèmes SolidWorks Corp., Concord, Massachusetts, USA). III) Creation of a 10node tetrahedral FE mesh in ANSYS Workbench (ANSYS, Canonsburg, Pennsylvania, USA). IV) FE model imported into ANSYS (ANSYS, Canonsburg, Pennsylvania, USA), where FE equations are solved. Rigid boundary conditions were applied at the patella ligament and the patella tendon attachments points. Force (Fr) was applied on the model were the distal femur contacts the patella. For determining this force a biomechanical model of the knee joint (Ward 2004; Ward 2005) was used. The joint moments were measured during stimulation at the beginning and three years into the FES treatment, with a non invasive pendulum test (Gallasch & Rafolt et al 2005). Moment arms and force directions in the sagittal plane were derived from the CT data. Isotropic, linear elastic, heterogeneous material properties were assigned to each node in the model with an in-house MATLAB script (The Mathworks, Natick, MA) based on the NI material mapping method introduced by Helgason et al. (2008b). The relationship (1) between Young’s modulus (E) and apparent bone density (ρ) was taken from Morgan et al. (2003): E = 6850ρ1.49 (MPa)

(1)

Poisson‘s ratio was set to 0.3. A CT dataset acquired after 3 years of daily home based FES treatment was also available for the same patient. Using the procedure described above, this CT dataset was used to create another FE model for comparison to the pre-treatment situation. After FE simulations, the equivalent Von Mises element strains were derived from the FE solutions and compared as shown in Fig. 1. Additionally, Young’s modulus distribution, patella bone total volume, weight and average Young’s modulus were derived directly from the CT data as shown in Fig. 3 and Tab. 1. A third CT dataset of a healthy, 37 year old male was available for comparison to the patient results but FE simulation was not carried out for this individual.

Figure 1. Mechanical strain stimulation according to FE simulations, before FES treatment and after 3 year of FES treatment. The red lines indicate the Frost interval [200, 2000] micro strains (Frost HM 1987) but the grey line indicates the Rubin and Lanyon threshold of 1000 micro strains.

3. RESULTS The maximum load on the patella during FES, derived from the biomechanical knee model, was found to be Fr = 60±2 N at the start of FES treatment and Fr = 123±3

466

Figure 3. Young‘s modulus distribution (MPa). Anterior view of the right patella. 1. Patella of a healthy subject. 2. Paraplegic patient after 5 years of paralysis. 3. Same Paraplegic patient after 5 years of paralysis and 3 years of FES treatment.

Figure 2. Histograms of Young’s modulus distribution. 1. Patella of a healthy subject. 2. Paraplegic patient after 5 years of paralysis. 3. Same Paraplegic patient after 5 years of paralysis and 3 years of FES treatment.

Sets

Trabecular bone Average Weight Young’s Modulus (MPa) (mg) 7210 2490 4820 790 5340 960

4. CONCLUSION The aim of the present study was to evaluate the osteogenic response of the patella of a paraplegic patient undergoing FES using the finite element method and image analysis. The results, as shown in Fig. 2, indicate that bone strain stimulus at the start of treatment was sufficient for bone formation according to published thresholds for bone maintenance (Rubin & Lanyon 1984 and Frost 1987). The results also indicate, that even after 3 years of FES treatment, strain stimulus was larger than was found at the beginning of treatment. This suggests that there is potential for further bone formation. Comparing the stiffness for the pre- and posttreatment situation as shown in Fig. 3, indicates that during FES the bone is adapting to the loads being applied, especially at the patella ligament and

Volume (mm3) 14100 13100 13600

Healthy Subject 5.y of paralysis 5.y of paralysis and 3 y of FES Table 1. Trabecular bone quantities (HU range 2001000 HU); Total weight (mg), average Young's (MPa) and total volume.

467

Kern H, Hofer C, Modlin M, Forstner C, RaschkaHögler D, Mayr W, Stöhr H.,2002. Denervated muscles in humans: limitations and problems of currently used functional electrical stimulation training protocols. Artificial Organs, 26(3): 216–218. Kern H, Salmons S, Mayr W, Rossini K, Carraro U., 2005. Recovery of long-term denervated human muscles induced by electrical stimulation. Muscle Nerve, 31(1):98–101. Lazo MG, Shirazi P, Sam M, Giobbie-Hurder A, Blacconiere MJ, and Muppidi M., 2001. Osteoporosis and risk of fracture in men with spinal cord injury. Spinal Cord, 39(4):208-214. Mandl T, Meyerspeer M, Reichel M, Kern H, Hofer C, Mayr W, Moser E., 2008. Functional Electrical Stimulation of Long-term Denervated, Degenerated Human Skeletal Muscle: Estimating Activation Using T2-Parameter Magnetic Resonance Imaging Method. Artificial Organs, 32(8): 604–608. Maïmoun L. Fattal C. Micallef J-P. Peruchon.E. P Rabischong., 2006. Bone loss in spinal cordinjured patients: from physiopathology to therapy. Spinal Cord 44:203–210. Morgan EF. Barnes GL. Einhorn TA. The Bone Organ System: Form and Function., 2008. third edition, OSTEOPOROSIS Chapter 1:3-25. Morgan EF. Bayraktar HH. and Keaveny TM. Trabecular bone modulus-density relationships depend on anatomic site., 2003. Journal of Biomechanics, 36(7) :897-904. Parfitt AM., 2008. Skeletal Heterogenity and the Purposes of Bone Remodeling: Implications for the understanding of Osteoporosis. third edition, OSTEOPOROSIS Chapter 2:71-89. Rubin CT. & Lanyon LE. Kappa Delta Award Paper. Osteoregulatory nature of mechanical stimuli: Function as a determinant for adaptive remodeling in bone., 1987. Journal of Orthopaedic Research Volume 5(2):300-310. Tadashi S. Kaneko, Jason S. Bell, Marina R. Pejcic, Jamshid Tehranzadeh Joyce H. Keyak., 2004. Mechanical properties, density and quantitative CT scan data of trabecular bone with and without metastases. Journal of Biomechanics, 37(4) :523530. Thimas SJ. Relative electron density calibration of CT scanners for radiotherapy treatment planning., 1999. The British Journal of Radiology, 72(860) :781-786. Ward SR and Powers CM. The Influence of Patella Alta on Patellofemoral Joint Stress During Normal and Fast Walking., 2004. Clin. Biomech, 19(10) :10401047. Ward SR. Terk MR. Powers CM. Influence of patella alta on knee extensor mechanics., 2005. Journal of Biomechanics, 38(12) :2415-2422.

quadriceps femoris insertion points. The trabecular bone appears to adapt to the load increase by increasing both mass and volume which results in increased average Young´s modulus. However, the influence of long term disuse on the cortical shell can clearly be seen in Fig. 3, where Young’s modulus distribution in the patella for the healthy and paralysed subjects is compared. According to our results it is possible to conclude that application of long term FES treatment on denervated degenerated muscles can be beneficial for bone growth in bones attached to the stimulated muscles. ACKNOWLEDGMENTS This work has been supported by The University Hospital research fund of Iceland Landspitali, Keilir Institute of Technology and Department of Biomedical engineering, University of Reykjavik. REFERENCES Clark JM. Jelbart M. Rischbieth H. Strayer J.Chatterton B. Schultz C. and Marshall R., 2007. Physiological effects of lower extremity functional electrical stimulation in early spinal cord injury: lack of efficacy to prevent bone loss. Spinal Cord, 45: 78-85. Frost HM., 1987. Bone ”mass” and the ”mechanostat”: a proposal. Anat Rec, 219(1) :1-9, Gallasch E & Rafolt D. Kinz G. Fend M. Kern H. and Mayr W., 2005. Evaluation of FES-Induced Knee Joint Moments in Paraplegics with Denervated Muscles. Artificial Organs, 29(3) :207-211, Gargiulo P. Helgason T. Reynisson PJ. Helgason B. Kern H. Mayr W. Ingvarsson P., Carraro U., 2011. Monitoring of Muscle Recovery in Spinal Cord Injury Patients Treated With Electrical Stimulation Using Three-Dimensional Imaging and Segmentation Techniques: Methodological Assessment, Artificial organs, 35(3): 275-281. Gargiulo P. Vatnsdal B. Ingvarsson P. Knútsdóttir SS. Gudmundsdóttir V. Yngvason S. Kern H. Carraro U. Thordur Helgason., 2009. Computational methods to analyze tissue composition and structural changes in denervated muscle undergoing therapeutic electrical stimulation, Basic Applied Myology, 19 (4):157-161. Helgason B. Perilli E. Schileo E, Taddei F. Brynjólfsson S. and Viceconti M., 2008a Mathematical relationship between bone density and mechanical properties: A literature review. Clinical Biomechanics, 23(2):124, 135-146. Helgason B. Taddei F. Pálsson H, Schileo E, Cristofolini L. Viceconti M. and Brynjólfsson S., 2008b. A modified method for assigning material properties to FE models of bones. Medical Engineering & Physics, 30(4):444-453. Jiang SD, Dai LY, and Jiang LS, Osteoporosis after spinal cord injury., 2006. Osteoporos Int, 17(2):160-192.

468

Applied Myology/The European Journal of Translational Myology since 1991, he founded and chair since 2005 the University of Padova Interdepartmental Research Center of Myology. Paolo Gargiulo is assistant professor at Reykjavik University and works as biomedical engineer and researcher at the University Hospital of Iceland. He studied electrical engineering at University Federico II in Naples, and obtained a PhD at Technical University of Vienna, Austria. His research interests are in the field of electrical stimulation medical modeling and rapid prototyping for clinical applications.

AUTHORS BIOGRAPHY Páll Jens Reynisson is a lecturer and engineer at Keilir Institute of Technology. He earned his bachelor’s degree in mechanical and industrial engineering at University of Iceland in 2007 and obtained a MSc. in biomedical engineering from Reykjavik University in 2010. Benedikt Helgason is a research associate at the Institute for Surgical Technology and Biomechanics, at the University of Bern in Switzerland. He earned his bachelor’s degree in civil engineering from the University of Iceland in 1993, a master’s degree in civil engineering from the Technical University of Denmark in 1996 and a Ph.D. degree in biomedical engineering at the University of Iceland in 2008. Stephen J. Ferguson is currently the head of the Biomechanics Division at the Institute for Surgical Technology and Biomechanics, at the University of Bern in Switzerland. He earned his bachelor’s degree in mechanical engineering at the University of Toronto in 1991, a master’s degree from the same university in 1994 and a Ph.D. degree from Queen’s University in 2000. Stephen J. Ferguson has been appointed a full professor at the ETH in Zurich from 1st of July, 2011. Thordur Helgason is Associate professor at Reykjavik University and works as biomedical engineer and researcher at the University Hospital of Iceland. He studied electrical engineering at University of Iceland, and obtained a Ph.D. at Technical University of Karlsruhe, Germany. His research interests are in the field of electrical stimulation biomedical technologies. Rúnar Unnþórsson is head of Keilir Institute of Technology. He studied mechanical engineering at University of Iceland, and obtained a CS in 1997, MSc in 2002 and Ph.D in 2008. Rúnar Unnþórsson has been teaching at University of Iceland since 2001. Rúnar is appointed assistant professor at the University of Iceland from 1st of august, 2011. Páll Ingvarsson is neurologist specialised at Goteborg University, Sweden. He works at Department of Rehabilitation Medicine, Landspitali-University Hospital,Reykjavik,Iceland. Helmut Kern is head of the "Department of Physical Medicine and Rehabilitation of the Wilhelminenspital" (Vienna, Austria) since 1984 and director of the research institute "Ludwig Boltzmann Institute of Electrical Stimulation and Rehabilitation" since 1988. Winfried Mayr is Associate Professor of Biomedical Engineering and Rehabilitation Technology. He works at Medical University of Vienna Center for Medical Physics and Biomedical Engineering. His research interests are in the field of Functional electrical stimulation (mobilization of paraplegics, phrenic pacing, EMG-controlled stimulation, pelvic floor applications, denervated muscles, application of FES in microgravity and bed-rest) Implant technology (Microelectronical and electromechanical implants). Ugo Carraro is Associate Professor of General Pathology at the Faculty of Medicine of the University of Padova, since 1983. Editor-in-Chief of Basic and

469

A HIGH RESOLUTION DISTRIBUTED AGENT-BASED SIMULATION ENVIRONMENT FOR LARGE-SCALE EMERGENCY RESPONSE Glenn I. Hawe, Graham Coates, Duncan T. Wilson, Roger S. Crouch

School of Engineering and Computing Sciences, Durham University, United Kingdom [email protected], [email protected], [email protected], [email protected]

mechanisms of the real system, limiting its potential to provide understanding and testable predictions regarding the problem it addresses. If a model is too complex, its analysis will be cumbersome and likely to get bogged down in detail.” This leads to the concept of the “Medawar zone” (Grimm et al. 2005), a region of complexity in which the ABM is not only useful for its intended purpose, but is also structurally realistic. Together, the usefulness and structural realism of an ABM determine the “model payoff”. The Medawar zone may be seen as an application of the Aristotelian notion of “the golden mean”, the most desirable region between two extremes, to ABM design. Whilst North and Macal (2007) assert that “realistic agent behaviors are the key to agent-based modeling”, they also state that “properly specified agent environments are critical for correct agent operation.” Thus it is important that both the agents and their environment are appropriately represented. In the remainder of this section, we briefly discuss different representations of agents and their environment in ABMs for large-scale emergency response, and consider their relation to model payoff. We also briefly discuss the high-level software organization of existing ABMs. Then in Section 2 we propose the use of two different representations for both the environment and the agents: one for around incident sites, and another for elsewhere. In Section 3 we describe our software which uses these two representations. Section 4 provides a summary, and describes future planned developments and use of the software.

ABSTRACT This paper describes the architecture of an agent-based simulation environment for large-scale emergency response. In an effort to increase “model payoff”, it uses two representations for both the environment and the agents. Around each incident, where topographical information is necessary, an operational level simulator program models the environment using the OS MasterMap topography and Integrated Transport Network (ITN) layers. At these incident sites, first responder agents are modelled with a rich repertoire of actions. The remaining area of interest, encompassing the locations of relevant resource bases (e.g. ambulance stations, hospitals and fire stations) which are outside of the incident sites, is modelled using only transport network information by a tactical level simulator program. This program also simulates the tactical level agents, communicates with each operational level simulator, and provides a viewer. A separate PreSimulator program allows new scenarios to be set up with ease. Keywords: agent-based simulation, emergency response 1. INTRODUCTION The simulation of emergency scenarios is an important part of the preparedness stage of the emergency management cycle (Haddow 2010). In silico simulation in particular finds numerous applications within emergency preparedness and response (Jain and McLean 2003, Longo 2010). Agent-based models are a popular way of simulating responses to large-scale emergencies (Khalil et al 2009). Roberts (2010) describes an agent-based model (ABM) design as: “...one in which analogs of those real-world entities that are to be modeled are represented as software agents, or objects, at a level of detail and resolution necessary to address the questions the model is required to answer.” The importance of finding the appropriate level of resolution in an ABM is emphasized by Grimm et al. (2005): “Finding the optimal level of resolution in a bottom-up model's structure is a fundamental problem. If a model is too simple, it neglects essential

1.1. Agent representations for large-scale emergency response The range of complexity which is possible when implementing agents, from simple production systems to cognitive architectures, is discussed by Gilbert (2007). An agent consists of a state (variables) and behavior (methods). The behavior of an agent manifests itself through the actions it decides to make. At the highest level, decision making may be descriptive or normative (Peterson 2009). Depending on which is used to implement the behavior of the agents, one of two conceptually different ABMs may arise:

470

1.

Descriptive: An ABM in which the behaviors of agents are designed to mimic that of their real-life counterparts.

2.

Normative: An ABM in which agents have behavior which is not based on reality. Instead their behavior is designed to be optimal, according to some criteria.

different research groups differ in what they deem appropriate. Müller (1999) gives eleven general guidelines for choosing the right agent architecture to apply to a specific problem, one of which is “Do not break a butterfly upon a wheel” (i.e. do not waste effort in developing complex agents when simpler agents suffice). While Sun (2006) points out that most social simulation tools “embody very simplistic agent models, not even comparable to what has been developed in cognitive architectures”, Gilbert (2006) questions when cognitive architectures are needed anyway. Grimm et al (2005) point out that many ABMs “try only one model of decision-making and attempt to show that it leads to results compatible with a limited data set”, and point out the flaws in doing so. One characteristic which existing ABMs for largescale emergency response do share however is that within each ABM, the representation of any particular agent is constant. For example, whether a firefighter agent is leaving the fire station, travelling to an incident scene, or inside the inner cordon, it is modelled using the same code, and thus at the same level of complexity, throughout the simulation.

An example of an ABM which has agents whose behavior is defined by normative decision making is the RoboCup Rescue Simulation (RRS) (Skinner and Ramchurn 2010). The aim in the annual RRS competition is to design ex novo behaviors for police, fire brigade and ambulance agents, along with their control centers, to optimize an objective function which combines the health of civilians and damage to property. As Carley et al (2006) says, “it is concerned with designing smart algorithms, not with investigating a current human social system as it exists and designing a public policy for it.” Such “smart algorithms” are often quite complex. A simple, yet crude, measure of the complexity of agent representation could be the number of lines of code taken to implement it. For example, thousands of lines of Java code were used to implement agent behaviors in RoboAkut (Akin 2010), the winning entry in the RRS 2010 competition. An interesting approach which yields agents of different complexities is described by Runka (2010). Agents are represented by decision trees, which evolve using genetic programming (Koza 1992). Different fitness functions yield different sized trees, corresponding to agents of different complexity. This paper is concerned with agents whose behavior is designed to mimic that of their real-life counterparts. A variety of ABMs exist which implement such agents, and complexity can vary greatly. For example, the rescuer agents in the ABS SimGenis (Saoud 2006) are quite simple in their implementation (despite being described as having “perceptive and cognitive intelligence”). A set of heuristic rules determine their behaviors. The casualty agents are even simpler, having only a discrete-valued health state, the evolution of which is modelled using a Markov chain. In PLAN-C (Narzisi 2007), rescue agents are also quite simple. For example, the pseudo-code in (Mysore 2006) is only a few lines long. Although pseudo-code, it is low-level enough to suggest that the actual (Java) code would not be significantly longer. More complex are the rescue agents in the AROUND project (Chu 2009), which learn their behavior from their human counterparts through interactive sessions. Weights in a utility function, which combines multiple objectives, are adjusted so as to select actions in a manner which is most consistent with that of the human experts. State of the art cognitive architectures, such as Soar (Lehman 2006) and ACT-R (Anderson 2007), do not appear to have yet been applied to large-scale emergency response ABMs. Just from these examples, it is evident that a range of complexities are possible for representing agents, and

1.2. Environment representations for large-scale emergency response Some large-scale emergencies, such as earthquakes, may cause widespread damage to the environment, whilst others, such as terrorist bombs, may cause localized damage. Some may even cause no damage to the environment, e.g. human pandemics. Thus, the most appropriate representation for the environment is dependent on the type of large-scale emergency being simulated, and in particular the damage it causes. For example, in the ABM EpiSimS (Del Valle 2006), which models the effect of different policies on the spread of human pandemics, only the transport network is modelled. RRS on the other hand, which simulates the response to an earthquake, models the buildings as well, as many will be damaged and need to be considered during the response effort. The representation of the environment in RRS is explored by Sato and Takahashi (2011). They found that the representation of topography influenced simulation results: when modelled at a lower resolution, buildings were found to take a longer time to burn; at a higher level of resolution, gaps between the smaller buildings prevented fire from spreading. In the context of military simulations, which use triangulated irregular networks (TIN) to model terrain, Campbell et al. (1997) point out that “users naturally favour high resolution, high fidelity models because of the realism they offer, but computers that run the simulations may not be able to store and process the amount of data that is associated with these high resolution models.” They propose “by identifying tactically significant (and insignificant) terrain, we can more effectively manage the TIN budget by suggesting areas of terrain that should be modelled at high and low fidelity”, i.e. the use of different resolutions within a single model.

471

Table 1: Activities in the National Occupational Standards for FireFighters in the U.K.

1.3. Software organization of ABMs for large-scale emergency response Many ABMs are single-process programs. However, some do make use of multiple programs, and in particular distributed memory parallelism. RRS makes use of functional parallelism. Different sub-simulators, each of which simulates one aspect of the earthquake response scenario (such as a fire subsimulator, and a flood sub-simulator), run on separate processors. Using more processors enables more functionality to be modelled; however it does not allow larger areas to be simulated. As well as functional parallelism, the IDSS ABM (Koto 2003), which is also designed for earthquake response simulation, uses data parallelism: large geographical regions are split into smaller ones, which are simulated in parallel. The use of up to 34 machines is reported for modelling an area affected by an earthquake. Data parallelism is also used in EpiSimS to split the large environment, spanning five US counties, into smaller regions, which are then simulated in parallel following a master-slave model. Del Valle et al. (2006) report distributing a single simulation over 106 processors. EpiSimS also reports the use of separate programs for “enhanced pre- and post-processing” (Del Valle et al 2006). An “InitializeHealth” program allows the user to specify different probability distributions on the population being modelled. A “graphic user interface enables it to be used by nonprogrammers”. This program is part of a suite of programs that are combined into a “Set-up Wizard”. A set of scripts, which call programs such as gnuplot and Excel, are then used for carrying out post-processing on the output files generated by the simulation.

Ref FF1 FF2 FF3 FF4 FF5 FF6 FF7 FF8 FF9

Activity Inform and educate your community to improve awareness of safety matters Take responsibility for effective performance Save and preserve endangered life Resolve operational incidents Protect the environment from the effects of hazardous materials Support the effectiveness of operational response Support the development of colleagues in the workplace Contribute to safety solutions to minimize risks to your community Drive, manoeuvre and redeploy fire service vehicles

2.

IMPROVING MODEL COMPLEXITY FOR LARGE-SCALE EMERGENCIES In this section, we propose the use of more than one representation for both agents and their environment, when simulating large-scale emergency response. 2.1. Agent representation In the vicinity of an incident site, where the casualties are (possibly trapped), first responders carry out a wide range of actions. Using the National Occupational Standards for firefighters in the U.K. (U.K. Firefighter NOS 2005) as an example, four broad groups of activities may be identified (out of nine) as being directly relevant at the time of an emergency. These four groups are highlighted in bold in Table 1. Using the detailed descriptions of these four activity groups, twelve distinct actions may be identified, as shown in Figure 1. Of these twelve distinct actions, only one is relevant away from the actual incident sites: “driveVehicle” (the action necessary to arrive at the incident site). Thus, away from the incident sites, it is unnecessary to model the full repertoire of twelve actions for firefighters. In this regime, the behavior of firefighters reduces to movement

Figure 1: Identification of FireFighter agent actions

472

1.

along the transport network, and the ABS is more akin to a traffic simulation. The same arguments hold for other first responder agents, such as paramedics and police.

2.

2.2. Environment representation The preceding discussion leads us to two different representations of the environment also. Around incident scenes, information is required (cues) in order for agents to discern which action to perform. The Ordnance Survey MasterMap (OS MasterMap 2011) topography and Integrated Transport Network (ITN) layers are used to model the topography and transport network of rectangular regions centred around each individual incident scene. Away from the incident scenes however, as the only action which agents are performing is the action of moving, only the transport network needs to be represented. Thus, only the ITN layer is used to model the larger area which surrounds the incident sites. This area is sufficiently large to capture all the hospitals, fire stations and police stations that may be involved in resolving the incident. Figure 2 illustrates our approach for the case of the London 2005 bombings. The locations of the bomb explosions are modelled using the topography and ITN layers. Edgware Road (Figure 2 (a)) and Liverpool Street (Figure 2 (b)) are modelled as separate 1 km2 regions. As the Tavistock Square and King’s Cross/Russell Square incidents were quite close, they are modelled together in a larger 3 km2 (1.5 km x 2 km) region (Figure 2 (c)). To capture all the hospitals and fire stations used, the road network in the larger 60 km2 area is modelled using the ITN layer as shown.

3.

A Pre-Simulator program, which is used to set up the emergency to be simulated. A Tactical level simulator program, which: a. simulates tactical level agents, b. simulates first responders when they are travelling along the road network, to and from incident sites, c. provides a viewer for the simulation, d. communicates with the operational level simulator programs. An Operational level simulator program, one instance of which simulates one individual incident site.

Figure 3 shows how these programs are organized. The Pre-Simulator program is used to set up the scenario to be simulated. The details are written to xml files which then serve as input to the Tactical level Simulator program.

Figure 3: High-level software organization More precisely, the Pre-Simulator is used to: 1.

2.

3.

Figure 2: Two representations for the environment

4.

3. SOFTWARE ORGANIZATION In this section, we describe the high-level organization of the software used to set up and perform simulations, using the different representations mentioned in the previous section. Three separate programs make up the agent-based simulation environment:

5.

6.

473

Specify the OS ITN file defining the transport network, and (optionally) specify a folder containing OS StreetView raster files (OS StreetView 2011) covering the same area. This information is saved to transportNetwork.xml. Identify nodes on this transport network which represent locations of resource bases (hospitals, fire stations, ambulance stations and police stations). This information is saved to resources.xml. Set the resources available at each resource base (e.g. the number of ambulances at each ambulance station). This information is saved to initialization.xml. Initialize the positions and crew of each individual resource. This information is also saved to initialization.xml. Specify the individual incident sites, and the OS MasterMap topography file(s) which define the topography in each. This information is saved to sites.xml. Set up the incidents (including casualty information held in text files) at each incident site. This information is saved to crisis.xml.

Figure 5: Setting up the incident sites, and creating incidents in the Pre-Simulator. The tactical level simulator takes the xml input files written by the Pre-Simulator as command line parameters, and uses them to create the virtual environment and populate it with agents. It also provides a viewer of the environment, as shown in Figure 6. Each incident site has its own dockable window showing the topography of the area as defined by its OS MasterMap topography files. These windows are docked in region “a” in Figure 6. The area outside the incident sites is represented by the transport network, but is visualized using the OS StreetView maps. This window is labeled “b” in Figure 6. The tactical level simulator also simulates the tactical level agents (in the strategic-tactical-operational command structure used for major incidents in the U.K. (LESLP 2007)). These agents are responsible for issuing a plan (usually a predetermined attendance) to the available resources. This is represented in the form of an evolving Gannt chart, which is shown in the window labeled “c” in Figure 6. Finally, the window labeled “d” in Figure 6 shows how the estimated total number of fatalities evolves with time.

Figure 4: Setting up the transport network, identifying resource bases, and initializing resources in the PreSimulator. Steps 1-4 of this process are illustrated in Figure 4, whilst steps 5-6 are illustrated in Figure 5. Note that, once the road network is loaded and displayed (Step 1), the user must select nodes as resource bases (Step 2). As it is difficult to identify the appropriate nodes using the road network alone, the first dialog in Figure 4 allows the user to (optionally) specify a folder containing OS StreetView raster files. If specified, these are superimposed onto the Pre-Simulator view which shows the transport network, allowing the user to easily identify the nodes of interest.

474

(a) Shared memory

Figure 6: Screenshot of the viewer The road network is represented as a graph in the tactical level simulator, using the Boost graph library (Siek, Lee and Lumsdaine 2002). This allows the use of Dijkstra’s algorithm, to determine responder agents’ paths to and from the incident sites (once issued with their part of the plan). When an agent enters an incident site, it stops being simulated in the tactical level simulator, and starts being simulated inside the appropriate operational level simulator. Its representation changes from a basic agent which can merely move along a transport network, to a more sophisticated agent which can perceive its environment, as shown in Figure 7, and select among a wide range of actions to perform. The actions for FireFighters have already been given in Figure 1. In a similar manner, eight actions for Paramedic agents and fourteen actions for Police agents have been identified from their National Occupational Standards and Major Incident Plans.

(b) Distributed memory Figure 8: Inter-process communication

4. SUMMARY AND FURTHER WORK The high-level architecture of an agent-based simulation environment, designed specifically for emergency response has been described. In an effort to target the most appropriate level of model complexity, it uses two different representations for both the agents and their environment. Here we describe four further ongoing developments. First, although agents have a rich repertoire of actions available in the operational level simulators, their action selection mechanism is still basic. Efforts are underway to model these mechanisms using naturalistic decision making (Klein 2008), in particular using a recognition-primed decision (RPD) model (Klein 2003, Warwick et al 2001). This has been shown to correspond well to the decision making of emergency first responders, such as firefighters (Burke and Hendry 1997, Klein et al 2010). Second, parallel to the implementation of an RPD model of decision making for operational level agents, validation and verification will be carried out. Practitioners from local Emergency Planning Units, involved from the initial stages of the project, will provide face validation, whilst past case studies, such as the London 2005 bombings, will be used for retrodiction. Third, a post-processor program will be developed to enable the analysis and understanding of simulation results.

Figure 7: An agent perceiving its environment in the operational-level simulator Finally, the tactical level simulator and operational level simulator(s) communicate with one another. The way they do this depends on whether the programs are running on the same machine or different machines, as shown in Figure 8. When on the same machine (Figure 8 (a)), they communicate using shared memory, using the QSharedMemory class from Qt (Qt 2011). When on different machines (Figure 8 (b)), they communicate using sockets, using the QTcpServer and QTcpSocket classes from Qt.

475

Systems for Crisis Response and Management – ISCRAM 2011. May 8-11, Lisbon, Portugal. Chu, T.-Q., Boucher, A., Drougal, A., Vo, D.-A., Nguyen, H.-P. and Zucker, J.-D., 2008. Interactive Learning of Expert Criteria for Rescue Simulations, Lecture Notes in Artificial Intelligence, 5347 pp. 127-138. Del Valle, S., Kubicek, D., Mniszewski, S., Riese, J., Romero, P., Smith, J., Stroud, P. and Sydoriak, S., 2006. EpiSimS Los Angeles Case Study. Technical Report LAUR-06-0666, Los Alamos National Laboratory. Gilbert, N., 2006. When Does Social Simulation Need Cognitive Models? In: R. Sun, ed. Cognition and Multi-Agent Interaction, Cambridge University Press, pp. 428-432. Gilbert, N., 2007. Computational Social Science: Agent-based social simulation. In: D. Phan and F. Amblard, eds. Agent-based modeling and Simulation. Bardwell Press, Oxford, pp. 115-134. Grimm, V., Revilla, E., Berger, U., Jeltsch, F., Mooij, W. M., Railsback, S. F., Thulke, H-H., Weiner, J., Wiegand, T. and DeAngelis, D. L., 2005. PatternOriented Modeling of Agent-Based Complex Systems: Lessons from Ecology, Science, 310 (5750) pp. 987-991. Haddow, G. D., Bullock, J. A. and Coppola, D. P., 2010. Introduction to Emergency Management, Butterworth-Heinemann. Jain, S. and McLean, C., 2003. A Framework for Modeling and Simulation for Emergency Response, Proceedings of the 2003 Winter Simulation Conference, pp. 1068-1076. December 7-10, New Orleans, Louisiana, USA. Khalil, K.M., Abdel-Aziz, M. H., Nazmy, M. T. and Salem, A. M., 2009. The Role of Artificial Intelligence Technologies in Crisis Response. Proceedings of the 14th International Conference on Soft Computing, pp. 293-298. June 18-20, Brno University of Technology, Czech Republic. Klein, G., 1998. Sources of Power: How People Make Decisions. MIT Press. Klein, G., 2008. Naturalistic Decision Making, Human Factors 50 (3), pp. 456-460. Klein, G., Calderwood R. and Clinton-Cirocco, A., 2010. Rapid Decision Making on the Fire Ground: The Original Study Plus a Postscript, Journal of Cognitive Engineering and Decision Making, 4 pp. 186-209. Koza, J. R., 1992. Genetic Programming: On the Programming of Computers by Means of Natural Selection, MIT Press. Koto, T. and Takeuchi, I., 2003. A distributed disaster simulation system that integrates sub-simulators. Proceedings of the First International Workshop on Synthetic Simulation and Robotics to Mitigate Earthquake Disaster. July, Padova, Italy. Lehman, J. F., Laird, J. and Rosenbloom, P., 2006. A Gentle Introduction to Soar, an Architecture for Human Cognition: 2006 Update. Available from:

Finally, the agent-based simulation environment is just one of two software components in the “REScUE” project (Coates et al 2011), being carried out at Durham University. The other component is a decision support system (DSS). The DSS has a two way communication with the tactical level simulator. It receives information about the emergency from tactical level agents as it becomes available, and uses this to generate plans for the responder agents. It then communicates these plans back to the tactical level agents, who may or may not decide to issue them to the operational level agents (who may or may not decide to adhere to the plan, depending on their most up-to-date knowledge of the emergency situation). It is a goal of the REScUE project to identify how to formulate near-optimal plans quickly, especially in the case of rapidly evolving, large-scale, unprecedented events where the practice of predetermined attendances and adhering to standard operating procedures may be far from optimal. ACKNOWLEDGMENTS The authors gratefully acknowledge the funding provided by the UK’s EPSRC (EP/G057516/1). Further, the authors thank practitioners from the Emergency Planning Units from Cleveland and Tyne & Wear, Co. Durham & Darlington Civil Contingencies Unit, Government Office for the North East, Fire and Rescue Services from Co. Durham & Darlington and Tyne & Wear, North East Ambulance Service, and Northumbria Police. REFERENCES Akin, H. L., Yilmaz, O. and Sevim, M. M., 2010. RoboAKUT 2010 Rescue Simulation League Agent Team Description. Available from: http://roborescue.sourceforge.net/2010/tdps/agents /roboakut.pdf Anderson, J., 2007. How Can the Human Mind Occur in the Physical Universe? Oxford University Press. Burke, E. and Hendry, C., 1997. Decision making on the London incident ground: an exploratory study, Journal of Managerial Psychology, 12 pp. 40-47. Campbell, L., Lotwin, A., DeRico, M. M. G. and Ray, C., 1997. The Use of Artificial Intelligence in Military Simulations, Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, Vol. 3, pp. 2607-2612, October 1215, Orlando, Florida. Carley, K. M., Fridsma, D. B., Casman, E., Yahja, A., Altman, N., Chen, L.-C., Kaminsky, B. and Nave, D., 2006. BioWar: scalable agent-based model of bioattacks. IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans 36 (2) pp. 252-265. Coates, G., Hawe, G.I., Wilson, D.T. and Crouch, R. S., 2011. Adaptive Co-ordinated Emergency Response to Rapidly Evolving Large-Scale Unprecedented Events (REScUE). Proceedings of the 8th International Conference on Information

476

Agents and Multi-Agent Systems, pp. 1647-1648. May 10-14, Toronto, Canada. Sun, R., 2006. Prolegomena to Integrating Cognitive Modeling and Social Simulation, In: R. Sun, ed. Cognition and Multi-Agent Interaction, Cambridge University Press, pp. 3-26. U.K. Firefighter NOS, 2005. Skills for Justice National Occupational Standards. Available from: http://www.skillsforjustice-ipds.com/nos.php Warwick, W., McIlwaine, S., Hutton, R. and McDermott, P., 2001 Developing computational models of recognition-primed decision making, Proceedings of the 10th conference on computer generated forces. May 15-17, Norfolk, VA, USA.

http://ai.eecs.umich.edu/soar/sitemaker/docs/misc/ GentleIntroduction-2006.pdf LESLP, 2007. London Emergency Services Liason Panel Major Incident Procedure Manual, The Stationery Office. Available from: http://www.leslp.gov.uk/docs/Major_incident_pro cedure_manual_7th_ed.pdf Longo, F., 2010. Emergency Simulation: State of the Art and Future Research Guidelines, SCS M&S Magazine, Vol. 1. Müller, J., 1999. The Right Agent (Architecture) to do the Right Thing. Lecture Notes in Artificial Intelligence, 1555, pp. 211-225. Mysore, V., Narzisi, G. and Mishra, B., 2006. Agent Modeling of a Sarin Attack in Manhattan. Proceedings of the First International Workshop on Agent Technology for Disaster Management, pp. 108-115. May 8-12, Hokkaido, Japan. Narzisi, G., Mincer, J., Simth, S. and Mishra, B., 2007. Resilience in the Face of Disaster: Accounting for Varying Disaster Magnitudes, Resource Topologies, and (Sub) Population Distributions in the PLAN C Emergency Planning Tool. Lecture Notes in Computer Science 4659, pp. 433-446. North, M. J. and Macal, C. M., 2007. Managing Business Complexity: Discovering Strategic Solutions with Agent-Based Modeling and Simulation. Oxford University Press. OS MasterMap, 2011. Ordnance Survey MasterMap Available from: http://www.ordnancesurvey.co.uk/ oswebsite/products/os-mastermap/index.html OS StreetView, 2011. Ordnance Survey StreetView Available from: http://www.ordnancesurvey.co.uk/ oswebsite/products/os-streetview/index.html Qt, 2011. A Cross-Platform Application and UI Framework. Available from: http://qt.nokia.com/ Peterson, M., 2009. An Introduction to Decision Theory, Cambridge University Press. Roberts, D., 2010. Distributed Agent Based Modeling, Linux Journal. Available from: http://www. linuxjournal.com/content/distributed-agent-basedmodeling Runka, A., 2010. Genetic Programming for the RoboCup Rescue Simulation System. M.S. Thesis, Brock University, Ontario. Saoud, N. B-B., Mena, T. B., Dugdale, J., Pavard, B. and Ahmed, M. B., 2006. Assessing large scale emergency rescue plans: An agent-based approach, The International journal of Intelligent Control Systems, 11 (4) pp. 260-271. Sato, K. and Takahashi, T., 2011. A Study of Map Data Influence on Disaster and Rescue Simulation’s Results In: Q. Bai and N. Fukuta, ed. Advances in Practical Multi-Agent Systems. Springer Berlin /Heidelberg, pp. 389–402. Siek, J. G., Lee, L-Q. and Lumsdaine, A., 2002. The Boost Graph Library, Addison Wesley. Skinner, C. And Ramchurn, S., 2010. The RoboCup Rescue Simulation Platform, Proceedings of the 9th International Conference on Autonomous

AUTHORS BIOGRAPHY Glenn I. Hawe is a Research Associate in the School of Engineering and Computing Sciences at Durham University, where he is currently developing an agentbased simulation environment for large-scale emergency response. He has a PhD in Electronics and Electrical Engineering from the University of Southampton, and an MPhys in Mathematics and Physics from the University of Warwick. Graham Coates is a Senior Lecturer in the School of Engineering and Computing Sciences at Durham University. He has a PhD in Computational Engineering Design Coordination from Newcastle University, and a B.Sc. in Mathematics from Northumbria University. Recently, his research interest in coordination has been extended into the area of emergency response, and an EPSRC (UK) grant has enabled a small team to be brought together to carry out a three year study. Duncan T. Wilson is a PhD student in the School of Engineering and Computing Sciences at Durham University, where he is currently working on a decision support system for large-scale emergency response. He has a B.Sc in Mathematics and an M.Sc in Operational Research from the University of Edinburgh. Prior to Durham he worked for the U.K. Government Operational Research Service. Roger S. Crouch is Head of the School of Engineering and Computing Sciences at Durham University. In 1994 he took up a lectureship (then senior lectureship) at Sheffield University where he set up the Computational Mechanics Group. In 2005 he moved to Durham. His research work has focused on four areas: (i) structural integrity of nuclear reactor vessels under elevated temperature and over-pressure, (ii) computational plasticity of geomaterials, (iii) wave slamming on breakwaters and (iv) simulating reactive processes in groundwater transport.

477

EIGENFREQUENCY BASED SENSITIVITY ANALYSIS OF VEHICLE DRIVETRAIN OSCILLATIONS Oliver Manuel Krieg(a), Jens Neumann(b), Bernhard Hoess(c), Heinz Ulbrich(d) (a)

BMW Group Research and Technology, Hanauer Str. 46, 80992 Munich, Germany (b) BMW AG, Hufelandstr. 4, 80788 Munich, Germany (c) BMW Group Research and Technology, Hanauer Str. 46, 80992 Munich, Germany (d) Institute of Applied Mechanics, Technical University of Munich, Boltzmannstraße 15, 85748 Garching, Germany (a)

[email protected], (b)[email protected], (c)[email protected], (d) [email protected]

Understanding and visualising the relevant sensitivities can seriously improve the design process since it helps identifying parameters that approximate good vehicle setups, gives profound understanding of the dynamics of the system and facilitates solutions for the described conflicts of objectives. Furthermore, no optimisation algorithms are needed, which usually require distinct optimisation criteria and boundaries that often do not exist explicitly, and will not necessarily lead to an improved understanding of the dynamics of the system. A methodology is presented in this paper performing different simplifications, analysing eigenfrequency sensitivities based on (Dresig and Holzweißig 2010) and deriving measures to palliate the conflict of objectives for shuffle and cyclic irregularity for low frequencies for the vehicle drivetrain design process. Numerous works are concerned with the discussed oscillation phenomena of a vehicle drivetrain. A complete overview would go beyond the scope of this paper. Therefore only selected works are presented here. In (Bencker 1998), experimental and simulative studies on shuffle are performed to identify palliative measures for the drivetrain. An analysis of a different engine torque excitation for shuffle follows in (Hülsmann 2007). Various works are concerned with active control of shuffle, e.g. (Best 1998), (Lefebvre, Chevrel, and Richard 2003), (Richard, Chevrel, and Maillard 1999). A holistic analysis of driveline oscillations due to cyclic irregularity is presented in (Gosdin 1985). Here, a parameter optimisation is achieved for the vehicle driveline for predefined boundaries. Various works are concerned with mechanical, semi-active or active components reducing cyclic irregularity, e.g. (Reik, Fidlin, and Seebacher 2009). New components for palliation as well as active control algorithm for shuffle control can both profit from the presented methodology. For the former, the discussed conflict of objectives is still present and any palliation helps the effectiveness of an additional

ABSTRACT For vehicle drivetrain design, there is a serious conflict of objectives between the oscillation phenomena shuffle and cyclic irregularities. The purpose of this paper is to illustrate a methodology to analyse and visualise the sensitivities of drivetrain eigenfrequencies in order to solve this conflict of objectives for the drivetrain design process for selected vehicle drivetrain concepts. Starting with a complex and detailed non-linear model, different simplifications are performed to finally visualise the sensitivities of relevant eigenfrequencies. This provides a profound understanding of the dynamic behaviour of the vehicle and enables engineers to identify parameter combinations that solve or palliate this conflict of objectives. Two exemplary measures are derived and the palliation effect is examined with the complex simulation models. Keywords: sensitivity, vehicle transient behaviour, eigenfrequency, oscillations 1. INTRODUCTION Due to the increasing CO2 requirements of the vehicle fleet consumption, there is a considerable trend of vehicle manufacturers to optimize the vehicle fuel economy. Nevertheless, there is a serious conflict of objectives for the design of vehicle drivetrains due to the fact that vehicle dynamics and comfort aspects also have to be taken into account. Therefore, the vehicle performance needs to meet the requirements and perturbing oscillations must not exceed an admissible threshold. In order to predict the behaviour of a future vehicle, complex non-linear models are used for the simulation of different components and settings. The pure implementation of a complex model however is not sufficient to gain a parameter setup that meets the requirements. The mere amount of parameters often prevents a straight-forward approach by simply adjusting one parameter after the other to approximate a good setup.

478

drivetrain component. For the latter, improved shuffle behaviour reduces the control requirements. Starting point is a detailed model based on the physical behaviour of the different elements of the drivetrain. This model including measurement comparisons is presented in Section 2. The desired vehicle behaviour and existing conflicts of objectives are described in Section 3. In Section 4, different model simplifications are derived and the eigenfrequency sensitivity analysis is performed. The results of the presented methodology are then used to identify palliative measures for the behaviour of a three cylinder drivetrain, which is derived from the six cylinder drivetrain from Section 2. Finally, the results are summarised in Section 5.

filling and emptying approach, also the conservation of linear momentum:

t

2

( A

)

x

( p A)

(2)

Ak f

2.2. Mechanical Drivetrain Implementation For the engine crank drive model, an analytical and a multibody approach are possible implementations. A deduction of an analytical implementation according to the projective Newton-Euler equations and an evaluation of both implementation approaches is presented in (Krieg, Förg, and Ulbrich 2011). All shafts of the drivetrain are modelled as rotational inertias and springs with small damping. The mechanical model of the drivetrain follows from Figure 1. Here, ji represents the inertias of the drivetrain, ci the stiffnesses and φi the degrees of freedom. The parameters u3 and u6 represent the gear ratio and the final drive ratio. Teng represents the engine torque, which is applied to the crankshaft j2.

2.1. Thermodynamic Engine Implementation The cylinder volume is the core element of the engine model. Here, the combustion takes place and the mechanical work is transferred to the crank drive. It is basically a homogeneous volume following the first law of thermodynamics applied to open systems (Müller and Müller 2005):

Q E W

x

Here, ρ represents fluid density, A the pipe cross section area, ω the fluid velocity and p the fluid pressure. Furthermore, t represents time, x position and kf is a coefficient for friction. A solution for the partial differential equation requires a discretisation method, e.g. the finite volume method (Dick 2009). A possible implementation of the pipes is presented in (Miersch 2003). Components implemented as characteristic maps and not physically are the turbine and the compressor of the turbocharger. Here, the mass flow is estimated according to the turbocharger shaft speed and the fluid pressure of the incoming and outgoing pipes of these components, as derived from measured data. The implementation of turbochargers is described e.g. in (Baines and Fredriksson 2007).

2. DRIVETRAIN MODELLING The examined prototype vehicle is a vehicle fitted with a six cylinder turbocharged engine. First, the engine model is presented, followed by the mechanical drivetrain model. Finally a measurement comparison is illustrated in this section.

U

( A )

(1)

Here, U represents the first derivate of the internal energy with respect to time, Q the heat flow and W the mechanical power done on the system. E represents the inner energy flow of matter entering and leaving the system. The conservation of mass and the caloric theory corresponding to the cylinder gases must also be taken into account for the model. Additional components as valves, a combustion model, heat transfer elements or the crank drive are also required to model the physical behaviour of the engine. A possible implementation of these elements is described in (Krieg, Förg, and Ulbrich 2011). Deviant from the filling and emptying approach presented in (Krieg, Förg, and Ulbrich 2011), the fluid oscillations of the intake and exhaust manifold are also considered here. Intake and exhaust manifold are modelled as sequence of pipes, which are discretised homogenous volumes. Incorporating detailed intake and exhaust manifold models can increase the quality of the simulation results for certain operating conditions but leads to more complex models and longer simulation duration. The implementation of these pipes follows the conservation of mass and energy and, in addition to the

crankshaft

φ1 j1

dual-mass flywheel

φ2 c1

j2

c2

right tire

φ3 j3

j9

φ4

c3 u3

j4

φ5 c4

j5

φ6 c5

j6

c7 j7

c6 u6

left tire

φ7

Diff

Teng torsional damper

φ9

j8

φ8

j10

φ10

c8

Figure 1: Mechanical Drivetrain System Tires are implemented as elements that calculate the force between wheel and street according to the differential speed of both. Data for tires are usually measured with a dedicated tire rig and implemented via a curve fitting algorithm. The longitudinal tire force follows according to (Pacejka and Bakker 1992):

479

s

tire ((F Fload )

s

Ftire

(3)

rtire wtire vvehicle rtire wtire

input shaft and

load

F

balance of momentum of the accelerated vehicle, as illustrated in Figure 2. The variable Fgravity represents the mass force due to gravity, h represents the height of the centre of mass of the vehicle and lf and lr the horizontal distance between front and rear wheel contact point to the centre of mass of the vehicle. The balance of momentum at the front wheel contact point rear and thus the tire load Fload

according to:

Ftire h Fgravity l f

rear Fload

lf

lr

tire 2 Fload

left wout

of the right and left

2.3. Measurement Comparison In order to verify the correctness of the drivetrain model, an accurate measurement comparison is required. Therefore, a Tip-In manoeuvre was measured, i.e. the acceleration pedal of a prototype vehicle with constant speed is quickly acted from a defined part load throttle to full throttle. The first gear is engaged here. The measurement results and the corresponding simulations for the vehicle are illustrated in the following figures. Figure 3 illustrates the intake manifold pressure, which is derived from a dedicated pressure sensor. For the simulated and measured manoeuvre, the acceleration pedal is acted at t 0.5 s . The figure shows that there is good consistence for the intake manifold pressure between measurement and simulation.

of the rear tire is calculated according to the

calculates the rear axle load

and

output shaft of the differential. The parameters for all shafts and additional compliant elements are measured on component rigs. All inertias are corrected according to the gear ratio for the equivalent degree of freedom.

(4)

Here, rtire represents the tire radius, wtire the tire rotational speed and vvehicle the vehicle speed. The tire tire load

right wout

(5)

1.8 Ftire

1.4

Fgravity

Ftire

Pressure [bar]

h front

Fload

rear

lf

lr

Fload

Figure 2: Vehicle Balance of Momentum

Tin win

Toutleft 2

(6)

right left wout wout 2

(7)

Here, and

right out

T

Tin and

Furthermore,

win

1 0.8

0.4 0.2

0

1

2

3

Time [s] Figure 3: Measurement Comparison of the Intake Manifold Pressure Figure 4 illustrates the engine torque. The measured engine torque is actually derived from an engine model on the engine control unit (ECU), which estimates the current torque according to diverse measured data, e.g. crankshaft speed or the intake manifold pressure. Apparently, there is also good consistence between measured and simulated engine torque. The first rise of the engine torque is very steep and results in an excitation that is similar to a torque step function. This occurs because of the rapid filling of the intake manifold and the cylinders after increasing the throttle diameter and the subsequent conversion into mechanical work by combusting a larger mass of air

represents the torque of the input shaft left Tout

1.2

0.6

Tire force is a quantity proportional to differential speed s, so it can also be seen as non-linear damper with high damping coefficient. The function represents the curve fitting algorithm, as shown in (Pacejka and Bakker 1992). The tire models illustrated in Figure 1 also contain an inertia that represents the rims. The equations for the differential follow according to:

Toutright 2

Simulation Measurement

1.6

of the right and left output shaft.

represents the rotational speed of the

480

and fuel. The following rise of the engine torque is a consequence of the turbocharger, which has a certain time delay because of its inertia.

3. DESIRED VEHICLE BEHAVIOUR Figure 5 illustrates the vehicle longitudinal acceleration for a Tip-In manoeuvre. The excitation of the drivetrain via the engine similar to a step function results in a dominant stimulus of the first eigenfrequency with approximately 2.5 Hz for the vehicle longitudinal acceleration. In fact, also higher eigenfrequencies of the drivetrain are stimulated by the step excitation. Nevertheless, for higher frequencies less energy is brought into the system for a step function excitation, higher frequencies oscillate with smaller magnitudes and they are quickly damped due to their higher rotational velocity. These effects explain the dominance of the first eigenfrequency for the vehicle longitudinal acceleration. The desired vehicle behaviour is described from a driver perspective. For the discussed Tip-In manoeuvre, a driver experiences oscillation of the longitudinal vehicle acceleration, also referred to as shuffle, buckling or Bonanza effect, as perturbing. Particularly the height of the oscillation amplitudes is considerable to the driver. The frequency of this oscillation on the other hand is of less importance. Shuffle oscillations are usually between 2-5 Hz and in fact it is difficult for the driver to resolve a difference between these low frequencies. A quick decline of the oscillation is of higher importance. Additionally, the driver appreciates a steep rise in the vehicle acceleration curve. For a higher frequency of the first drivetrain eigenfrequency, the oscillation amplitude is lower, the damping effect is stronger and the acceleration curve is steeper. As a result, a driver prefers higher frequencies for the first eigenfrequency of the vehicle drivetrain. From this point of view, a consequence for the vehicle drivetrain design could be to choose e.g. shafts with high stiffness in order to move the first eigenfrequency to higher frequencies. This perspective however is too narrow for a drivetrain design. Driving a vehicle in other use cases also results in exciting higher drivetrain frequencies. In particular for drivetrains with three or even less cylinder engines, the second eigenfrequency is excited by the cyclic irregularity of the engine during stationary operation for low engine speeds n < 1,500 rpm in 5th or 6th gear. For higher engine speed or engines with more cylinders, the cyclic irregularity of the engine is also a problem, but the oscillations are transferred to the driver with smaller amplitudes. For these oscillation phenomena, other effects than the eigenfrequency affect the comfort perception of the driver as well. These aspects are not examined here. The focus of this paper is shuffle and cyclic irregularity due to excitation of drivetrain eigenfrequencies. Even though shuffle and cyclic irregularity are examined with different gears, they illustrate a serious conflict of objectives for the vehicle drivetrain design. A simple optimization of shuffle, e.g. choosing stiffer shafts, would result in a higher frequency of the first and second eigenfrequency of the drivetrain. A higher

450 Simulation Measurement

400

Torque [Nm]

350 300 250 200 150 100 50 0

0

1

2

3

Time [s] Figure 4: Measurement Comparison of the Engine Torque A measure comparison for the mechanical drivetrain is also required, as follows in Figure 5. Here, the drivetrain model is acted by the measured engine torque. The longitudinal acceleration of the vehicle is illustrated. The drivetrain model shows good consistence between measurements and simulation. Note that several control algorithms for the damping of oscillations within the ECU are not considered here and were switched off for the measurements to avoid masking the drivetrain behaviour with interference of these algorithm interactions for examined frequencies.

8 Simulation Measurement

7

Acceleration [m/s2 ]

6 5 4 3 2 1 0 -1

0

1

2 3 Time [s] Figure 5: Measurement Comparison of the Vehicle Acceleration

481

second eigenfrequency however is then closer or identical to the cyclic irregularity of the engine for certain vehicle speeds. Vice versa, a simple optimization of the cyclic irregularity behaviour could worsen shuffle. An efficient drivetrain design therefore must take both effects into account and resolve this conflict of objectives.

(10)

d diag( j1 , j 2 , j3 , j 4 , j5 , j6 , j7 , j8 )

M

There are only small deviations between the two models of first and fifth gear, e.g. gearing stiffness c3 or gearing ratio u3. The simplified model is illustrated in Figure 6.

4. SENSITIVITY ANALYSIS The eigenfrequency methodology is presented in this section. First, simplifications of the drivetrain model are performed, followed by the application of the sensitivity analysis based on (Dresig and Holzweißig 2010). Subsequently, palliative measures are derived and examined with the detailed model.

φ1

φ2

φ3

φ4

φ5

φ6

c7 φ7

j7 j1

c1

j2

c2

j3

c3 u3

c4

j4

j5

c5

j6

c6 u6

Diff j8

4.1. Simplifications In order to examine the eigenfrequencies, simplifications are performed here. The thermodynamic engine model is of less interest since it essentially defines the excitation, the interaction is negligible. The crank drive is simplified as part of the crankshaft inertia. An additional simplification is concerned with the tire inertia and the vehicle mass. In relation to the drivetrain inertia, the vehicle mass is very high. Therefore the error caused by attaching the vehicle mass to the inertial frame is small. Furthermore, the tires are neglected and the rims are also fixed to the inertial frame. Due to the fact that there is generally only low damping for the drivetrain, damping effects are neglected for the sensitivity analysis. These simplifications actually influence the eigenfrequencies of the drivetrain and therefore a final comparison of the simplified model with the detailed drivetrain model is required. The simplified drivetrain model is a one dimensional chain of inertias and springs:

M x C x 0

Figure 6: Reduced Mechanical Drivetrain Model 4.2. Sensitivity Algorithm The presented sensitivity algorithm is based on (Dresig and Holzweißig 2010). A modal decomposition of the differential equation system of Equation (8) returns the modal inertia as follows i and stiffness i (Ulbrich 1996):

C

c1 c2 c2

0

0

0

0

0

0

0

c2 c 2 c3

0

0

0

0

0

u 3 c3

0

0

0

0

2 u 3 c3 u 3 c3 c 4 c4 c4 c 4 c5 0

0

0 u

0

0

0

0

c5

0

0

0

0

0

0

0

0

0

0

c5 c 6

6 c6 2

6 c6 2

(12)

M vi

vi represents the eigenvector of the

i

viT C vi

i

viT M vi

i

is then

(13)

vi 0 :

is negligible ~2

2 i

i

2 i

viT (C

C ) vi

viT

M ) vi

(M

(14)

0 u

6 c6 2 u2 6 c6 4

Assuming that variations of the inertias are much smaller than the inertias themselves M M , and can be estimated after some transposition i

(9)

u2 6 c6 c7 4 u u2 u2 6 6 6 c6 c6 c6 c8 4 4 2

u

i

Small variations of the inertias M or the stiffnesses C lead to small variations of the eigenfrequency i , and the change of the eigenvector

0

0

c5

(11)

viT

2 i

0

0

viT C vi

eigenfrequency i . The eigenfrequency calculated according to:

(8)

0

i

Here,

Here, x represents the vector for the rotational positions and x the vector for the rotational accelerations of the inertias. The matrix of stiffness C follows according to: c1 c1 c1 0

φ8

c8

according to (Dresig and Holzweißig 2010): 2 i

The matrix of inertia M is diagonal according to:

482

2 i

viT viT

C vi C vi

v iT

M vi

viT

M vi

(15)

60

The variation matrices for stiffness C and inertias M are a sum of the variations of all inertias j l and stiffnesses c k : M l0

jl jl

(16)

Ck0

ck ck

(17)

l

C k

rel. Sensitivity [%]

M

50

The matrices C k 0 and M l 0 represent the introduced matrices for stiffness and inertias with all elements equal to zero, except element k or l respectively, e.g. for the inertias:

Ml0

diag( d 0,  , 0, jl , 0,  , 0)

2 i

ki k

viT ki

C k 0 vi

viT C vi viT M l 0 vi

li

viT M vi

20

0

(18)

of element k or l respectively:

30

ck ck

25

li l

jl jl

1

2

3

4 5 6 Inertia ji

7

8

Figure 7: Sensitivity of Inertias for Shuffle

(19)

rel. Sensitivity [%]

2 i

i

30

10

This finally leads to the following sensitivity coefficients for stiffness ki and inertia li for eigenfrequency

40

(20) (21)

These coefficients describe the variation of the eigenfrequency for a small relative parameter variation. Thus they are regarded as sensitivity coefficients for that eigenfrequency.

20 15 10 5 0

4.3. Drivetrain Eigenfrequency Analysis For the presented drivetrain models for the first eigenfrequency of the first gear and the second eigenfrequency of the fifth gear, the sensitivity coefficients are illustrated in the following figures. In order to solve the discussed conflict of objectives, the task now is to move the first eigenfrequency of the first gear to higher frequencies and vice versa move the second eigenfrequency of the fifth gear to lower frequencies. For a spring and mass system, an eigenfrequency is moved to higher frequencies by decreasing inertias or increasing stiffnesses, compare Equation (19). Figure 7 illustrates the sensitivity coefficients of the inertias and Figure 8 those of the stiffnesses for shuffle. The shuffle eigenfrequency is 1 2.5 Hz .

1

2

3

4 5 6 Stiffness c i

7

8

Figure 8: Sensitivity of Stiffnesses for Shuffle For the shuffle eigenfrequency, there is a high sensitivity for the crankshaft inertia and the inertia of the dual-mass flywheel j2 and j3. Furthermore, there is a high sensitivity for the rubber joint and the sideshaft stiffness c4, c7 and c8. The same sensitivity analysis was performed for the relevant eigenfrequency for cyclic irregularity 17.1 Hz , as illustrated in Figure 9 for the inertias and 2 1 Figure 10 for the stiffnesses.

483

80

principle. From Figure 8 and Figure 10 it is apparent that the stiffness c7, i.e. the stiffness of the right sideshaft, has a major influence on the shuffle eigenfrequency and a minor influence on the eigenfrequency of cyclic irregularity. On the other hand, stiffness c2, i.e. the stiffness of the dual-mass flywheel, has a minor influence on the shuffle eigenfrequency and a major influence on the eigenfrequency of cyclic irregularity. A promising combination to solve the conflict of objectives could now be to increase stiffness c7 and decrease stiffness c2. In order to contain the symmetry of the drivetrain, stiffness c8, i.e. the stiffness of the left sideshaft, also needs to be modified according to stiffness c7. Here, the stiffnesses of the driveshafts are doubled and the stiffness of the dual-mass flywheel decreased to half of the original value. This results in modification 1:

rel. Sensitivity [%]

70 60 50 40 30 20 10 0

1

2

3

4 5 6 Inertia ji

7

8

Figure 9: Sensitivity of Inertias for Cyclic Irregularity

70

c2mod1

0.5 c2

(22a)

c7mod1

2c7 2

(22b)

mod1 8

mod1 7

c

(22c)

c

rel. Sensitivity [%]

60 Additionally, another modification is examined to ensure that modification 1 is not a coincidence. The new parameter setup is referred to as modification 2:

50 40 30 20

c4mod 2

2 c4

(23a)

c

0.8 c7

(23b)

c

c7mod 2

(23c)

mod 2 7 mod 2 8

Excitation 1.5 Order

10 0

1st & 2nd EF Original 1

2

3

4 5 6 Stiffness c i

7

1st & 2nd EF Modification 1

8

1st & 2nd EF Modification 2

Figure 10: Sensitivity of Stiffnesses for Cyclic Irregularity

40 35

Figure 9 illustrates that there is a high sensitivity of the eigenfrequency for cyclic irregularity for the secondary mass of the dual-mass flywheel j3. Furthermore, the stiffness of the dual-mass flywheel c2 also shows high sensitivity for that eigenfrequency. With the presented diagrams, it is now possible to identify combinations that raise the first eigenfrequency and decrease the second eigenfrequency as well. In order to derive a parameter combination as palliation measure, a first parameter is selected from the illustrations, which has a high sensitivity for shuffle and a low sensitivity for cyclic irregularity. Additionally, a second parameter is chosen with low sensitivity for shuffle and high sensitivity for cyclic irregularity.

Frequency [Hz]

30 25 20 15 10 5 0

4.4. Modifications of the Mechanical Drivetrain Two different combinations will be discussed in the following section in order to illustrate the methodology

0

500 1000 Engine Speed [rpm] Figure 11: Campbell Diagram

484

1500

The values for the parameter variation are derived so that the effect of one measure does not overcompensate the effect of the other. The Campbell diagram in Figure 11 shows the result of the drivetrain eigenfrequencies (EF) for both modifications. The shuffle eigenfrequency of modification 1 is mo mod1 2.8 Hz , the eigenfrequency for cyclic 1 mod mo 1 2

350 300

Torque [Nm]

irregularity is

400

1 .5 Hz . Furthermore, the shuffle 15

eigenfrequency of modification 2 is

mod mo 2 1

2.6 Hz , the mod mo 2 2

eigenfrequency for cyclic irregularity is 116.9 Hz . Obviously, the aimed task to move both eigenfrequencies in the desired directions worked for both modifications. A weakness of the described methodology is the difficulty in predicting the precise frequency of the eigenfrequency of the new setup. Hence, the methodology helps to identify, which parameters should be modified in which direction, but not how high the parameter variation should be.

250 200 150 100 6 Cylinder Engine 3 Cylinder Engine

50 0

2 3 Time [s] Figure 12: Engine Torque for Original and 3 Cylinder Engine

4.5. Simulative Measure Verification Due to the performed simplifications it is finally required that the results also persist for the detailed model. The eigenfrequency of cyclic irregularity as examined here for low frequencies is a problem for the vehicle drivetrain because it is close to the excitation of the engine. For a three cylinder engine, it is particularly difficult since it excites the drivetrain with the 1.5 order. The excitation of a three cylinder engine is also illustrated in Figure 11. A detailed three cylinder model is derived in this work from the presented six cylinder model of section 2. Therefore, three cylinders are removed and additional modifications for the exhaust and the intake manifold are performed to generate the three cylinder model. In particular, parts of the exhaust and intake manifold system with two parallel paths of the six cylinder model are removed. Those parts with one common path are physically divided in half, e.g. the throttle cross section or the intake manifold volume. This model is used to obtain a realistic engine excitation. Figure 12 illustrates the engine torque of the model for a Tip-In manoeuvre. This diagram shows that the three cylinder engine has a comparable steepness of the first torque raise which is followed by an engine torque which is approximately half as high as the six cylinder engine. In Figure 13 the acceleration of the vehicle model is illustrated. Here, the original drivetrain model is excited with the three cylinder torque of Figure 12. Furthermore, the drivetrain model modification 1 is also excited by this engine torque. First remark is that due to the lower torque the absolute values of the acceleration and its oscillations are lower than those of the drivetrain of Section 2. Nevertheless, since the oscillations are observed in relation to the mean acceleration, the oscillations are still a problem.

0

1

Additionally, Figure 13 shows that the aimed increase of the shuffle eigenfrequency worked for modification 1. Consequently, also the amplitude of the oscillations is reduced and the suggested measure worked for the simulation of shuffle as expected.

4 Original Modification 1

Acceleration [m/s2 ]

3.5 3 2.5 2 1.5 1 0.5 0

0

1

2

3

Time [s] Figure 13: Vehicle Acceleration with Modification 1 Next step is the simulation of cyclic irregularity. Here, the three cylinder engine is again used to generate the excitation. For cyclic irregularity, a constant excitation for a dedicated engine speed is required. According to Figure 11, a low engine speed results in an excitation close to the eigenfrequency. For that reason n = 1,000 rpm is used as excitation. Figure 14 illustrates the torque acting on a distinct drivetrain shaft, which is relevant for cyclic irregularity, for the original drivetrain and the drivetrain modification 1 in order to evaluate oscillations due to cyclic irregularity.

485

Original Modification 2

300

300

250

250

200

200

Torque [Nm]

Torque [Nm]

Original Modification 1

150 100 50 0

150 100 50

0

0

0.2

0.4 0.6 0.8 1 Time [s] Figure 14: Cyclic Irregularity for n = 1,000 rpm, Modification 1

0.4 0.6 0.8 1 Time [s] Figure 16: Cyclic Irregularity for n = 1,000 rpm, Modification 2

The illustration shows that the amplitude of the torque oscillation of the modified drivetrain decreased from 70 Nm to 30 Nm. Thus, the exemplary system modification worked in the desired way and helped to palliate the conflict of objectives between shuffle and cyclic irregularity. For modification 2, the drivetrain was excited in the same way as described for modification 1 above. Figure 15 and Figure 16 illustrate the simulation results for drivetrain modification 2.

The vehicle acceleration in Figure 15 shows that the effect of modification 2 for shuffle is marginal. The mod 2 . The reason for that is the minimal change of 1mo decrease of the first stiffness almost completely compensates the increase of the second stiffness. Improved behaviour is observable for cyclic irregularity in Figure 16. The effect of modification 2 is smaller than that of modification 1.

Original Modification 2

Acceleration [m/s2 ]

3 2.5 2 1.5 1 0.5 0

0

1

2

0.2

5. CONCLUSION For vehicle drivetrain design, oscillations represent a major aspect in order to gain a setup that meets the comfort requirements of the driver. Shuffle and cyclic irregularities due to the engine excitation cause a serious conflict of objectives. A methodology is presented that can help to solve this conflict of objectives for selected drivetrain setups, based on an understanding of the drivetrain eigenfrequency sensitivities. First, a detailed model of the drivetrain is presented in Section 2, containing a complex engine model for the vehicle excitation and a mechanical drivetrain model including tires. A measurement comparison for the engine torque, intake manifold pressure and the vehicle acceleration is presented to verify the correctness of the models. In Section 3, the desired vehicle behaviour and the conflict of objectives for the palliation of shuffle and cyclic irregularity is described. The methodology is presented in Section 4, starting with a simplification of the drivetrain model and an eigenfrequency sensitivity analysis based on (Dresig and Holzweißig 2010). The following visualisation facilitates the identification of palliation measures for the described conflict of objectives. Two exemplary palliation measures are

4 3.5

0

3

Time [s] Figure 15: Vehicle Acceleration with Modification 2

486

Dick, E., 2009. Introduction to Finite Volume Methods in Computational Fluid Dynamics. In: Wendt, J. F., ed. Computational Fluid Dynamics. Springer, Berlin: pp. 87-103. Dresig, H., Holzweißig, F., 2010. Dynamics of Machinery: Theory and Application. Berlin: Springer. Gosdin, M., 1985. Analyse und Optimierung des dynamischen Verhaltens eines PkwAntriebsstranges. Düsseldorf: VDI. Hülsmann, A., 2007. Methodenentwicklung zur virtuellen Auslegung von Lastwechselphänomenen in PKW. Munich: Technische Universität München, Dissertation. Krieg, O.M., Förg, M., Ulbrich, H., 2011. Simulation of Domain-Coupled Multibody and Thermodynamic Problems for Automotive Applications. Proceedings of Multibody Dynamics 2011. July 47, Brussels, Belgium. Lefebvre, D., Chevrel, P., Richard, S., 2003. An HInfinity-Based Control Design Methodology Dedicated to the Active Control of Vehicle Longitudinal Oscillations. IEEE Transactions On Control Systems Technology, Volume 11, No. 6: pp. 948-956. Miersch, J., 2003. Transiente Simulation zur Bewertung von ottomotorischen Konzepten. Munich: Hieronymus. Müller, I., Müller, W.H., 2005. Fundamentals of Thermodynamics and Applications. Berlin: Springer. Pacejka, H.B., Bakker, E., 1992. The Magic Formula Tyre Model. Vehicle System Dynamics: International Journal of Vehicle Mechanics and Mobility, Volume 21, Issue S1: pp. 1-18. Reik, W., Fidlin, A., Seebacher, R. 2009. Gute Schwingung – Böse Schwingung: Proceedings of 6. VDI-Fachtagung Schwingungen in Antrieben: pp. 3-15. Richard, S., Chevrel, P., Maillard, B. 1999. Active Control of future vehicles drivelines. Proceedings of the 38th Conference on Decision & Control: pp. 3752-3757. Ulbrich, H., 1996. Maschinendynamik. Wiesbaden: Teubner.

derived from these illustrations. In order to show the conflict of objectives, a three cylinder engine is derived from the examined models, which particularly suffers from the described conflict, since the excitation frequency is close to the second drivetrain eigenfrequency for low engine speed. The setup of the mechanical drivetrain is not changed any further to illustrate the methodology principle. Simulations for the detailed models show that the suggested, exemplary measures can palliate the described conflict of objectives for the observed drivetrain. The example shows improved vehicle behaviour for the performed time-based simulations. Nevertheless, a final measurement comparison is required. It must eventually be clarified for other conditions of use that the suggested measures will not worsen the vehicle behaviour for those use cases. This can also include oscillations of cyclic irregularity for higher frequencies since other effects affect the driver perception here as well. It shall also be remarked that the described method is not the usual method to derive a drivetrain setup. In particular, the vehicle mass of the three cylinder engine remained identical to the six cylinder engine. Also the gear and final drive ratios remained the same. For a drivetrain design in the industry, the gear and final drive ratio, the vehicle mass and additional parameters are adjusted to achieve a coherent vehicle setup. In this paper, these adoptions are neglected in order to demonstrate the methodology principle. It is the dedicated objective of the presented method to palliate the conflict of objectives between shuffle and cyclic irregularity and not to derive a drivetrain setup for new vehicle concepts. Despite to an optimisation algorithm, the comparison of the sensitivities for the different use cases provides a profound understanding of the dynamic vehicle drivetrain behaviour. Furthermore, optimisation algorithms require distinct optimisation criteria and boundaries that often do not exist explicitly. In particular, for the heterogeneous design process of a vehicle with different responsibilities spread around research and development departments, a methodology providing a profound understanding of the dynamic behaviour is superior compared to a singular optimum. REFERENCES Baines, N., Fredriksson, C., 2007. The Simulation of Turbocharger Performance for Engine Matching. In: Pucher, ed., H., Kahrstedt, J., ed. Motorprozesssimulation und Aufladung 2: Engine Process Simulation and Supercharging. Expert, Berlin: pp. 101-111. Bencker, R., 1998. Simulationstechnische und experimentelle Untersuchung von Lastwechselphänomenen an Fahrzeugen mit Standardantrieb. Munich: Hieronymus. Best, M.C., 1998. Nonlinear optimal control of vehicle driveline vibrations. UKACC International Conference on Control ´98 (Conf. Publ. No. 455).

487

IMPROVING JOB SCHEDULING ON A HETEROGENEOUS CLUSTER BY PREDICTING JOB EXECUTION TIMES USING HEURISTICS Hannes Brandstätter-Müller(a) , Bahram Parsapour(b) , Andreas Hölzlwimmer(c) , Gerald Lirk(d) , Peter Kulczycki(e) (a, b, c, d, e) Upper

Austria University of Applied Sciences School of Informatics, Communication, and Media Department of Bioinformatics

(a) hannes.brandstaetter@. . . (b) bahram.parsapour@. . . (c) andreas.hoelzlwimmer@. . . (d) gerald.lirk@. . . (e) [email protected]

ABSTRACT In this paper, we propose the scheduling system for the Bioinformatics Resource Facility Hagenberg (BiRFH). This system takes advantage of the fact that the facility offers tailored solutions for the customers, which includes having a limited amount of different programs available. Additionally, the BiRFH system provides access to different hardware platforms (standard CPU, GPGPU on NVIDIA Cuda, and IMB Cell on Sony Playstation machines) with multiple versions of the same algorithm optimized for these platforms. The BiRFH scheduling system takes these into account and uses knowledge about past runs and run times to predict the expected run time of a job. That leads to a better scheduling and resource usage. The prediction and scheduling use heuristic and artificial intelligence methods to achieve acceptable results. The paper presents the proposed prediction method as well as an overview of the scheduling algorithm.

of algorithms with different input data. This allows runtime prediction using heuristics and therefore an improved deadline scheduling. The “Bioinformatics Resource Facility Hagenberg” is a resource management and scheduling system targeting the needs of microbiology and bioinformatics related high performance computing. The main features are: (1) integration and management of different hardware platforms, (2) scheduling jobs so that the available hardware is used, and (3) making the management and job creation more accessible to non-technical users. To fully enable the necessary features, and supported by the fact that the BiRFH service is designed to be used with a limited number of algorithms rather than allowing arbitrary code to be uploaded and executed like on many other standard compute clusters, BiRFH requires the algorithms to be adapted and to support some defined methods. The BiRFH framework seeks to mitigate the necessary development effort. The system is based on a framework that is to be included in each algorithm if possible, i. e., if the source code is available for modification. Including the framework directly into the source allows the use of more advanced features and more control over the program. Should the source for a program be unavailable, the framework also supports the creation of wrapper programs that can in turn execute the desired program. Regardless of how the framework is applied, it allows the use of the best available hardware for the selected algorithms with the trade-off of higher development effort to enable the algorithm on as many hardware platforms as possible. As a further feature, BiRFH uses heuristic scheduling and resource management algorithms in order to optimize the cluster’s throughput. There are many scheduling systems and resource managers for high-performance cluster computing available (see Table 1). Most of them are designed for uniform hardware. Almost no system allows the coupling of compute platforms having different processor architectures in a way that e. g. allows the migration of a running algorithm from one type of platform to a different one. BiRFH offers this possibility for algorithms with available source

Keywords: algorithms, bioinformatics, high performance computing, molecular biology 1. INTRODUCTION Scheduling and resource management are fundamental tasks when running a high performance computing system. Resource management and scheduling systems for different processor technologies and architectures in a single cluster are not very common although they offer great possibilities to the user. Our system software “Bioinformatics Resource Facility Hagenberg” (BiRFH) allows efficient control and management of the so-called “MetaHeterogeneous Cluster”. BiRFH allows one not only to drive classic heterogeneous clusters (i. e., systems that comprise nodes that vary just in CPU speed and RAM size), it allows one to integrate and operate different processor architectures simultaneously. Our resource facility currently consists of standard Intel CPUs (Intel 2005), NVIDIA GPUs (NVIDIA 2010), and IBM Cell Broadband Engines (IBM 2006). There are many strategies available for scheduling jobs on a cluster. We focus on the special case where jobs are based on a small number 488

2. DATA MINING AND HEURISTICS The terms makespan and flowtime (Pinedo 2008) are commonly used when classifying the success of the output of a scheduler. More advanced scheduling techniques, i. e., most scheduling methods beyond simple load balancing, require a knowledge of the expected time to complete a task. Some solutions (Xhafa and Abraham 2008) require an estimation by the user, which is sometimes too complicated a task for non-IT scientists, and on the other hand does not account for different versions of an algorithm optimized for the available heterogeneous hardware. Therefore, the BiRFH approach does not require run time estimations by the user. Most algorithms exhibit some form of correlation between the input data, other given parameters and the time the program needs to run. The BiRFH system gathers performance data on the various algorithm implementations as they are executed. The collected data consists of the runtime measurements, i. e., how long the execution of the task took in real-time. Additionally, the consumed CPU time is also measured. Should the hardware be under-subscribed, algorithms can be executed during this idle time to generate additional measurements, especially for combinations of parameters that are expected to complete “holes” in the available data. This performance data, combined with data about the input and other parameters, is used by machine learning algorithms to predict the temporal behavior of subsequent runs, especially for new parameter combinations or input data. As the parameter values, especially the input data, usually consist of file names or other non-numeric values, it can not be used directly for the prediction of the expected runtime. Therefore, the BiRFH framework calls a method that is expected to be provided for each algorithm. This method should provide a numeric value that represents each non-numeric parameter value. If, for example, one algorithm parameter is an input file name, the method produces a number signifying the “weight”, or expected impact on the runtime, of the contained data. Simple file size is sometimes not sufficient to use when predicting the impact on run time, e. g. when processing a file containing sequences in FASTA format, sometimes the number of sequences, other times the average or maximum length of the sequences is more significant. The amount of computation time to produce this number should be kept within a reasonable time frame, but is left to the algorithm developer to decide. If a simple file size is not sufficient, then partial, sampling or even a total file analysis should be done. This is only appropriate if the computational run time is very long compared to the time needed to load the data from the file, and not just slightly longer than the full analysis of the input file itself would take. The next chapter, 2.1 Heuristics, contains an example dataset and shows how the prediction works.

SLURM

Maui









Cycle Scavenging



Heterogeneous Platforms



Priority Based Scheduling



Hibernation Support

()

Resource Based Scheduling

()





()













Advanced Resource Reservation Topology Awareness

BiRFH

SGE

Workload Manager

MPI2

Feature

Condor

code by implementing a data exchange that can also be used for hibernation, i. e., the freeing of used system resources by writing the current state of the calculations to the hard drive. There are some other approaches to enable heterogeneity for compute-intensive applications. These include OpenCL (Munshi 2011) and C++ Accelerated Massive Parallelism (AMP) (Sutter 2011; Moth 2011). These focus on enabling single applications flexible access to any available computing device rather than distributing instances of algorithms over the available hardware. The BiRFH approach focuses more on user guidance and supporting the storage of data on the remote compute system. It is notable, however, that Microsoft’s C++ AMP moves the definition of heterogeneity more in the direction of heterogeneous platforms than the previously common heterogeneous systems, i. e., including different processor architectures. BiRFH focuses on algorithms and computations for biomolecular and bioinformatics applications. The reasons for focusing especially on bioinformatics lie (1) in the near-exponential growth of available data in the currently booming field (Howe et al. 2008), (2) in the demands for making high performance computing available to nontechnical users and (3) the availability of bioinformatics knowledge in the project team. Moreover, the BiRFH system supports a scheduling mechanism that is (1) based on the temporal behavior of algorithms and (2) also based on the size and the inner structure of input data to be processed.











Table 1: Some features of well-established resource managers and scheduling systems, in addition to the BiRFH system, from (Hoelzlwimmer 2010).

2.1. Heuristics The machine learning algorithms that provide the best results on the current training data sets are Artificial Neural 489

Networks (ANNs). These are very versatile and produce accurate predictions for the available data. Other evaluated machine learning algorithms are the regression algorithms offered by the Weka toolkit (Hall et al. 2009). To evaluate the available regression algorithms, a sample data set with 2400 measurement instances was created. This data set consists of measurements of a simple algorithm that reads a file containing several FASTA formatted sequences. Then, some string operations are performed and an output file is written. This algorithm is executed with three different input files (one with 500 sequences, one with 1000 and another one with 1500) and also with different parameters influencing the string operations. The collected data can be presented as CSV:

the assumption that a regression method can most likely predict the run times accurately from this data set. Weka yielded the following results with 10-fold cross validation: Classifier

wall,InSize,InCount,MaxMut,WinSize,GC-Cont 2288.74,89725,500,1,5,0 3279.26,89725,500,1,5,0.8 2540.13,181156,1000,1,5,0 4627.02,564117,1500,1,5,0.7 5089.33,564117,1500,5,16,0.6 ...

CC

MAE

RMS

RAE

MultilayerPerc.

0.9256

280.06

358.72

50.3 %

LinearRegr.

0.8714

385.02

464.75

69.7 %

IsotonicRegr.

0.7620

502.95

613.38

65.8 %

SimpleLinRegr.

0.6626

551.43

709.47

72.2 %

PACERegr.

0.8713

385.04

464.79

50.4 %

LibSVM

0.8452

403.02

507.48

52.7 %

Table 2: Results in the Weka Toolkit for various prediction methods: MultilayerPerceptron, LinearRegression, IsotonicRegression, SimpleLinearRegression, PACERegression, LibSVM (CC : Correlation Coefficient, MAE : Mean absolute error, RMS : Root mean squared error, RAE : Relative absolute error)

Wall represents the wall clock time used to complete this algorithm. InSize is the size of the input file, InCount is the number of FASTA sequences, MaxMut, WinSize and GC-Cont are parameters that influence how the input is processed and may or may not have an influence on the run time. These is the data available after the algorithm runs have finished, and as mentioned in the chapter above, some of these data points are available before the real computation is started. In this case, everything except the wall clock time and the InCount is available before the algorithm is started for the calculation run. Getting the file size is not as compute intensive and can substitute the exact count of input sequences for this case.

Taking one of the validation results to better visualize the resulting prediction quality. 5000

run time ms

4000 3000 2000 1000 0 1000 0

50

100

150

200

sample number 400

Figure 2: Actual and predicted values of a validation run using MultilayerPerceptron

count

300

200

0.2

100

0.1

0 2500

3000

3500

4000

4500

5000

5500

relative error

2000

run time ms

Figure 1: Run times of 2400 calls to the same algorithm with different input data and parameters

0.0

0.1

0.2

Looking at the run times gathered, see Figure 1, it is easy to spot three distinct peaks which indicate the run time effect of the two different input file sizes. There is also a third peak in the upper run time regions that is most likely caused by one of the parameters. The remaining variance due to other parameter variations. This reinforces

0

50

100

150

200

sample number

Figure 3: Relative errors of the validation run using MultilayerPerceptron

490

3. SCHEDULING The task of scheduling and resource management working together is to ensure that computation tasks are completed in a timely manner and that the available hardware is used as efficient as possible. There are many strategies available to fulfill these requirements. As mentioned before in chapter 2, makespan and flowtime are used as indicators for the quality of a scheduling attempt. Having a reliable prediction for the expected run time enables more advanced scheduling features. Earliest Deadline First (EDF) is one of the most popular and widely used scheduling strategies, and has some parts in common to the BiRFH approach. The EDF strategy requires tasks to have a deadline, usually either the latest possible start time or the latest acceptable finishing time. The scheduler then orders the tasks by their deadlines and executes those with the earliest deadline first. Usually resource managers can also interrupt running tasks should a new task with an earlier deadline become available (Kruk et al. 2007).

Using libSVM, the results are similar but with a higher margin of error. Figures 4 and 5 show the best result achieved by tweaking the libSVM parameters “cost”, “gamma” and “epsilon” with epsilon-SVR in a similar fashion as above with MultilayerPerceptron.

5000

run time ms

4000 3000 2000 1000 0 1000 0

50

100

150

200

sample number

Figure 4: Actual and predicted values of a validation run using libSVM

start

0.2

relative error

wait

are there any jobs to schedule ?

no

0.0

yes 0.2

0.4

get run-time predictions 0

50

100

150

all run-time predictions vailable ?

no

200

sample number

yes

Figure 5: Relative errors of the validation run using libSVM

create initial schedule

W

Standard ANN training algorithms usually require the basic structure to be predefined, i. e., not only the number of inputs and outputs, but also the number of internal layers, the nodes per layer and how the nodes are connected. This would prevent maximal flexibility of the training for the expected diversity of the dataset, as the complexity is an important factor in the success of the neural network training: too little complexity, and the solution quality suffers, but too much complexity would lead to over-fitting during the training. To work around this limitation and make the ANN approach viable for a multitude of different algorithms, training algorithms that start with an empty network and add complexity as needed to reach a neural network for the given training set are used in the BiRFH system. The FANN library (Nissen 2003), which was chosen as ANN implementation, provides such a training algorithm called Cascade2 (Nissen 2007).

wait for event and switch on kind of event new job

resource gain

resource loss

job ended

job terminated

N

G

L

E

T

Figure 6: A simplified flowchart of the scheduling system

The basic program flow in the scheduling logic is displayed in figure 6. After an initial check for pending jobs, and calculating the run time estimation for these, the scheduler distributes these jobs among the available resources according to the scheduling policy. Then, the scheduler waits for one of the following events to occur: 491

(N) new jobs are submitted to be scheduled and executed, (G) an additional machine comes online and is available to execute jobs, (L) a currently available machine goes offline, (E) a currently running job ends or a running job terminates unexpected (T). The scheduler then handles these events accordingly and returns to the waiting state. Some paths lead to a rescheduling, others do not require a reordering. Depending on the circumstances, the reordering can affect the currently running processes as well.

already is handled in the third path. This path assumes that the resource is new or has been offline long enough that it can be regarded as new. All pending jobs are rescheduled, as are all currently running jobs if there is a significant reduction in overall time to completion. The next path, (L), deals with the sudden, unplanned loss of a resource: all currently running and pending jobs of this resource are rescheduled over the remaining resources. Should the resource come back online in time, i. e., before the “replacement job” is completed, and the job on that resource is healthy (i. e., was not broken by the loss of connectivity), the now duplicate backup job is canceled and the pending jobs are rescheduled.

N

get run-time predictions

no

all run-time predictions vailable ?

L

reschedule lost jobs

yes priority of new job low

high

schedule jobs without interfering with existing ones

reschedule, stop, or migrate running jobs

resource avilable in time ? yes

no

discard duplicate jobs, reschedule pending jobs

W

Figure 7: Scheduling for newly submitted jobs W

Figure 9: The scheduler handles the unexpected loss of a resource

In the path (N) new jobs are submitted, these are again run through the prediction. After the predictions are available for all new tasks, these are scheduled according to their priority. If their priority is default normal, then no special additional steps have to be performed and the tasks are scheduled without interfering with currently running tasks. Should the new tasks have high priority, the currently running tasks are included in the rescheduling, which could lead to low priority tasks to be paused, migrated or even aborted and restarted at a later point. G

T

reschedule all pending jobs

reschedule, stop, or migrate running jobs

W

W

E

ended sooner/later than expected ? yes

no

reschedule, stop, or migrate running jobs

W

Figure 8: Scheduling for the case of the gain of a resource (left) or an erroneous job termination (right)

Figure 10: Scheduling when a job ends without error

Path (G) (Figure 8, left) is caused by new hardware resources coming available for the system. The special case where these resource has been known to the system

The remaining two paths are triggered every time a job ends. If the job ends cleanly (E), the run time is compared to the estimate, if the difference is below the 492

threshold, no further scheduling is needed. If the difference exceeds the threshold, a scheduling run is performed to ensure optimal resource usage. If the job did not terminate cleanly (T) (see Figure 8), subsequent dependent runs are removed from the queue and an error notification is sent to the originating user. The remaining jobs, including the running ones, are then rescheduled again. Under one special condition, in the case where two algorithms are running simultaneously and one algorithm produces output that is immediately used by the second algorithm— called streaming read/write—the consuming job has to be terminated if the producing job experiences an error and terminates. The scheduling algorithm itself is designed to be fully modular. It reads the information about the pending jobs as well as the current system status from a database, and after reordering the jobs according to the scheduling rules and policies, writes the new orders back to the database, where the resource manager reads them and relays the orders to the compute nodes. That way, different scheduling strategies and optimizations can be evaluated without big changes in the whole BiRFH system. FCFS

FCFS

BiRFH

R1

R2

R3

R1

R2

R3

T1

T2

T3

T1

T5

T3

T4

to finish and task T5 is the longest with 4 time units. The simple FCFS scheduling on the left produces an execution sequence that would take 7 time units. Knowledge of the expected run times can rearrange the tasks in such a way that the total execution time would be 5 time units. The above example is based on the assumption of three completely identical compute nodes. Our scheduling approach aims to improve this strategy for MetaHeterogeneous Clusters, i. e., clusters consisting of different hardware architectures like CPU, GPUs and others, by considering multiple versions on the different hardware platforms and deferring some algorithms expecting a more powerful platform to become available. This is enabled by the run time predictions of the heuristic analysis.

R1

R2

R1

T1

T2

T1

R2

T2

T2 T5

BiRFH

Figure 12: Advantage of knowledge of runtimes with two different hardware platforms

T4

A simple scenario (see Figure 12) where this would be of benefit: Task T2 is a task that finishes fastest on the first system, e. g. a GPU, but the GPU is currently occupied by task T1. Using a standard deadline first algorithm, this task would now run on the second available system, e. g.the CPU. If the R1 node is expected to be ready before the task on the R2 system would be finished, and the sum of the expected calculation for both tasks on the R1 system would put the finishing time earlier, the task would be put on hold instead of using the free CPU. Even if not all the tasks are known from the beginning, knowledge of the runtimes enables better scheduling in many cases. Figure 13 shows an example: the compute nodes R1 and R2 are again different hardware platforms, which results in different runtimes for the tasks. The FCFS version also displays a possible variant of the basic FCFS strategy, where tasks only are scheduled to hardware platforms where they take the least time to finish. This is shown with T3 (transparent on the lower left). Naively, this should yield better results, but even in this example, using all the available resources still provides better results. The BiRFH result on the right again shows optimized resource usage with reduced overall runtime.

Figure 11: A simple scenario where knowledge of run times yield better scheduling than a basic First Come First Serve

Knowing about the expected runtime can improve the utilization and therefore reduce the overall time needed to complete several tasks. Figure 11 displays a hypothetical comparison of a simple EDF variant, where the tasks are prioritized in the order that they become available, i. e., a First Come First Serve scheduling. In this scenario, 5 tasks with different runtimes are available: Tasks T1 and T4 need 2 time units, tasks T2 and T3 take 3 time units 493

FCFS R1

R2

R1

R2

T1

T2

T1

T2

T3 T4

can be adapted to some kind of stepwise progress, the framework defines the methods to pause and serialize the current memory content to the hard disk for transfer. The counterpart implementation on the target hardware then can read the progress and continue with the calculations. All the above examples target the least “real world” run time, or minimal makespan, for optimization. Other possible targets, especially when using heterogeneous hardware, could be the conservation of power. If a deadline is set and multiple hardware platforms are able to execute the task, the job could run on a platform that consumes less power during the calculations. The power consumption factor, combined with the run time, can be added to the optimization parameters. As the overall design of the system allows for the individual models to be easily interchangeable, the prediction model and scheduling methods can gradually be replaced by new ones that yield better results. The underlying database offers a simple interface that enables this modularity. Therefore, future development also encompasses the implementation and evaluation of different prediction models and scheduling strategies.

BiRFH

T4 T3

T3

Figure 13: Prioritizing algorithm versions on heterogeneous platforms with minimal makespan in the FCFS strategy does not yield better results Some scheduling scenarios require a complete rearrangement of the tasks, even running tasks could be blocking an optimal arrangement. In these cases, there has to be an assessment if the running task should be canceled and moved to a different hardware or if it would be better to wait for the task to finish. Knowledge of the expected finishing time can be very helpful in these cases. Figure 14 visualizes a scenario where a new task T3 is added after the tasks T1 and T2 have already been started. This task is only available on the hardware platform of R2, but this system is being used by T2. Therefore, T2 is canceled and restarted on node R1 after T1 has finished. Even though some computational effort is lost, the overall time is significantly shorter than waiting for T2 to finish. before

4. CONCLUSION AND FUTURE WORK The use of heuristics can improve the scheduling quality given some special circumstances, like, in our case, the limited number of different algorithms and the degree of heterogeneity of the hardware. Future work includes the implementation and evaluation of other computational intelligence strategies as well as the inclusion of the pipeline concept, i. e., the consideration of follow-up calculations or calculations using different algorithms based on the same input data, which are required by follow-up calculations. The research project “Bioinformatics Resource Facility Hagenberg” is supported by grant #821037 from the Austrian Research Promotion Agency (FFG).

after

R1

R2

R1

R2

T1

T2

T1

T2

REFERENCES Hall, Mark, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer, Peter Reutemann, and Ian H. Witten. 2009. “The WEKA Data Mining Software: An Update.” SIGKDD Explorations 11:10–18.

T3

Hoelzlwimmer, Andreas. 2010, September. “A Scheduling Framework Prototype for Heterogeneous Platform High Performance Computing.” Master’s thesis, University of Applied Sciences, Hagenberg.

T2

Howe, Doug, Maria Costanzo, Petra Fey, Takashi Gojobori, Linda Hannick, Winston Hide, David P. Hill, Renate Kania, Mary Schaeffer, Susan St Pierre, Simon Twigger, Owen White, and Seung Yon Rhee. 2008. “Big data: The future of biocuration.” Nature 455:47–50.

Figure 14: Canceling and rescheduling a task.

IBM. 2006. “Cell Broadband Engine Programming Handbook.” Technical Report, IBM.

Instead of losing the computational progress so far it would be preferable to keep it while migrating to a different hardware platform. This can be enabled when the source code of the algorithm is available. If the algorithm

Intel. 2005, November. Intel Pentium 4. http:// www.intel.com/support/processors/ 494

pentium4/sb/cs-007993.htm. last visit: July 28, 2010. Kruk, Lukasz, John Lehoczky, Kavita Ramanan, and Steven Shreve. 2007, December. Heavy Traffic Analysis for EDF Queues with Reneging. Moth, Daniel. 2011, June. Blazing-fast code using GPUs and more, with C++ AMP. Session at AMD Fusion Developer Conference 2011, http: //ecn.channel9.msdn.com/content/ DanielMoth_CppAMP_Intro.pdf. Munshi, Aaftab. 2011, January. The OpenCL Specification, Version 1.1.

NVIDIA. 2010, June. NVIDIA CUDA http: Reference Manual Version 3.1. //developer.download.nvidia.com/ compute/cuda/3_1/toolkit/docs/ last visit: CudaReferenceManual.pdf. June 28, 2010. Pinedo, Michael L. 2008. Scheduling: Theory, Algorithms, and Systems. 3rd Edition. Springer. Sutter, Herb. 2011, June. Heterogeneous Parallelism at Microsoft. Keynote at AMD http: Fusion Developer Summit 2011, //developer.amd.com/documentation/ presentations/assets/4-Sutter_ Microsoft-FINAL.pdf.

Nissen, S. 2003, October. “Implementation of a Fast Artificial Neural Network Library (FANN).” Technical Report, Department of Computer Science, University of Copenhagen. Nissen, Steffen. 2007, October. “Large Scale Reinforcement Learning using Q-SARSA(λ ) and Cascading Neural Networks.” Master’s thesis, University of Copenhagen.

Xhafa, Fatos, and Ajith Abraham. 2008. Chapter Metaheuristics for Grid Scheduling Problems of Metaheuristics for Scheduling in Distributed Computing Environments, edited by Fatos Xhafa and Ajith Abraham, 1–38. Springer.

495

AGENT-BASED SIMULATION OF ELECTRONIC MARKETPLACES WITH ONTOLOGY-SERVICES Maria João Viamonte(a), Nuno Silva(a), Paulo Maio(a) (a)

GECAD - Knowledge Engineering and Decision Support Research Group Institute of Engineering of Porto Portugal (a)

{mjv, nps, pam}@isep.ipp.pt

ontology problem of electronic negotiations (Viamonte and Silva, 2008). Consequently, given the increasingly complex requirements of applications, the need for rich, consistent and reusable semantics, the growth of semantically interoperable enterprises into knowledgebased communities; and the evolution; the adoption of semantic web technologies need to be addressed (Silva and Rocha, 2004). In that sense, a suitable approach to address this interoperability problem relies on the ability to reconcile vocabulary used in agents’ ontologies. In literature, this reconciliation problem is referred as Ontology Matching (Euzenat and Shvaiko, 2007). In order to provide help in the conversation among different agents, we are proposing what we call ontology-services to facilitate agents’ interoperability. More specifically, this work presents the AEMOS Agent-Based Electronic Market with Ontology-Service System, a multi-agent market simulator with ontology services. The system includes agents that provide services that allow other agents to communicate with each other in order to reach an agreement, ensuring that both parties are able to understand the terms of negotiation. The AEMOS system is an innovative project (PTDC/EIA-EIA/104752/2008) supported by the Portuguese Agency for Scientific Research (FCT).

ABSTRACT Agent technology has been successfully applied to the Electronic Commerce domain, but the diversity of the involved actors leads to different conceptualizations of the needs and capabilities, giving rise to semantic incompatibilities between them. It is hard to find two agents using precisely the same vocabulary. They usually have a heterogeneous private vocabulary defined in their own private ontology. In order to provide help in the conversation among different agents, we are proposing what we call ontology-services to facilitate agents’ interoperability. More specifically, this work presents a multi-agent market simulator with ontology services. The system includes agents that provide services that allow other agents to communicate with each other in order to reach an agreement, ensuring that both parties are able to understand the terms of negotiation. Keywords: Intelligent Agents, Simulation, Electronic Markets and Ontology Mapping 1. INTRODUCTION With the increasing importance of Electronic Commerce across the Internet, the need for software agents to support both customers and suppliers in buying and selling good/services is growing rapidly. It is becoming increasingly evident that in a few years the Internet will host a large number of interacting software agents. Most of them will be economically motivated, and will negotiate a variety of good and services. It is therefore important to consider the economic incentives and behaviors of ecommerce software agents, and to use all available means to anticipate their collective interactions. Even more fundamental than these issues, however, is the very nature of the various actors that are involved in Electronic Commerce transactions. The involved actors lead to different conceptualizations of the needs and capabilities, giving rise to semantic incompatibilities between them. It is hard to find two agents using precisely the same vocabulary. They usually have a heterogeneous private vocabulary defined in their own private ontology. This leads to different conceptualizations of the needs and capabilities, giving rise to semantic incompatibilities between them. This problem is referred to as the

2. AEMOS SYSTEM AEMOS system is an Agent Based Electronic Market where agents can customize their behaviors adaptively by learning each users preference model and business strategies. Unlike traditional tools, agent based simulation does not postulate a single decision maker with a single objective for the entire system. Rather, agents representing the different independent entities in electronic markets are allowed to establish their own objectives and decision rules. Moreover, as the simulation progresses, agents can adapt their strategies, based on the success or failure of previous efforts. AEMOS includes a complex simulation infrastructure; able to cope with the diverse time scales of the supported negotiation mechanisms and with several players competing and cooperating with each

496

The simulator was developed based on “A Model for Developing a MarketPlace with Software Agents (MoDeMA)” (Viamonte, Ramos, Rodrigues and Cardoso, 2006). The following steps compose MoDeMA: • Marketplace model definition, that permits doing transactions according to the Consumer Buying Behavior Model;

other. In each situation, agents dynamically adapt their strategies, according to the present context and using the dynamically updated detained knowledge (Viamonte, Ramos, Rodrigues and Cardoso, 2006). AEMOS is flexible; the user completely defines the model he or she wants to simulate, including the number of agents, each agent’s type, ontology and strategies. Figure 1, figure 2 and figure 3 shows the AEMOS System Interface.

Figure 1: AEMOS system Interface – Internal Market Configuration



Identification of the different participants, and the possible interactions between them;



Ontology specification, that identifies and represents items on transaction;



Agents architecture specification, and information flows between each agents module;



Knowledge Acquisition, defining the process that guarantees the agent the knowledge to act on pursuit of its role;



Negotiation Model, defining the negotiation mechanisms to be used;



Negotiation Protocol, specification of each negotiation mechanism rules;



Negotiation Strategies, specification and development of several negotiation strategies;



Knowledge Discovery, identification and gathering of market knowledge to support agents’ strategic behavior.

2.1. Multi-Agent Model Multi-agent model includes three main types of actors as illustrated in figure 4. External Agents (Buyers, Sellers) Internal Agents (MM, MDM, MF, OMi)

Figure 2: AEMOS system Interface – Buyer Configuration

Inter-Agent Communication Platform (OAA) Figure 4: AEMOS system layers External Agents represent entities whose behavior is intended to be simulated and studied. There are two different types of external agents: • •

Buyers (B) are agents representing entities desiring to acquire products; Sellers (S) are agents representing entities desiring to sell products.

Internal Agents provide services that allow external agents to communicate with each other in order to reach an agreement, ensuring that both parties are able to

Figure 3: AEMOS system Interface – Seller Configuration

497

understand the terms of negotiation. The main internal agents are: •







PPg i DT Agts→ Agtb

represents the proposal offered by

the S agent Agts to the B agent Agtb at time Τ, at the

Market Manager (MM) is responsible for the market management. Manages all internal agents, register external agents and manages agents associations. In order to participate in the market an agent must first register with the MM agent. Usually there is only one MM agent per marketplace; Market Data Manager (MDM) registers information about all external agents participating in the market. When an external agent register in the market, the MDM agent collects its information, which is later provided when necessary. This agent is also responsible for writing statistical reports that enable to validate the correct functioning of the market. Normally there is only one MDM agent per marketplace; Market Facilitator (MF) is the agent responsible for the information integration process in the messages exchanged between external agents. It is an intermediate agent during the negotiation process that ensures, or tries to ensure that both parties are able to understand each other. Multiple MF agents can exist per marketplace. These agents are initialized by the MM agent when necessary. When an external agent is registered an MF agent is associated, from that moment all messages related to the negotiation process are sent for the associated MF; Ontology Matching intermediary (OM-i) is the agent that supports the information integration process. For that, this agent request the services (e.g. perform the information transformation according with the approved alignment) provided by several ontology matching specialized agents. Multiple OM-i agents can exist per marketplace, being initialized by the MM agent when necessary. When a MF agent is initialized an OM-i agent is associated, from that moment all the requests related to the information integration are sent to the associated agent.

negotiation period D for a specific product. The B agent evaluates the proposals received with an algorithm that calculates the utility for each one, Agtb U PPgi

U Agtb

PPg

DT

i Agts→ Agtb ; if the value of PPgi for at time Τ is greater than the value of the counter-proposal that the B agent will formulate for the next time Τ, in the same negotiation period D, then the B agent accepts the offer and negotiation ends successfully in an agreement;

CPg

DT

i Agtb→ Agts is made by otherwise a counter-proposal the B agent to the next time Τ. The S agent will accept a buyer counter-proposal if

U

Agts

CPgi is greater than the value of the the value of counter-proposal that the S agent will formulate for the next timeΤ; otherwise the S agent rejects the counterproposal. On the basis of the bilateral agreements made among market players and lessons learned from previous bid rounds, both agents (B and S) revise their strategies for the next negotiation rounds and update their individual knowledge module.

3. THE ONTOLOGY-SERVICES MODEL To provide a transparent semantic interoperability between all Electronic Commerce actors, an ontologyservices infrastructure was added to AEMOS. Thus, the AEMOS system architecture recognizes three new types of actors: • Ontology Matching Service (OM-s) agent is able to specify an alignment between two ontologies based on some ontology matching algorithm. There are several OM-s on the marketplace, each one providing the same service but based on distinct matching approaches (e.g. syntactic, lexical, structural); • Ontology Matching Information Transformation (OM-t) agent is responsible to transform any information represented according to one ontology (i.e. source ontology) to a target ontology using an already specified alignment between those two ontologies. Multiple OM-t agents can exist per marketplace. When an OM-i agent is initialized an OM-t agent is associated, from that moment all the requests related to the information translation are sent to the associated agent; • Ontology Matching Repository (OM-r) agent registers the agreed ontology alignments specified between agents’ ontologies. These alignments are applied to enable further agents’ interactions. These actors deploy a set of complementary features among themselves whose goal is to automate and improve the quality of the results achieved in the

2.2. Bilateral Contracts at AEMOS In bilateral contracting B agents are looking for S agents that can provide them the desired products at the best price. We adopt what is basically an alternating protocol (Faratin, Sierra and Jennings, 1998). Negotiation starts when a B agent sends a request for proposal. In response, a S agent analyses its own capabilities, current availability, and past experiences and formulates a proposal. Seller’s agents can formulate two kinds of proposals: a proposal for the product requested; or a proposal for a related product, according to the B agent preference model.

498

single alignment between agents’ ontologies which need to be accepted by both agents. On the other hand, on scenario 3c each agent generates its own alignment according to its internal preferences. Due to agents’ different preferences and interests, the resulting alignments may have contradictory and inconsistent perspectives about candidate correspondences. Conflicts about correspondences are addressed by agents through the generic argumentation process described in (Maio, Silva, and Cardoso 2011a). In (Maio, Silva, and Cardoso 2011b) it is described how agents can exploit that argumentation process for ontology matching purposes. At the end, both agents need to inform the OM-i agent about the agreed alignment. On scenarios 2a, 2b, 3a and 3b the agent lacking OM capabilities needs to delegate such responsibility to the OM-i agent. Yet, because the other agent has OM capabilities two distinct alignments exist. Resulting conflicts about correspondences are addressed either: (i) solely by OM-i agent if none of the agents have OMN skills (scenario 2a and 3a) or (ii) by a negotiation process between the agent with OMN skills and the OM-i agent in representation of the other agent (scenario 2b and 3b). The Information Transformation phase is the set of information transactions through OM-i that transforms (i.e. converts) information described according to the sender’s ontology to be described according to the receiver’s ontology. This process is very methodical in accordance to the specified ontology alignment.

electronic commerce transactions. The OM-i agent is responsible to manage all these services and consequently to hide the resulting complexity of that task from the marketplace (namely from the MF agent). Figure 5 depicts the types of interactions between the marketplace internal agents (i.e. MF and OM-i) and the external agents (i.e. B and S). 0)

5HJLVWUDWLRQ

5HJLVWUDWLRQ

,QIRUPDWLRQ 7UDQVIRUPDWLRQ 2QWRORJ\SXEOLFDWLRQ %

2QWRORJ\SXEOLFDWLRQ 20L

2QWRORJ\PDWFKLQJ

6 2QWRORJ\PDWFKLQJ

TXHU\ 20V

20W

20U

2QWRORJ\PDWFKLQJQHJRWLDWLRQ

Figure 5: Marketplace actors and interactions The Registration phase is initiated by the B or S agent and allows these agents to identify themselves to the marketplace and specify their roles and services. The Ontology Publication phase is the set of transactions allowing B and S to specify their ontologies to the marketplace. The Ontology Matching phase is the set of transactions driven by OM-i to align the ontologies of agents B and S. This phase is crucial for agents’ interoperability and depends on the agents’ ontology matching capabilities. By ontology matching (OM) capabilities of an agent we mean the ability it has to generate an alignment between two ontologies reflecting its own preferences and interests (e.g. alignment requirements), being such alignment achieved by its own or in cooperation with a set of ontology matching specialized agents (available or not in the marketplace). Yet, agents having OM capabilities may (or may not) have ontology matching negotiation (OMN) skills. Therefore, according to the agents’ OM capabilities and OMN skills, for each pair of agents B and S six distinct scenarios are possible: 1. None of the agents’ have OM capabilities; 2. Only one of the agents (say AgOM) have OM capabilities: a. AgOM do not have OMN skills; b. AgOM have OMN skills; 3. Both agents have OM capabilities: a. None of the agents have OMN skills; b. Only one of the agents have OMN skills; c. Both agents have OMN skills. On the first scenario, the OM-i agent is fully responsible for the ontology matching task. Even though, OM-i agent may take into consideration a set of preferences about the ontology matching process specified by both agents. Therefore, OM-i generates a

4. AN ONTOLOGY MAPPING EXAMPLE For this example we consider a simple market with only two external agents (one B and one S). The B agent intends to purchase 10 units of the same product (mp3 player) using for its representation the Ontology Consumer Electronics Ontology (CEO). The S agent provides the desired product in sufficient quantity however uses for its representation the Ontology MP3Player (MP3_Player). The interaction between agents is shown in Figure 6. When MF receives a request for proposals for a product, as there are no S agents who use the same ontology as the B agent, MF makes a request to the OMi to suggest S agents that may be able to satisfy the request. OMi selects S agents that use some of the ontologies that can be mapped with the B agent’s ontology. From this selection results a list of S agents (in this case only one) where for each S agent is associated a proposal for the mapping of its ontology with the B agent’s. Then the MF asks B and S agents for approval of the proposed mapping. If both approve, confirms the approval to OMi and ask him to represent the B request data according to the S ontology. The transformed data is replaced on the original request and it’s forward to the S agent.

499

agents, no matter which machine they are running on or which programming language they are programmed in. OAA is not a framework specifically devoted to develop simulations; some extensions were made to make it more suitable, such as the inclusion of a clock to introduce the time evolution mechanism of the simulation. Each agent is implemented in Java, as a Java thread. The model can be distributed over a network of computers, which is a very important advantage to increase simulation runs for scenarios with a huge amount of agents.

Figure 6: Data Integration Interaction If the S formulates proposals, the MF makes a new request to the OMi so it represents the proposals data according to the B ontology. During this process it’s registered the approved mapping by the agents for the type of product. When a new communication is made by these agents related with this type of product (e.g. to close deal) the information is transformed using the approved mapping. In case that the Players (B and S agents) don´t approved the ontology mapping proposed by the system an Ontology Matching Negotiation Process (Maio, Viamonte and Silva, 2011) is used in order to obtain an agreement. It is envisaged that in the ontology matching negotiation phase agents adopt the argument-based negotiation process presented in (Maio, Viamonte and Silva, 2011). In real scenarios with more Players (B and S agents) the process descripted above is replicated.

5.

Figure 7: The OAA Facilitator

Figure 8: The AEMOS MarketPlace 6. CONCLUSIONS AEMOS project is an innovative project that proposes a semantic information integration approach for agentbased electronic markets based on ontology-based technology, improved by the application and exploitation of the trust relationships captured by the social networks. We intent face the problem of the growth of electronic commerce using software agents to support both customers and suppliers in buying and selling products. The diversity of the involved actors leads to different conceptualizations of the needs and capabilities, giving rise to semantic incompatibilities between them.

IMPLEMENTATION

The AEMOS system was developed in Open Agent Architecture (OAA) (http://www.ai.sri.com/~oaa/) and in Java. The OAA platform, figure 7, is a framework for integrating a community of heterogeneous software agents in a distributed environment. It is structured to minimize the effort involved in creating new agents, written in various languages and operating platforms; to encourage the reuse of existing agents; and to allow the creation of dynamic and flexible agent communities. The OAA’s Interagent Communication Language is the interface and communication language shared by all

500

Maio, Paulo, Nuno Silva, and José Cardoso. 2011a. “EAF-based Negotiation Process.” in The 4th International Workshop on Agent-based Complex Automated Negotiation (ACAN) at AAMAS. Taipei, Taiwan. Maio, Paulo, Nuno Silva, and José Cardoso. 2011b. “Generating Arguments for Ontology Matching.” in 10th International Workshop on Web Semantics (WebS) at DEXA. Toulouse, France. (CEO) Ontology Consumer Electronics Ontology. Available from: http://www.ebusinessunibw.org/ontologies/consumerelectronics/v1.owl)

Ontologies have an important role in Multi-Agent Systems communication and provide a vocabulary to be used in the communication between agents. It is hard to find two agents using precisely the same vocabulary. They usually have a heterogeneous private vocabulary defined in their own private ontology. In order to provide help in the conversation among different agents, we are proposing what we call ontology-services to facilitate agents’ interoperability. More specifically, AEMOS project proposes an ontology-based information integration approach, exploiting the ontology mapping paradigm, by aligning consumer needs and the market capacities, in a semi-automatic mode, improved by the application and exploitation of the trust relationships captured by the social networks. Yet, it is our conviction that the marketplace must encourage agents to play an important role in the required matching process. Even though, that cannot be a mandatory issue and therefore the marketplace must be equipped to deal with agents having different ontology matching capabilities. It is envisaged that by taking part in the matching process agents may become more confident in the underlying communication process and in face of that consider the electronic commerce exchanged data (e.g. RFP and PP) more reliable (safe) and consequently become more proactive in the marketplace.

(MP3_Player) MP3_ Player. Available from: http://daisy.cti.gr/svn/ontologies/AtracoProject/Atr acoUserProfile/Y2IntegrationFeelComfortable/MP3_Player.owl AUTHORS BIOGRAPHY MARIA JOÃO VIAMONTE is an adjunct professor of informatics at the School of Engineering of the Polytechnic Institute of Porto. Her research interests are in multi-agent simulation, agent mediated electronic commerce and decision support systems. She received her PhD in electrical engineering from the University of Trás-os-Montes and Alto Douro. She is coordinator and technical leader of several research projects Contact her at GECAD - Knowledge Engineering and Decision Support Research Group at the School of Engineering of the Polytechnic Institute of Porto, Rua Dr. António Bernardino de Almeida, 4200-072 Porto, Portugal; [email protected]

ACKNOWLEDGMENTS The authors would like to acknowledge FCT, FEDER, POCTI, POSI, POCI, POSC, POTDC and COMPETE for their support to R&D Projects and GECAD Unit. REFERENCES Viamonte, M. J. and Silva, N., 2008. Semantic WebBased Information Integration Approach for an Agent Based Electronic Market. Semantic Web Methodologies for E-Business Applications: Ontologies, Processes and Management Practices. Chapter VIII (150-169). Publisher: IDEA Book, Editor(s): Roberto García. Silva, N. and Rocha, J., 2004. Semantic Web Complex Ontology Mapping. Web Intelligence and Agent Systems Journal, vol. 1, no. 3, p. 235ʊ248, 2004. Euzenat, J. and Shvaiko, P., 2007. Ontology Matching. Firt ed., vol. 1, 1 vols. Heidelberg, Germany: Springer-Verlag, 2007. Viamonte, M.J., Ramos, C., Rodrigues, F. And Cardoso, J, 2006. ISEM: A Multi-Agent Simulator For Testing Agent Market Strategies. IEEE Transactions on Systems, Man and Cybernetics – Part C: Special Issue on Game-theoretic Analysis and Stochastic Simulation of Negotiation Agents, vol. 36, no. 1, pp. 107-113. Faratin, P., Sierra, C. and Jennings, N., 1998. Negotiation Decision Functions for Autonomous Agents. Int. J. Robotics and Autonomous System, 24 , 3, 1998, 159-182

NUNO SILVA is Coordinator Professor of informatics at the School of Engineering of the Polytechnic Institute of Porto. His research interests are Information Integration, Knowledge Engineering and the Semantic Web. He received his PhD in electrical engineering from the University of Trás-os-Montes and Alto Douro, Portugal. He is coordinator and technical leader of several research projects. Contact him at GECAD Knowledge Engineering and Decision Support Research Group at the School of Engineering of the Polytechnic Institute of Porto, Rua Dr. António Bernardino de Almeida, 4200-072 Porto, Portugal; nps@ isep.ipp.pt PAULO MAIO is an assistant professor of computer engineering at the Polytechnic Institute of Porto’s Institute of Engineering. He is also a PhD student at the University of Trás-os-Montes and Alto Douro. Current research interests are focus in the ontology matching problem applied to multi-agent systems promoting agents’ interoperability. Contact him at Instituto Superior de Engenharia do Porto, Departamento de Engenharia Informática, Porto, Portugal; [email protected].

501

KEY ISSUES IN CLOUD SIMULATION PLATFORM BASED ON CLOUD COMPUTING Lei Ren(a), Lin Zhang(a), Yabin Zhang(b), Yongliang Luo(a), Qian Li(c) (a)

School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191, China; (b) Beijing Simulation Center, Beijing 100854, China; (c) School of Information Technology, Shandong Institute of Commerce and Technology, Jinan 250103, China (a)

[email protected], (b)[email protected], (c)[email protected]

heterogeneous simulation resources (Bohu 2007). Despite the increasingly rich simulation resources over networks, it is even more difficult for users to find what they just need to complete their simulation tasks. Now networked M&S platform technology is facing new challenges from a user-centric viewpoint. One of the great challenges is that simulation users need to take full advantage of a variety of simulation resources in distributed and heterogeneous environments efficiently and transparently (Lei and Lin 2010). Simulation resources are the basis for modeling and simulation, and they include models, computing devices, storage, data, knowledge, software, and simulator needed by simulation systems. The simulation resources usually distribute in dispersed locations across different organizations over networks. To leverage the required simulation resource, users need to check its location, and then negotiate with the owner before they can access to it. This is a very inefficient operation mode for resources sharing. In addition, the distributed simulation resources often run in heterogeneous environments. Diverse hardware platform, operating system, and programming environment set up obstacles for users to integrate them together to support collaborative simulation applications. So the key problem is how to shield distribution and heterogeneity of simulation resources to provide a transparent access mechanism and build an efficient resources sharing environment. Moreover, users need to get on-demand simulation services according to their personalized requirements and use them ubiquitously. Currently the networked M&S technology has shifted the focus from computingcentric angle to user-centric. However, current complex simulation systems always bring about a heavy burden caused by the deployment and configuration work. Users have to spend much time installing hardware drivers, operating systems, and software tools to establish a specific simulation application. The system development process is cumbersome and cannot meet user's need of customizing a simulation system simply and rapidly. In addition, there is a growing demand for using mobile network terminals (e.g., pad computer, smart phone, etc.) to participate simulation anywhere

ABSTRACT Networked modeling & simulation platform can provide important support for collaborative simulation applications by integrating simulation resources over networks. Facing the increasing rich simulation resources over networks, the new challenge is that users need to get simulation services on demand simply and use simulation resources in a more efficient, transparent, and ubiquitous way. This paper presents Cloud Simulation Platform (CSP) that introduces the idea of Cloud Computing into networked modeling & simulation platform. We discusses the key issues in CSP including the operating principle of Cloud Simulation model, CSP architecture and components, and the key technologies in CSP. CSP provides a promising solution to the problems that networked modeling & simulation platform is facing, and this paper presents a map for future research on CSP. Keywords: cloud simulation, cloud computing, cloud manufacturing, high performance simulation, networked modeling and simulation platform 1. INTRODUCTION The rapid development of Modeling & Simulation (M&S) technology has profound influence on understanding the complicated world in recent years. M & S platform technology provides effective support for modeling, computing, analyzing, evaluating, verifying, and forecasting in many application domains such as simulation in military training, industry, and economics. To deal with the large amount of simulation resources distributing over networks, networked M&S platform technology has become one of the most powerful tools to support sharing and collaboration of distributed simulation resources (Zhang and Chen 2006), especially in large-scale simulation applications such as virtual prototyping of space shuttle. With the rapid development of distributed computing technology, more and more types of simulation resources can be connected into networks and take part in collaborative simulation process. This has resulted in a sharp increase in complexity of networked simulation systems, as well as increasing difficulty in making use of distributed and

502

anytime. Users don't want to know too many technical details about how to call a simulation function remotely, as well as how to integrate diverse simulation resources to implement a specific function. Thus users can pay more attention to simulation applications themselves instead of technical details. Therefore, the networked M&S technology had better provide a new paradigm by which simulation resources could be integrated to construct a simulation system according to the needs of users flexibly and dynamically. And the simulation system can be accessed and used through ubiquitous terminals anytime anywhere. To address these issues, the idea of Cloud Computing (Buyya et al. 2009) may provide an opportunity to give impetus to the development of networked M&S technology. Cloud Computing refers to a pattern of IT service delivery and utilization. In Cloud Computing, users may access the scalable IT resources on demand via networks by using a computer, smart phone and other interactive terminals, and they don't need to download and install applications on their own devices because all IT resources (computing and storage) are maintained by cloud servers. In this paper we introduce the idea of Cloud Computing into networked M&S technology and present Cloud Simulation Platform (CSP). The paper firstly presents the operating principle of Cloud Simulation model. Then the CSP architecture and the components are illustrated, and the key technologies in CSP are discussed to indicate the future research areas. The key issues discussed in this paper can contribute to the future research on CSP. 2. RELATED WORK There are many different definitions for Cloud Computing. For example, it refers to a large-scale distributed computing paradigm that is driven by economies of scale, in which a pool of abstracted, virtualized, dynamically-scalable, managed computing power, storage, platforms, and services are delivered on demand to external customers over the Internet (Foster et al. 2008). Cloud Computing is considered as a new business paradigm describing supplement, consumption and delivery model for IT services by Utility Computing (Mladen 2008) based on the Internet. The typical examples of public Utility Computing include Google AppEngine, Amazon Web Services, Microsoft Azure, IBM Blue Cloud, Salesforce. So far, Cloud Computing is regarded as the sum of IaaS (Infrastructure as a Service), PaaS (Platform as a Service), and SaaS (Software as a Service). Cloud Computing providers can deliver on-demand services covering IT infrastructure, platform, and software vie virtualization and service encapsulation technology. Thus, it can meet the needs of ever-rising scale of computing and storage of consumers and lead to decrease in IT investment cost at the same time. Cloud Computing introduces a promising model and technical approach to the problems networked M&S platform facing. For example, virtualization technology (Thomas et al. 2005) can support deep encapsulation of a logic

entity of IT infrastructure (e.g., CPU, memory, disk, and I/O) and software into a pool of virtual machines, thus it can achieve high efficient and transparent utilization of resources. 3. CLOUD SIMULATION MODEL Figure 1 illustrates the operating principle of Cloud Simulation model. A simulation cloud refers to a cluster of virtualized simulation resources and services. The physical simulation resources can be mapped into virtualized resource templates by virtualization technology (Lei and Lin 2010). For example, package ANSYS software, Window XP, 4 CPUs, 1G memory, and etc. into a virtual machine template. And the functions of a simulation resource can be encapsulated into standard services that are loosely coupled and interoperable. The simulation clouds can shield the complexity caused by the distribution and heterogeneity of resources, and accomplish unified management of the standard services. Simulation clouds Virtualized simulation resources and services

Networks

Cloud Simulation Platform Virtualized simulation environment on demand

Virtualized simulation environment on demand

Ubiquitous UI

User

Ubiquitous UI

Virtualized simulation environment on demand

Ubiquitous UI

User

User

Simulation tasks Simulation task description

Service discovery & composition

Simulation running

Interaction & visualization

Ă

Figure 1: Cloud Simulation Model Users don't need to deploy their own simulation environments involving hardware and software meeting specific needs. A user can submit a request of simulation tasks to CSP by using ubiquitous interactive terminals (e.g., smart phone, pad computer, etc.), and the request will be parsed by CSP to generate formal requirements of simulation resources. CSP launches a search to match the needed resources in the simulation clouds according to the resource requirements. The selected resources then will be accumulated to establish a virtualized simulation environment, and CSP can complete the deployment of the resources automatically by the way of instantiating the virtualized resource templates. In the virtualized simulation environment,

503

users feel like facing a real system designed for his specific needs, and the simulation functions can be acquired in a unified form of service, e.g., web service (Papazoglou et al. 2007), to support collaborative simulation. In run-time simulation process, CSP can assemble services and schedule resources dynamically over networks, and users can interact with the virtualized system as well as monitor the visual feedback remotely. Once a crash or unrecoverable error occurred in some simulation resource, CSP can carry out live migration (Christopher et al. 2005) from the trouble resource to another healthy one. This ability of

fault tolerance can achieve a high reliable collaborative simulation, and the whole process is transparent to users. 4. CLOUD SIMULATION PLATFORM ARCHITECTURE Figure 2 shows the layered architecture of CSP. It consists of four layers: Virtualized Resource Layer, Middleware Layer, Simulation Service Layer, and Ubiquitous Portal Layer. CSP can integrate a broad range of simulation resources as Figure 2 illustrates, and provide support for simulation applications such as military training, product development, and conceptual prototype verification.

Figure 2: Cloud Simulation Platform architecture

504

to encapsulate simulation resources into virtualized templates, thus the corresponding physical resources can be managed through the exposed interfaces. The service encapsulation tool is used to package the functional interfaces of the simulation resources into standard services, and this tool provides a development environment for users to customize the domain-specific services. The high-level modeling tools gives support for users to complete high-level modeling work of simulation tasks with interactive visual tools. The simulation run-time service is responsible for the run-time process management. There are four types of services supporting run-time simulation. The discovery & match module is used to search the suitable services and resources in response to simulation requests. Another module is responsible for establishing the virtualized simulation environment dynamically meeting the needs of simulation tasks, and deploying the resources automatically. The collaborative resource scheduling module is used to parse the description of coordinated tasks, manage the time steps of concurrent tasks, and dispatch the services and resources to complete the simulation tasks. The simulation application service provides the support of common functions that simulation applications need. The module of virtual UI is used to generate personalized virtual desktop interfaces. Users can access CSP and interact with applications through virtual UI by network browsers. The remote visualization module is responsible for rendering visual graphics that show the live image of simulation task progress. The module of batch job scheduling is used to manage the task queue and schedule the tasks. The license scheduling module is used to manage the licenses of commercial software, including license reservation, dispatch and recovery. The accounting & charging module is used to account the simulation resource usage according to the rate set by the resource provider and charge the users. The module of user permission management provides mechanism for user management, such as account maintenance, identity authentication, role assignment, and access control. 4.4. Ubiquitous Portal Layer The Ubiquitous Portal Layer provides interfaces supporting ubiquitous UI for users to access CSP services. This layer offers interface adapter for interactive terminal devices including PC, pad computer, and smart mobile phone, so users can acquire simulation services without time and space constrains. The CSP portals include simulation service portal, virtual desktop portal, batch job processing portal, and collaborative simulation task portal. The simulation service portal is the homepage where users can search for simulation services and customize their own simulation applications. To better support typical simulation application patterns, three kinds of portal mentioned above are offered on the homepage. In the virtual desktop portal, users can customize preferred system environment including hardware and software, then CSP find the matched virtualized

4.1. Virtualized Resource Layer The Virtualization Resource Layer is responsible for connecting a variety of simulation resources over networks and mapping them to virtualized templates according to a unified description specification. The simulation resources consist of computing resource, software resource, model resource, data resource, knowledge resource, license resource, network resource, and simulator device resource. The physical resources such as computing resource, software resource, and model resource can be mapped to virtual machine templates where the fine-grained resources (e.g., CPU, memory, etc.) can be regrouped to create new virtualized resources. These virtualized resources are logical partition and the abstract composition of the real ones, and the templates should be described according to the unified description specification. The provider of the resources can register the templates in the registration center of CSP to establish a virtualized resource pool. The pool seems like a cloud where the raindrops of simulation resources aggregate and move. 4.2. Middleware Layer The Middleware Layer serves as middleware with the common functions that links the Virtualization Resource Layer with the Simulation Service Layer. This layer includes three types of middleware to implement resource management, service management, and collaborative simulation, respectively. The resource management middleware is used to manage the virtualized resource templates pool as well as the physical resources. It takes charge of allocating resources dynamically according to the scheduling commands issued from the upper layer. The distributed and heterogeneous resources are under control of this middleware and the real-time running status can be monitored. The service management middleware acts as a service container to support unified service description, registration, discovery, match, composition, and remote call. This middleware shields the complexity caused by distribution and heterogeneity, and makes the functions of the resources independent from specific infrastructure. The collaborative simulation middleware offers support for run-time simulation process management. This middleware serves as a special service-oriented RTI that is responsible for maintaining simulation runtime status, allocating data in federations and controlling the synchronization of multiple tasks. 4.3. Simulation Service Layer The Simulation Service Layer is the core functional component to provide frequently-used services supporting simulation applications. The services are divided into three categories: simulation application development service, simulation run-time service, and simulation application service. The simulation application development service is used to design and customize simulation applications with three tools. The resource virtualization tool is used

505

templates and instantiate the virtual machines. Finally CSP returns the remote virtual UI to users for interaction. In the batch job processing portal, users submit batch job files to CSP, then CSP parses the file and find the needed simulation services and resources to deploy simulation environment. Once the environment is ready, the tasks are uploaded to the virtual machines, and CSP returns result to users when the job is done. In the collaborative simulation task portal, users can establish the high-level model of simulation tasks. The formal description of the model is parsed by CSP, then CSP discover the needed services and resources to create simulation federation. In the simulation process, users can monitor the run-time status through the portal. In addition, users can start, pause, continue, and stop the simulation progress in the portal. 5. KEY TECHNOLOGIES IN CLOUD SIMULATION PLATFORM 5.1. Simulation Resource Virtualization and Ondemand Use To support high efficient sharing and flexible use of large-scale simulation resources, virtualization is one of the key techniques (Lei et al. 2011). Virtualization technology can decouple simulation applications from needed simulation resources, thereby allowing the finegrained resources to integrate on demand flexibly. The key issues in simulation resource virtualization include simulation resource taxonomy, unified formal description of simulation resource, virtualized simulation resource template, composition verification, management of large-scale virtualized resource pool, mapping approach from physical simulation resource to virtualized resource, and remote management of simulation resource status. 5.2. Service-oriented Simulation Resource Publication and Intelligent Match Service computing technology plays a important role in CSP. The unified service encapsulation of simulation resource make it possible to shield the complexity derived from the distributed and heterogeneous resources. Moreover, standard service interfaces can implement effective inter-operability in collaborative simulation based on standard protocols. To realize efficient and intelligent search for simulation resources, semantics-based service match technique (Martin et al. 2007) is essential to CSP. The key issues include formal description of service of simulation resource, semantics description of simulation service, semantic service encapsulation approach, simulation service publication, Ontologybased service match method, and semantic composition of simulation service. 5.3. Simulation System Dynamically Construction and Deployment One of the advantages of CSP is the capability of constructing a simulation environment on demand dynamically and rapidly. This technique facilitates the time-consuming deployment work for complex simulation system to a large degree. In addition, it can

optimize the utilization of simulation resources in runtime simulation along with the fluctuating resource needs. The key issues include formal description of simulation task requirement, automatic parsing of resource requirement, intelligent match of virtualized resource template, virtualized resource composition optimization, automatic deployment of simulation resource, and virtualized simulation resource instantiation and run-time management. 5.4. Fault Tolerance and Migration in Run-time Simulation One of the most important targets of CSP is to achieve high reliability, because a variety of failures and errors are inevitable in collaborative simulation progress over unstable networks. CSP should have the ability of fault tolerance to ensure the simulation tasks proceed at the lowest cost once failures occurred. Migration technique offers a mechanism that enables fault tolerance in runtime simulation. The key issues include run-time simulation monitoring, risk evaluation and fault prediction, runtime simulation failure detection, optimal selection for migration target, migration cost evaluation, live migration in run-time simulation, and post-migration simulation tasks synchronization. 5.5. Utilization Accounting The distinguishing characteristic of CSP different from other networked M&S platform, such as Grid simulation platform, is that CSP gives support for business transaction between simulation resource consumers and providers. And the transaction billing accords with the actual usage of simulation resources, just like electricity utility charging. The simulation resource utilization accounting technology is the key to accomplish the target. The key issues include multiple-level accounting model of simulation resource utilization, fine-grained simulation resource utilization statistics, transaction and rate standard management, and user account security management. 5.6. Ubiquitous UI and Remote Virtual Interface User operating environment is no longer confined to PC desktop in CSP. CSP provides more powerful support for simulation in mobile environments relying on ubiquitous computing and virtualization technology. Users can access CSP by using mobile terminals such as pad computer and smart phone, and customize their virtual UIs to implement remote interaction. The key issues include ubiquitous terminal adapter standard, interaction context perception and processing, adaptive visualization in small UIs, personalized portal customization, virtual desktop customization and automatic building, and remote interaction and visualization in virtual UIs. 6. CONCLUSION Networked M&S platform gives powerful support for collaborative simulation applications by means of integrating simulation resources over networks. Facing the new challenge of increasing difficulty in making use

506

Computing and Information Technology, 16(4), 235-246. Papazoglou, M., Traverso, P., Dustdar, S., and Leymann F., 2007. Service-Oriented Computing: State of the Art and Research Challenges, IEEE Computer, 40(11), 38-45. Thomas, A., Larry, P., Scott, S., and Jonathan, T., 2005. Overcoming the Internet impasse through virtualization. IEEE Computer, 38 (4), 34-41. Zhang, H. and Chen, D., 2006. An approach of multidisciplinary collaborative design for virtual prototyping in distributed environments. Proceedings of the 10th International Conference on Computer Supported Cooperative Work in Design, pp. 1-6. May 3-5, Nanjing, China.

of massive simulation resources in distributed and heterogeneous network environment, users need a way to leverage the rich simulation resources to build complex simulation applications rapidly and simply. This paper introduced the idea of Cloud Computing into networked M&S platform and discussed the key issues in Cloud Simulation Platform. We discussed the operating principle of Cloud Simulation model, CSP architecture, and the key technologies in CSP. The proposed CSP provides a promising approach that users can get simulation services on demand and use simulation resources in a more efficient, transparent, and ubiquitous way. And this paper provides a research map for future research on CSP. We are developing a CSP prototype now and we plan to establish a M&S application of aircraft design on CSP in the future.

AUTHORS BIOGRAPHY Lei Ren, Ph.D. Dr. Lei Ren received Ph.D. degree in 2009 at the Institute of Software, Chinese Academy of Sciences, China. From 2009 to 2010 he worked at the Engineering Research Center of Complex Product Advanced Manufacturing System ˈ Ministry of Education of China. He is currently a researcher at the School of Automation Science and Electrical Engineering, BeiHang University. He is a member of SCS and SISO. His research interests include high performance simulation platform, integrated manufacturing systems, Cloud Simulation, Cloud Manufacturing and Cloud Computing. Lin Zhang, Ph.D., Professor. Lin Zhang received M.S. degree and the Ph.D. degree in 1989 and 1992 from the Department of Automation at Tsinghua University, China, where he worked as an associate professor from 1994. From April 2002 to May 2005 he worked at the US Naval Postgraduate School as a senior research associate of the US National Research Council. Now he is a full professor in BeiHang University. His research interests include system modeling and simulation, integrated manufacturing systems, and software engineering. Yabin Zhang is now a Ph.D. candidate at the School of Automation Science and Electrical Engineering, BeiHang University. His research interests include Cloud Simulation, and Cloud Computing. Yongliang Luo is now a Ph.D. candidate at the School of Automation Science and Electrical Engineering, BeiHang University. His research interests include service-oriented manufacturing and integrated manufacturing systems. Qian Li received the B.S. degree and the M.S. degree in 2002 and 2008 from the School of Computer Science and technology at Shandong University, China. She is now a researcher at the School of Information Technology, Shandong Institute of Commerce and Technology. Her research interests include ubiquitous user interface, VR and visualization.

ACKNOWLEDGMENTS The research is supported by the Fundamental Research Funds for the Central Universities in China, and the NSFC (National Science Foundation of China) Projects (No.61074144, No.51005012) in China. REFERENCES Bohu, L., 2007. Research and Application on Virtual Prototyping Engineering Grid. System Modeling and Simulation, 2 (15): 304-308. Buyya, R., Yeo C.S., Venugopal S., Broberg J., and Brandic I., 2009. Cloud computing and emerging IT platforms: Vision, hype, and reality for delivering computing as the 5th utility. Future Generation Computer Systems, 25 (6), 599-616. Christopher, C., Keir, F., Steven, H., Jacob, G.H., Eric, J., Christian, L., Ian, P., and Andrew, W., 2005. Live Migration of Virtual Machines. Proceedings of the 2nd ACM/USENIX Symposium on Networked Systems Design and Implementation, pp. 273-286. Boston (CA, USA). Foster, I., Zhao, Y., Raicu, I., and Lu, S., 2008. Cloud Computing and Grid Computing 360-degree compared, Proceedings of Grid Computing Environments Workshop, pp. 1-10. Nov. 12-16, Austin (Texas, USA). Lei, R., and Lin, Z., 2010. VSim: A Virtual Simulation Framework for High Performance Simulation[C]. Proceedings of Summer Simulation Multiconference, pp. 38-44. July 11-15, Ottawa (Ontario, Canada). Lei, R., Lin Z., Fei T., Xiaolong, Z., Yongliang, L., and Yabin, Z., 2011. A methodology toward virtualization-based high performance simulation platform supporting multidisciplinary design of complex products. Enterprise Information Systems, 5 (3), 1-24. Martin, D., Burstein, M., Mcdermott, D., and Mcilraith, S., 2007. Bringing semantics to web services with owl-s, World Wide Web, 10 (3), 243-277. Mladen, A.V., 2008. Cloud Computing – Issues, Research and Implementations. Journal of

507

RESEARCH ON KEY TECHNOLOGIES OF RESOURCE MANAGEMENT IN CLOUD SIMULATION PLATFORM Ting Yu Lin (a), Xu Dong Chai (b), Bo Hu Li (c) (a)

(b)

School of Automatic Science and Electrical Engineering, BeiHang University, Beijing 100191 China Beijing Simulation Center, Second Academy of Aerospace Science & Industry Co., Beijing 100854 China (c) School of Automatic Science and Electrical Engineering, BeiHang University, Beijing 100191 China (a)

[email protected], (b) [email protected], (c) [email protected]

resource, licensing resource, model resource, equipment resource and capability resource. The distribute resources are accessed through virtualization middleware, service-oriented middleware and sensing middleware, and appeared as virtual resources, service resources and physical resources to the upper layer, shown in the Figure 1.

ABSTRACT For the diversity and the life-cycle dynamics of the resources on the cloud simulation platform, this paper discuss how to cross and integrate all types of resources management systems to realize the centralized management of distributed resources and then supply services of unifying managed resources to distributed users efficiently. This paper proposed the guideline and architecture of the cloud simulation platform at first. Then, from the perspective of resource operation and running, present the unified simulation modeling to support the dynamic management for the life cycle of the resources. After that, based on the concept of the resource group, this paper proposes the resource selection technology to support the efficient allocation for types of the resources. Finally, this paper introduces an application in the collaborative design and simulation for the multi-disciplinary virtual prototype.

Figure 1: Resources in Cloud Simulation Platform In terms of virtual resource, both Xen (Barham et al. 2003) and VMware (Anderson et al. 2005) are popular technology. The virtual machine cluster is managed through VMMs on each nodes and management center on the master node. In terms of service-oriented resources, Apache axis24 is used as service container on each node to deploy and monitor the services, and Platform ego5 is used to manage the services of entire cluster. In terms of physical resources, the performance of physical cluster are monitored by Ganglia6, and the management of them are implemented by distribute agents. However, the diversity of resources on the cloud simulation platform makes it difficult to aggregate resources on demand. We need to integrate all types of resource management systems, and take the resources’ collaboration into account, to achieve the optimal selection and combination of various resources. Cloud simulation platform uses resource management middleware to achieve the unified management of various resources, and uses dynamic construction modular to achieve the dynamic combination of various resources, so as to solve the problem effectively. Based on the analysis of related work in Section II, we propose the guideline and the architecture of the resource management in cloud simulation platform in Section III. The resource management of cloud simulation platform involves several key technologies,

Keywords: cloud simulation platform, resource management, unified modeling, resource selection 1. INTRODUCTION Cloud Simulation is a service-oriented, intelligent, agile, green, new networktized simulation paradigm. Combining with the emerged information technologies such as cloud computing, service-oriented, virtualization, high performance computing, and developing the existing networktized modeling and simulation (M&S) technology, cloud simulation encapsulates the simulation resource and capability as virtualization and service-oriented forms, and then constructs the cloud service pool of simulation resource and capability so as to implement unified, centralized management and operation, which will support users to access the services of simulation resource and capability on demand at any time through the network for their various activities during the life cycle of simulation. The paper (Bo Hu Li. et al. 2009) has fully introduced the technical content, application mode, the architecture and key technologies of the cloud simulation platform. There are kinds of resources on cloud simulation platform, including physical machine resource (CPU, memory, storage), virtual machine resource, software

508

including unified modeling techniques, selection techniques, combination technology, and operation/running technology, fault-tolerant technology, evaluation technology and so on. This paper focuses on the modeling technology and selection technology. In section IV, we present the unified formal description model of the simulation resource and resource instance from the perspectives of operation and running. In section V, we present the selection technology based on the resource group. Then, we introduce an application of the resource management in the field of multidisciplinary virtual prototype. Finally, in section VII, we give a summary and future outlook.

security and high availability of operating environment for the upper applications. With the rapid development of virtualization technologies recently, we can shield the differences of the hardware architecture, operating environment and other simulation resource, and uncouple the coupling relationship between them, making the construction and operation of simulation systems more flexible; we can provide a unified encapsulating standard for kinds of the heterogeneous simulation resources, making share and transparent use of simulation resources as much as possible; we can migrate and expand simulation operating environment dynamically and efficiently, so as to enhance the reliability and stability of the simulation runtime environment. After merging the latest technology of information field, such as virtualization technology, into the simulation grid, we have developed Cloud Simulation, which makes a better solution to the shortcomings of the simulation grid. It is known that the cloud simulation is the further development of the simulation grid, therefore, some technology of resource management in the simulation grid can be learned and continue to be adopted. Chang Feng Song, et al. (2009) proposes a resource selection model and algorithm in the environment of the simulation grid. For the "All-to-All" task model that any federate in the distribute interactive simulation possibly interacts with any other federate, the paper aims to select the grid node with CPU frequency as large as possible, communication delay as small as possible, communication speed as quick as possible from the grid nodes contained the required grid services. The work presents a Resource Selection Model (RSM) in matrix, and then uses various intelligent optimization algorithms to select a set of optimal resource instances from eligible simulation resources. However, since there are subtle changes in the resource form, the resource management of cloud simulation platform needs to fine-tune. For example, in the past, the task was to select M optimal nodes from N candidate nodes contained the required resource services (N>M), however, one node can use virtualization technology to build multiple copies of resource services at present. In the past, the required resource services were just fixed on N (finite) candidate nodes, however, the number of candidate nodes now can be much larger than N, since the introduction of virtualization technology which can build virtual nodes dynamically. These differences are related to the adjustment of the unified modeling technology and the selection technology for the simulation resources, which is the focus of this paper.

2. RELATIVE WORK Since 1983, the development of distributed simulation technology has experienced four periods which are Simulator Networking (SIMNET), Distributed Interactive Simulation (DIS), Aggregate Level Simulation Protocol (ALSP) and High Level Architecture (HLA) (Dahmann et al. 1998). HLA, whose latest progress is the HLA-Evolved8-10, is a popular technology of distributed simulation. It can provide a flexible general-purpose simulation framework for the M&S of complex systems, and can improve the Interoperability and reusability of the simulation model and simulation system. Therefore, it has been widely applied in M&S. After making good solution of the Interoperability and reusability on the level of simulation model such as components, federates and services, in order to manage and use all kinds of simulation resources better, Grid technology (Foster 2002) is introduced into simulation, which is so called Simulation Grid (Bo Hu Li et al. 2006). Simulation Grid, such as Cosim Grid (Bo Hu Li et al. 2006), NessGrid (Pearlman et al. 2004) and FederationX Grid (Einstain 2005), is a new generation of M&S support system, combined with the new network technologies such as Internet technology, Web Service technology and Grid technology and the traditional M&S support technologies such as HLA technology. In the simulation grid, the simulation models are packaged as loose coupling simulation services with interoperability deployed and shared on the grid node, and are dynamically discovered and invoked in the runtime. As a result, the traditional pattern of chimney-like development and resource utilization are broken, and establish new resource management and utilization patterns of autonomous, dynamic sharing and on-demand collaborative. However, from the view of application, current simulation grid services are almost fixed on some grid nodes, rather than created resource copy flexibly. In addition, current simulation grid services cannot penetrate the underlying hardware facilities so as to share fine-grained resources, including CPU, memory, software, etc, and it is difficult to satisfy multi-users to access all kinds of M&S services through the Internet anytime, anywhere. Again, the heterogeneous and loose environment of simulation grid cannot provide the

3.

GUIDELINE AND ARCHITECTURE OF RESOURCE MANAGEMENT Before discussion the simulation resource management technology about unified modeling and selection of resources in detail, we need first to introduce the guideline and the architecture of cloud simulation platform.

509

The guideline of cloud simulation platform is shown as follow: • Centralized management of distributed resources Fully integration of networked and shared computing resources, storage resources, software resources, License resources, knowledge / model resources, equipment resources and capability resources conducts unified management; • Supplying services of unifying managed resources to distributed users The aim of system is to organize simulation resources and capabilities quickly and flexibly, so as to provide services transparently on demand at anytime and anywhere. • Effective collaboration of resources on the life cycle of M&S The resource management covers the whole life cycle of M&S to support four application patterns, which include multi-users complete the design and analysis tasks independently, multi-users complete the simulation tasks collaboratively, multi-users complete the muti-analysis process collaboratively and multiusers access simulation capability, which is the combination of the intellectual resource and the traditional simulation resource, on demand. This paper does not take the fourth pattern into consideration. The architecture of cloud simulation platform is shown in the Figure2.

monitor agents and so called physical computing resources. Management layer includes physical machine management, virtual machine management, service management, software and their License Management. They will be integrated by resource management middleware, which is deployed on a dedicated server, so as to manage resource uniformly. With the expansion of the resources' scale and terrain, the deployment of the resource management middleware will be distribute, however the pattern of management and allocation of the resources keeps still. Supply layer includes statistics and analysis modular and dynamic construction modular, which both deploy on the dedicated servers respectively. Dynamic construction modular prepares hardware, software and service environment ready for the upper application, and statistics and analysis modular provides the reporting service about resources for the users. From the architecture of cloud simulation platform mentioned above, we can summarize that the resource management of cloud simulation platform covers physical machine resource management, virtual machine resource management, service resource management, software resource management and license management, shown in the Figure 3. In essence, license resource management subordinates to software resource management.

Figure 3: Types of Resource Management

Figure 4: Life Cycle of Resource Management In order to achieve the centralized management of distributed resources, as well as supply services of unifying managed resources for distributed users, resource management of cloud simulation platform covers the entire life cycle of resource, includes the unified registration, flexible configuration, real-time monitoring, on-demand allocation, efficient deployment, transparent running, timely retrieve and safe destruction for all kinds of simulation resources, shown in the Figure 4.

Figure 2: Architecture of Cloud Simulation Platform There are three layers in the architecture, which are Supply Layer, Management Layer and Resource Layer. Resource layer includes distributed computing resources, storage resources, service resources, which including service-oriented simulation software, simulation models, simulation equipments, and their license resources. Parts of the cluster is deployed virtualization middleware and become virtual computing resources so as to build virtual computing nodes on demand, other parts of the cluster is deployed

4.

UNIFIED MODELING OF SIMULATION RESOURCES FOR OPERATION AND RUNNING To support the unified management for the life cycle of various simulation resources, we first need to model all

510

kinds of simulation resources on the cloud simulation platform uniformly. Generally, there are two perspectives of unified modeling, the one is resource operation, and the other is resource running. The former concerns about the resource's grouping, ownership, availability and state of allocation; the later describes the static configuration, dynamic performance and running state of the whole life cycle for kinds of simulation resources. Figure 5 shows the formal description model of simulation resource and resource instance, as well as the relationships between them, from the perspective of resource operation. The resource attributes include resource ID, resource name, resource type and set of resource instances. Resource type can be divided into physical machine resource (identified as HPC), virtual machine resource (identified as VM), service resource (identified as SVR), and software resource (identified as SFW), from the analysis mentioned above. The set of resource instances achieve the association between resource and resource instances. A simulation resource includes a number of resource instances, and can be reverse indexed by any resource instance.

life cycle of resource instances, including NOSTATE, IDLE, CONFIGURED, RESERVED, OCCUPIED, UNAVAILABLE, ERROR and DESTROYED. The conversion between the various states is shown in Figure6. Figure 7 shows the formal description model of simulation resource and resource instance, as well as the relationships between them, from the perspective of resource running. The resource attributes include basic information (such as resource ID), the character information (such as operating system type), resource type and set of resource instances. The resource instance attributes include static configuration (such as the number of CPU), dynamic performance (such as CPU utilization) and running state of the whole life cycle. Similarly, the set of resource instances achieves the association between resource and its instances. A simulation resource includes a number of resource instances, and can be reverse indexed by any resource instance. Resource Model Basic Infomation

Resource

Feature Infomation

Resource Instance Set

Basic Info

Resource Instance

Static Configuration

Concept

Attribute

State

Resource Instance Model

Dynamic Performance

Property

Mapping

Figure 7: Formal Description Model from the perspective of Resource Running In fact, there are subtle differences for different types of simulation resources and resource instances in their formal description models, from the perspective of resource running. There are detailed definitions of resource and resource instance for the physical machine, virtual machine, service and software respectively, shown as follow. Definition I: Physical Machine Resource and Resource Instance

Figure 5: Formal Description Model from the perspective of Resource Operation

Figure 6: State Transition between Resource Instances



In addition, the resource instance attributes include instance ID, instance name, resource ID, resource group ID, ownership ID and state of allocation. Resource group ID is an important basis for the allocation of resources. Resource instances contacted with each other closely and executed collaboratively will be assigned to the same resource group. Ownership ID is an important symbol described the occupancy status and availability of the resource instance. From the perspective of resource operation, the state of allocation is an important property of the dynamic management for the

Physical Machine Resource < Basic Information , Character Information > Physical Machine Node < Basic Information , Static Configuration , Dynamic Performance , State {HPC_NOSTATE, HPC_INI, HPC_RUNNING, HPC_SHUTOFF, HPC_ERR}>

Definition II: Virtual Machine Resource and Resource Instance Virtual Machine Resource

511

cluster inside which the point to point communication is based on infiniband, so that the communication performance can almost always meet with the requirement of the collaborative simulation. Therefore, for the resource selection in a resource group, which usually refers to select suitable physical machines to deploy simulation models or startup virtual machines, we focus on considering the performance of each node from the perspective of CPU num and memory size. In the resource group, the algorithm of selecting and allocating the resources for collaborative simulation was shown as follow with the form of pseudo-code.

< Basic Information , Character Information > Virtual Machine Node < Basic Information , Static Configuration , Dynamic Performance , State {VM_NOSTATE, VM_INI, VM_RUNNING, VM_PAUSE, VM_BLOCKED, VM_SHUTOFF, VM_ERR }>

Definition III: Service Resource and Resource Instance

Variable Declaration

Service Resource < Basic Information < Resource ID, Resource Name >, Character Information < Is Parallel, OS Type > > Serviece Instance < Basic Information , Static Configuration < UDDI, CPU Num, Memory Size, Host ID >, Dynamic Performance < >, State {SVR_NOSTATE, SVR_INI, SVR_IDLE, SVR_RUNNING, SVR_PAUSE, SVR_SHUTOFF, SVR_ERR}>

rCPUNum: the requirement of the cpu num, which is an integer rMEMSize: the requirement of the memory size, which is an integer rIndex: the requirement of the computing performance, which is a real number pCPUNumˈpCPUSpeed, pCPUUtil: the cpu number, its speed and its utilization in the physical compute node pMEMSizeˈpMEMUtil: the memory size and its utilization

Definition IV: Software Resource and Resource Instance

in the physical compute node

Software Resource < Basic Information < Resource ID, Resource Name >, Character Information < Is Parallel, OS Type > > Software Instance < Basic Information , Static Configuration < UDDI, CPU Num, Memory Size, Host ID (List)>, Dynamic Performance < >, State {SFW_NOSTATE, SFW_INI, SFW_RUNNING, SFW_PAUSE, SFW_SHUTOFF. SFW_ERR }>

pCPUIndex: the available performance of the cpu, which is a real number pMEMIndex: the available performance of the memory, which is a real number pIndex: the available performance of the computing, which is a real number Selection Algorithm For i = 1 to M //M is the requirement number of the computer node

5.

RESOURCE SELECTION TECHNOLOGY BASED ON RESOURCE GROUP In the unified modeling of simulation resources mentioned above, we introduce the concept of resource group. Administrators can classify the simulation resource instances closely as a resource group. For example, a virtual machine template was deployed in a high performance cluster with shared storage, and then a virtual machine can be start up at any node of the cluster according to the mission requirement. Then the high performance cluster with the virtual machine template can be classified as a resource group. As another example, considering the security, the providers of some simulation models or services usually do not allow them migrating and being deployed free on the simulation platform, so as to isolate and publish them as the form of SOA. Then any of the simulation models or services can be classified as a separate resource group to reflect its isolated feature. In general, there is high communication efficiency in a resource group, such as the high performance

rIndex[i] = Į1 * rCPUNum[i] +Į2 * rMEMSize[i] //Į1 and Į2 are the weighting factors EndFor Rearrange the array rIndex in descending For j = 1 to N //N is the candidate number of the computer node pCPUIndex[j] = pCPUNum[j] * pCPUSpeed[j] * (1 – pCPUUtil[j]) pMEMIndex[j] = pMEMSize[j] * (1 – pMEMUtil[j]) End For For i = 1 to M For j = 1 to N If (rCPUNum[i] * pCPUSpeed[j] < pCPUIndex[j]) Then pIndex[j] = ȕ1 * pCPUIndex[j] //ȕ1 is the weighting factor

512

groups instead of the grid nodes. After selecting a resource group, we can use the above method to select specific resource instances in the resource group for the simulation task. 6. APPLICATION The resource management technologies of cloud simulation platform described in this paper have played an important role in the field of multi-disciplinary virtual prototype engineering in their preliminary application. Based on the cloud simulation platform COSIM-CSP (Bo Hu Li et al. 2009), the application about collaborative design and simulation of the landing gear system of an aviation aircraft is shown as follow. The multi-disciplinary virtual prototype of the landing gear system consists of several sub-system models, such as electronic control model, multi-body dynamics model and hydraulics model and so on. These models refer to various commercial software, such as control system design/simulation tools MATLAB / SIMULINK, dynamics system design/simulation tools MSC ADAMS, hydraulics system design/simulation tools MSC EASY5 and structure design tool CATIA. The top-level system model of the virtual prototype constructed by the COSIM-CSP was shown in Figure9.

ELSE pIndex[j] = ȕ1 * pCPUIndex[j] * į1 //į1 is the punishment factor End If If (rMEMNum[i] < pMEMIndex[j]) Then pIndex[j] = pIndex[j] + ȕ2 * pMEMIndex[j] //ȕ2 is the weighting factor ELSE pIndex[j] = pIndex[j] + ȕ2 * pMEMIndex[j] * į2 //į2 is the punishment factor End If End For Select the node j, the pIndex of which is max, as the target node for the requirement i pCPUIndex[j]

=

pCPUIndex[j]

-

rCPUNum[i]

*

pCPUSpeed[j] pMEMIndex[j] = pMEMIndex[j] - rMEMNum[i] End For

Where, F cpu: a diagonal matrix for computing performance

F Comm : a matrix for communication performance

X is a N dim row vector with the element 0 or 1 [x 1, x 2, Ă , x N ], x i = 1indicates that the resource group i was selected, x 1+ x 2 + Ă + x N = M, The equations would Have different return values for different X.

Figure 8: Resource Selection Model for Selecting Resource Groups In the cloud simulation platform, every user corresponds to a prior resource group for utilization according to his identity, which is configured by the administrators. When the amount or performance of specific resource in the resource group can not satisfy the requirement of collaborative simulation task for a user, cloud simulation platform will find the specific resource in other resource groups. However, the cloud simulation platform has the feature of large-scale virtualization, the communication between resource groups may be on the WAN or on the Internet. At this situation, communication efficiency should be considered as an important indicator. If there are suitable resource instances in more than one resource groups to meet with the remaining simulation requirement, which cannot be satisfy in the prior resource group, both computing performance and communication performance need to been considered. The RSM (Chang Feng Song, et al. 2009) in simulation grid mentioned above may be learned and continue to be adopted to select the resource group, shown in the Figure 8. What we need to do is just setting the weights of communication performance between the resource

Figure 9: Top-level System Model of the Virtual Prototype Constructed by the COSIM-CSP According to the formal description of unified modeling mentioned above, we can describe the requirements of various sub-system models of the virtual prototype as follow. - - Electronic Control Model Windows XP - MATLAB/SI MULINK 1 1 - Hydraulics

513

service; model resource includes top-level model, the electronic control model, multi-body dynamics model, the hydraulics model, structure model. In the process of implementation, the overall system designer and the sub-system designers complete the top-level system modeling and sub-system design respectively in their own virtual desktops, which are the research environment customed, and then upload the corresponding model files through the application portal of COSIM-CSP. Based on the resource management technologies mentioned above, COSIM-CSP constructs the running environment of the collaborative simulation dynamically, and builds the system-level virtual prototype automatically, so that achieving the aggregation and collaboration of the simulation resources. The process is shown in the Figure 11.

Model CentOS5.4 - MSC EASY5 1 1 - Multi-body Dynamics Model Windows XP - MSC ADAMS 1 1 - Structure Model Windows XP - CATIA 2 2

Figure 11: The Process of the Implementation 7. CONCLUSION AND FUTURE WORK The research of this paper proposes the guideline and the architecture for the resource management of the cloud simulation platform. In this framework, in order to support the centralized management for the life cycle of all kinds of simulation resources, we present the formal description of the unified modeling from the perspectives of resource operation and resource running. In the unified modeling, we propose clearly the concept of resource group, so as to distinguish the selection methods in the resource group and inter-groups. The resource management technologies of cloud simulation platform described in this paper have been verified in the collaborative design and simulation process for the multi-disciplinary virtual prototype. In future work, we will do more research on simulation performance evaluation, and enhancing the flexibility of resource selection, which will select the most appropriate resource instances instead of best performance ones. In addition, we need further research on the on-demand aggregation and high efficient collaboration for the simulation capabilities.

Figure 10: Virtual Resource Environment In the COSIM-CSP, the overall system designer and the sub-system designers can see the virtual resource environment shown in the Figure 10, including virtual computing environment, virtual storage space, collaborative scheduling / management service, virtual machine template resource, licensing resource and model resource. Specifically, virtual machine template resource includes templates installed COSIM top-level modeling software, MATLAB / SIMULINK software, MSC ADAMS software, MSC EASY5 software, CATIA software respectively; cooperative scheduling / management service includes execution management tool, collaboration middleware, remote visualization

ACKNOWLEDGMENTS This paper is supported by the National 973 plan (No. 2007CB310900).

514

REFERENCES LI Bo-hu, CHAI Xu-dong, HOU Bao-cun, et al., 2009. A Networked Modeling & Simulation Platform Based on the Concept of Cloud Computing – “Cloud Simulation Platform”. Journal of System Simulation 21(17): 5292-5299 . (in Chinese) Barham, P., et al., 2003. Xen and the art of virtualization. Proceedings of the 19th ACM Symposium on Operating Systems Principles, 164177. Oct. 2003, New York USA. Anderson, T., et al., 2005. Overcoming the Internet impasse through virtualization. IEEE Computer 38 (4): 34-41. Apache, 2009. Axis2/Java - Next Generation Web Services. Available from: http://ws.apache.org/ axis2/[2010]. Platform, 2007. Platform EGO white papers. Available from: http://platform.com/Products/platformenterprise- grid-orchestrator/whitepapers[2010]. Ganglia, 2010. Ganglia Monitoring System. Available from: http://ganglia.sourceforge.net/[2010]. Dahmann, J., et al., 1998. The DoD High Level Architecture: An Update. Proceedings of the 1998 Winter Simulation Conference, 797-804. 10 Dec. 1998, Washington, D. C., USA IEEE, 2008. Draft Standard for Modeling and Simulation (M&S) High Level Architecture (HLA)Object Model Template (OMT) Specification. Available from: http://ieeexplore.ieee.org/servlet/ opac?punumber=4478265 [2008]. IEEE, 2009. Draft Standard for Modeling and Simulation (M&S) High Level Architecture (HLA)Framework and Rules. Available from: http://ieeexplore.ieee.org/servlet/opac?punumber= 5347324 [2009]. IEEE, 2009. Draft Standard for Modeling and Simulation (M&S) High Level Architecture (HLA)Federate Interface Specification. Available from: http://ieeexplore.ieee.org/servlet/opac?punumber= 5347330 [2009]. Foster, I., 2002. The Grid: A New Infrastructure for 21st Century Sciece. Physical Today 55 (2): 42-47. Bo Hu Li, Xudong Chai, Baocun Hou, et al., 2006. Research and Application on CoSim (Collaborative Simulation) Grid. Proceedings of MS-MTSA2006, 156-163. July 2006, Alberta, Canada. Pearlman, L., et al. 2004. Distributed Hybrid Earthquake Engineering Experiments: Experiences with a Ground-Shaking Grid Application. Proceedings of 13th IEEE Symposium on High Performance Distributed Computing, 14-23. June 2004, Hawaii, USA. Einstein, A. 2005. FederationX: A Technical Brief. Available from: http://www.magnetargames.com/. Chang Feng Song, Bo Hu Li and Xudong Chai, 2009. Node selection in simulation grid. Journal of Beijing University of Aeronautics and Astronautics 35(1): 56-60. (in Chinese)

AUTHORS BIOGRAPHY Ting Yu Lin was born in 1984. He received his B.S.degree in BeiHang University. He is currently a Ph.D. candidate at the School of Automatic Science and Electrical Engineering, BeiHang University, Beijing, China. His research interests include multi-disciplinary virtual prototype and intelligent distributed simulation. Xudong Chai was born in 1969. He is a researcher and deputy director at Beijing Simulation Center of Second Academy of Aerospace Science & Industry Co. and council members of Chinese System Simulation Association and National Standardization Technical Committee. His research interests include automatic control and simulation. Bo Hu Li was born in 1938. He is a professor at School of Automatic Science and Electrical Engineering, BeiHang University, and Chinese Academy of Engineering, and the chief editor of “Int. J. Modeling, Simulation, and Scientific Computing”. His research interests include multi-disciplinary virtual prototype, intelligent distributed simulation and cloud manufacturing.

515

A FRAMEWORK FOR ENHANCED PROJECT SCHEDULE DESIGN TO AID PROJECT MANAGER’S DECISION MAKING PROCESSES Sanja Lazarova-Molnar(a), Rabeb Mizouni(b), Nader Kesserwan(a) (a)

Faculty of Information Technology, United Arab Emirates University, United Arab Emirates (b) Department of Software Engineering, Khalifa University, United Arab Emirates (a)

{sanja,nkesserwan}@uaeu.ac.ae, (b)[email protected]

the real needs of a project. The models developed still suffer from many limitations that often make them not representative to real world situations. Typically, project schedules are described in very strict terms, using Gantt charts or PERT. In real life, even small projects face risks and may, consequently, deviate from their original plans. As a consequence, even good projects can fail (Matta and Ashkenas 2003). For instance, in the software industry it has been reported (Denning and Riehle 2009) that approximately one-third of software projects fail to deliver anything, and another third deliver something workable but not satisfactory. In order to have a more realistic and effective project scheduling, management frameworks need to incorporate uncertainties on the one hand, and guide the managers to what actions to take when such uncertainties arise. This is the issue that we address in our paper, i.e. to lay out a strategy to create an optimal enhanced project schedule. Our objective is to answer the needs of managers by providing a framework that helps the generation of a more realistic and insightful project planning. The proposed framework supports flexible and efficient project schedule modeling and simulation. It combines: (a) a novel model for describing project schedules in a more realistic way, accommodating uncertainties, and (b) facilities for model’s simulation and assessment with respect to predetermined project goals. The objective of the framework is to provide managers with answers to the following types of questions:

ABSTRACT Good schedule increases the chances of a project meeting its goals. The most popular formalisms for describing project schedules are very rigid and inflexible in modeling changes due to uncertainties. In this paper we describe a framework to support enhanced project schedule design. The proposed framework is based on the Enhanced Project Schedule (EPS) model. In addition to an initial Gantt Chart, EPS allows definition of Remedial Action Scenarios (RAS), which contain guidelines of actions to consider when uncertainties arise. This creates a dynamic and evolving schedule. It is meant to guide the project manager in the decision making processes throughout the project implementation. The process of selection of the remedial action scenario is an optimization one, based on simulation. We illustrate the dynamics of the EPS design framework by an example. Keywords: project schedule, proxel-based simulation, remedial action scenarios, uncertainty 1. INTRODUCTION During the past few decades project management has evolved into a discipline that studies planning, scheduling and controlling of activities that directly contribute to the achievement of project’s objectives. The pressure of time-to-market along with the increasing complexity of present-day projects, have contributed project management to become one of the main factors for projects success. A growing number of companies use various advanced project management tools and methods to ensure the project quality expected by customers, delivered within reasonable deadlines and at the lowest possible cost. Many attempts have been conducted to improve the project scheduling prediction (Herroelen and Leus 2005; Arauzo, Galán et al. 2009; Huang, Ding et al. 2009; Jing-wen and Hui-fang 2009; Sobel, Szmerekovsky et al. 2009). Many of them are based on analytical models and simulation. Tools, such as Microsoft Project and Primavera Project Planner, are typically suggested to help managers in planning and controlling their projects. Existing frameworks and methods, however, fail, or, are insufficient; to answer

1) What is the best Remedial Action Scenario (RAS) to adopt if some uncertainties arise during the implementation of the project? 2) What are the features that can be implemented within the deadline of the projects? 3) What are the best and robust deadlines to consider that take into consideration the deviation from the original scheduling because of uncertainties. As shown in Figure 1, our framework is based on two main modules, and two supporting ones:

516

side a standard Gantt chart and an EPS, as shown in Figure 2. For comparison, Figure 2(a) illustrates a simple project schedule, modeled using a classical Gantt chart. The project schedule consists of four tasks (Task1, Taks2, Task3, and Task4) and two available teams (Team A and Team B). All tasks have predefined executors leading to one possible scenario of execution. Such model is in fact rigid and it is not able to anticipate the occurrence of any unpredictable events. Figure 2(b) illustrates the EPS. While having the same number of tasks and teams, two majors features are added:

1) Multi-RAS EPS Proxel-Based Simulator is a simulator based on the proxel-based simulation method. 2) Result Visualization Module: responsible for visualizing and interpreting the results of the simulation with respect to the goals specified by the manager. 3) User Interface Module that supports and facilitates the input of project schedules. 4) A Data Storage Module that manages project schedule data.

1) “floating task” (Task 2), which is a non-vital task that can be executed by any of the two teams, albeit with different duration distribution functions (based on teams’ expertise).

The rest of the paper is organized as follows. In the next section we describe the EPS model. In Section 3 we present the framework that supports the generation of EPS models and we describe the different modules, focusing on the key ones. Section 4 demonstrates the idea of the framework by an example. Finally, Section 5 concludes the paper.

2) fuzzily described guidelines, provided below the schedule, which are meant to accompany the project schedule as RAS (remedial action scenario).

2.

WHAT IS AN ENHANCED PROJECT SCHEDULE? To illustrate the concept of an enhanced project schedule, we provide an example that displays side-by-

Data Storage Deadlines Initial Plan Uncertainties PROJECT MANAGER

Proxel-Based Simulator

Results Visualization Module

Project Goals

User Interface ENHANCED PROJECT SCHEDULING FRAMEWORK Figure 1: EPS Design Framework Architecture

517

Simulation Results

(a) No additional rules

(b) EPS with a Remedial Action Scenario Mar 2010

Mar 2010

ID

ID

Task Name 1

1

Task 1

2

Task 3

3

Task 2

4

Task 4

2

3

4

5

6

7

8

Task Name

9

1

Team A

Team A

TeamB

Teams A and B

1

Task 1

2

Task 3

3

Task 2

4

Task 4

2

3

4

5

6

7

8

9

Team A

Team A, or B

TeamB

Teams A and B

1. If duration of Task 2 performed by team B is “very short” then start Task 3 by team B. 2. If duration of Task 1 is “too long” and it completes “shortly after” team B started to work on Task 3, then Task 3 is cancelled and both teams start working on Task 4. Figure 2: Illustration of: a) Classical Gantt Chart and b) Enhanced Project Schedule with RAS

In our previous work (Lazarova-Molnar and Mizouni 2010; Lazarova-Molnar and Mizouni 2010) we successfully modeled and simulated the type of scenarios described in Figure 2(b). There, we also developed an approach to analyze and simulate the effects of the uncertainties and remedial actions on the duration of project. As expected, on-the-fly decisions make a significant difference in the duration of the project and need to be considered and, if possible, predetermined. To account for resource re-allocations we have also defined a new type of tasks, which we termed as “floating task”. This task was a typically a noncrucial task for the success of the project, which could be implemented by a number of teams, albeit with different duration distribution functions, and based on their availabilities.

Gantt Chart

Remedial Action Scenario (RAS)

Enhanced Project Schedule

Figure 3: Enhanced Project Schedule Components Let us take the example of a simple software development project schedule, subject to various uncertainties. The enhanced schedule would consist of a Gantt chart, where each task that corresponds to a requirement implementation, is associated with a probability distribution function for its duration, as well as a set of fuzzy rules that describe the remedial strategy under certain conditions (e. g. if task A finishes in a very short time than proceed to task B, else skip task B). The set of fuzzily specified guidelines are obtained by simulating a set of possible RAS, and accordingly selecting the most optimal one (similar to the simple example presented in Figure 2(b)). Once the simulation of the chosen RAS provides good results with respect to project goals (e.g. complete as many tasks as possible, complete in as short time as possible, or minimize budget), the resulting EPS is communicated to the project manager to aid his/her decision making process. We see the fuzziness as a great advantage as it leaves a certain degree of freedom to the project manager as well, to involve his/her knowledge/perceptions he/she might have. The proxel-based simulation method allows for a great flexibility in schedule description and provides solutions taking into account anticipated uncertainties. This helps us in picking the best RAS that specifies the

2.1. EPS Model Description We propose the definition of a schedule to include the uncertainties that can arise and their quantification using statistical probability distributions. In addition to this, we formalize the remedial actions that managers can take. Every schedule along with the set of remedial actions (RAS) creates what we term as: enhanced project schedule (as shown in Figure 3). The RAS consists of a set of fuzzy if-then production rules. These rules make the project evolving and thus, the sequencing of tasks, dynamic and changing. Once the enhanced project schedule is designed, we simulate each RAS using the proxel-based method and pick the best one based on the success criteria for the project. “The probability that the project is delivered before deadline” and “the probability that the project is implemented within this budget” are examples of success criteria.

518

evaluation. This makes it straightforward to compare various RAS, as well as test for the best RAS to counteract uncertainties, as described by ‫ܨ‬. Note that F can be an empty set too, which would imply sticking to the original project schedule provided by IGC. Once an optimal remedial action scenario is selected, it is associated with the initial project schedule and handed to the project manager as a decision making aid. The goal of the proposed framework is to support the generation of the EPS. This process is further demonstrated by a simple example in Section 4.

optimal set of recommended remedial actions when uncertainties occur. 2.2. EPS Formal Model An Enhanced Project Schedule (EPS) is described as follows: ‫ ܵܲܧ‬ൌ ሺ‫ܣ‬ǡ ܲǡ ܶǡ ‫ܦ‬ǡ ܹǡ ‫ܨ‬ǡ ‫ܥܩܫ‬ሻ x x

x x

x x x

‫ ܣ‬ൌ ሼ‫ܣ‬ଵ ǡ ‫ܣ‬ଶ ǡ ǥ ǡ ‫ܣ‬௡ ሽ, set of tasks, where each task corresponds to a task in the project schedule, ܲ ൌ ሼܲଵ ǡ ܲଶ ǡ ǥ ǡ ܲ௠ ሽ, set of precedence constraints, that are actually tuples of two tasks where the completion of the first one is a prerequirement for commencing the second one, e.g. ሺ‫ܣ‬௫ ǡ ‫ܣ‬௬ ሻ would mean that completing of ‫ܣ‬௫ is a pre-requirement for beginning ‫ܣ‬௬ , ܶ ൌ ሼܶଵ ǡ ܶଶ ǡ ǥ ǡ ܶ௟ ሽ, set of teams available for the execution of the project, ‫ ܦ‬ൌ ሼ݀ଵ ǡ ݀ଶ ǡ ǥ ǡ ݀௦ ሽ, set of probability distribution functions that correspond to duration of tasks performed by the competent teams, ܹ ൌ ሼ‫ݓ‬ଵ ǡ ‫ݓ‬ଶ ǡ ǥ ǡ ‫ݓ‬௧ ሽ, set of mappings of distribution functions to competent teams and tasks, ‫ ܨ‬ൌ ሼ݂ଵ ǡ ݂ଶ ǡ ǥ ǡ ݂௥ ሽ, set of fuzzy rules that define the remedial action scenario, ‫ ܥܩܫ‬െ ‫ݐݎ݄ܽܥݐݐ݊ܽܩ݈ܽ݅ݐ݅݊ܫ‬, initial sequencing of tasks that satisfies the set of precedence constraints provided by ܲ,

2.3. Multi-RAS EPS To facilitate the generation of the optimal EPS, the framework needs to analyze a number of various RAS that could potentially accompany a given EPS. For this purpose, we define the term Multi-RAS EPS (MEPS) as follows: ‫ ܵܲܧܯ‬ൌ ሺ‫ܣ‬ǡ ܲǡ ܶǡ ‫ܦ‬ǡ ܹǡ ‫ܨ‬Ԣǡ ‫ܥܩܫ‬ሻǡ where the only difference to the standard EPS is that ‫ ܨ‬ᇱ ൌ ሼ‫ܨ‬ଵ ǡ ‫ܨ‬ଶ ǡ ǥ ǡ ‫ܨ‬௤ ሽ represents a set of RAS, where each ‫ܨ‬௜ ǡ ݅ ൌ ͳǤ Ǥ ‫ݍ‬ǡ is a defined as a set of fuzzy rules. This defines the central input to the framework. 3. THE EPS DESIGN FRAMEWORK Main functionalities of the proposed framework for EPS design are the following:

whereܲ ‫ ܣ ك‬ൈ ‫ܣ‬, and ܹ ‫ ܣ ك‬ൈ ‫ ܯ‬ൈ ‫ܦ‬. Also, ‫ ܣ‬ൌ ‫ܣ‬௖ ‫ܣ ׫‬௡ , where ‫ܣ‬௖ is the set of cancelable tasks and ‫ܣ‬௡ is the set of non-cancelable tasks. Cancelable task is a task that is non-vital for the success of the project, and thus, not compulsory, however, useful for the value of the project. Non-cancelable tasks are the ones that are crucial for the success of the project. This differentiation is important for the realistic simulation of project schedules. Each fuzzy rule is made up of two parts: condition and action, formally expressed as “ܿ‫ฺ ݊݋݅ݐ݅݀݊݋‬ ܽܿ‫”݊݋݅ݐ‬. Conditions can be described either by using strict terms, or fuzzy ones. An action can typically be canceling or interrupting some of the tasks, or one of the various types of rescheduling. This is the fact that makes our schedule description evolving, rather than rigid and inflexible. Two examples of fuzzy rules are:

1.

Support the expressive description of EPS,

2.

Support definitions of project goals (e.g. minimize duration; maximize number of completed tasks; etc.),

3.

Run simulations for a single project schedule in combination with a number of RASs (MultiRAS EPS), and

4.

Select the RAS that best meets the specified project goals to accompany the initial project schedule to yield the final EPS.

The simulation method of choice is the proxelbased simulation (Horton 2002; Isensee, LazarovaMolnar et al. 2005) as it is highly flexible and provides high accuracy. Its additional advantage is that the simulation is carried out directly, based on the user model, i.e. EPS, without building the state space prior to that. The resulting, simulation-based calculated, optimal EPS will definitely take into account many of the uncertainty factors, thus reducing the risk in the project. In addition, it will provide managers with more insight and guidance when making decisions during project’s implementation. A high-level diagram of the data-flow process that underlies the EPS supporting framework is presented in Figure 4. It shows that the inputs to the program are the Multi-RAS EPS and the Project Goals, and it produces an Optimal EPS as a final product. In the following, we provide detailed description of the most complex module of the framework, i.e. the

‫ܣ‬௫ ‫ܣ݈݁ܿ݊ܽܿ ֜ ݃݊݋݈݋݋ݐݏ݁݇ܽݐ‬௬ or ‫ܣ‬௫ ܿ‫ܣݎ݁ݐ݂ܽݕ݈݇ܿ݅ݑݍݏ݁ݐ݈݁݌݉݋‬௬ ֜ ݈ܿܽ݊ܿ݁‫ܣ‬௭ . Both are examples for typical proceedings during project execution. However, in our approach we formalize their modeling, assessment and quantitative

519

EPS Proxel-Based Simulator as previously mentioned in Section 2, and a brief description of the remaining modules.

Multi-RAS EPS Project goals

EPS Proxel-based Simulator

x

Team information,

x

Distribution functions in use,

x

Mappings of distribution functions to teams and tasks,

x

Multiple RAS,

x

Deadline, and

x

Initial state.

The proxel-based simulation of a given project schedule in combination with each provided RAS is the core element of the tool. Algorithm 1 provides more details of how this is performed. It describes the dynamics of the proxel-based simulation for a single-RAS EPS. This is further repeated for each provided RAS. The basic computational unit, i.e. the proxel, for each EPS is formed based on the information in the input file. The general simplified proxel format is the following:

Results Visualization Module Optimal EPS

ܲ‫ ݈݁ݔ݋ݎ‬ൌ ሺܵ‫݁ݐܽݐ‬ǡ ‫ݐ‬ǡ ܲ‫ݎ‬ሻ where:

Figure 4: High-level diagram of the framework

ܵ‫ ݁ݐܽݐ‬ൌ ሺܶܽ‫ݎ݋ݐܸܿ݁݇ݏ‬ǡ ‫ݎ݋ݐܸܿ݁݁݃ܣ‬ǡ ‫ݏ݇ݏܽܶ݀݁ݐ݈݁݌݉݋ܥ‬ሻ, and

3.1. EPS Proxel-Based Simulator Module The Proxel-Based Simulator is the key-module of the framework. This is the module that performs simulation of the provided Multi-RAS EPS. During simulation, statistics that correspond to project’s goals are collected. As previously stated, our simulation method of choice is the proxel-based simulation (Lazarova-Molnar 2005). The proxel-based method is a simulation method based on the method of supplementary variables (Cox 1955). It was introduced and formalized in (Horton 2002; Lazarova-Molnar 2005). The advantages of the proxel-based method are its flexibility to analyze stochastic models that can have complex dependencies and the accuracy of results, which is comparable to the accuracy of numerical solvers (Stewart 1994). The proxel-based method expands the definition of a state by including additional parameters which trace the relevant quantities in one model following a previously chosen time step. Typically this includes, but is not limited to, age intensities of the relevant transitions. The expansion implies that all parameters pertinent for calculating probabilities for future development of a model are identified and included in the state definition of the model. In order to apply the proxel-based simulation algorithm, this module needs to process the information contained in the input file, i.e. the Multi-RAS EPS. In summary, it contains the following information: x

Maximum simulation time,

x

Time step,

x

Task information,

x x x x x

Task Vector is a vector whose size is equal to the number of teams available and records the task that each team is working on, Age Vector tracks the length that each team has been working on the task specified in the Task Vector, correspondingly, Completed Tasks stores the set of completed tasks, t is the time at which the afore-described state is observed, and Pr stores the probability that the schedule is in the afore-specified state at time t.

Algorithm 1 demonstrates the on-the-fly building of the state-space of the project schedule model. Thus, there is no need for any pre-processing to generate the statespace. It is directly derived from the input file specification. The initial state proxel is derived from the initial state that is specified in the input file as well. The algorithm operates by using two interchangeable data structures, Proxel_Tree[0] and Proxel_Tree[1], that store the proxels from two subsequent time steps (regulated by the switch variable). If two proxels represent the same state, there is only one proxel stored, and their corresponding probabilities are summed up.

520

EPS Design Framework File

Results

About Change simulation parameters:

Minimize project duration

Maximum simulation time:

Maximize number of completed tasks

Time step:

Maximize team utilization

Add Custom Goal

Simulate

Figure 5: EPS Design Framework GUI prototype

Algorithm 1: Proxel-based simulation of enhanced project schedules Input: EPS, Project Goals Output: Simulation Results switch = 0 insert Initial State Proxel in the Proxel_Tree[switch] switch = 1 - switch while (maximum simulation time has not been reached) { px = get_proxel(Proxel_Tree[switch]); for (each task in the Task Vector(px)) { check task precedence & team availability; generate next state S; compute probability for S in computed_prob search for the S in the Proxel_Tree[1-switch]; if (S found) { px1 = found_proxel(S); probability(px1) = (probability(px1) ) + computed_prob; } else { generate new proxel px2(S); insert proxel in Proxel_Tree[1-switch]; } delete px from Proxel_Tree[switch]; increase simulation time by one time step; calculate statistics with respect to project goals; switch = 1- switch; } }

3.2. Supporting Modules The remaining modules, i.e. the Results Visualization Module and the two supporting ones, are trivial. The Visualization Module is charting the (transient or steady-state) solutions of the simulation with respect to project goals. An example of such solution is provided in the following Section 4. The Graphical User Interface Module facilitates file-based and graphical input of Multi-RAS EPS, along with the set of project goals. The project goals are meant to be selected from a list of most commonly used ones. The list would include project goals as: x minimize duration, x maximize number of tasks completed, etc. as well as allow the user to specify custom goal by using a scripting language. To illustrate our idea, a prototype of the framework GUI is shown in Figure 5. The Data Storage Module ensures efficient memory manipulation and stores the statistics and intermediate solutions of the simulation experiments. 4. EXPERIMENTS To demonstrate the proposed framework, we demonstrate the processing of an example Multi-RAS EPS. The example EPS contains 4 tasks, identified as: Task 1, Task 2, Task 3, and Task 4. Each task can be performed by one of the two teams: Team A or Team B. Tasks 1, 2 and 3 have fixed human resource allocation, i.e. performing team, and Task 4 is a cancelable floating task, and can be performed by either team A or B. The initial Gantt chart (IGC) of the sample project schedule is shown in Figure 6, where the green-colored tasks are cancelable and the team capable of carrying out task is labeled on the task itself. In addition to this the project schedule has a predefined deadline '.

521

Mar 2010

ID

Task Name 1

1

Task 1

2

Task 3

3

Task 2

4

Task 4

2

3

4

5

6

7

8

#tasks 1 Task1 () 10 false false 2 Task2 () 05 false false 3 Task3 (1) 10 true false 4 Task4 () 10 true true #teams 10 A 20 B #distributions 100 U (2.0,10.0) 200 N (7.0,1.0) 300 U (2.0,8.0) 400 W (3.5,1.5) 500 U (2.0,5.0) #ttd 1 10 100 2 20 200 3 20 300 4 10 400 4 20 500 #ras (a) 1 fuzzy1 (11.25,15.0) C 3 1 fuzzy1 (11.25,15.0) C 4 2 fuzzy1 (11.25,15.0) C 3 2 fuzzy1 (11.25,15.0) C 4 (b) 3 non_avail (20) C 3 (c) #deadline 20 #initial (1,2)

9

Team A Team B

Team B

Team A or B

Figure 6: Example of an Initial Gantt chart of the example project schedule (the green-colored tasks are cancelable) The multi-RAS enhanced project schedule features three RAS. In our case we choose among the two remedial action scenarios (a and b) and the default sequencing (seen as an empty set of fuzzy rules, c), which are defined as follows: a) If the duration of tasks 1 or 2 is close to the deadline ', then do not start working on any of the tasks 3 or 4 and do not interrupt the other team if they have already started to work on either of the latter two tasks. b) If the team assigned to a certain non-floating and cancelable task (Task 3 in our case) is unavailable at the time it can be initiated, then cancel the task. c)

No guidelines are provided and the manager is instructed to follow the original schedule.

The performance measure, according to which we assess the three RAS, is: x

Figure 7. Example Input File

“the probability of completing the project before the deadline”

The input file consists of: 1) Definition of all tasks and their duration probability distribution function, 2) Definition of all fuzzy functions in use, 3) Deadline of the project, and 4) Initial state(s)

The project goal that is supported by this performance measure is defined as “Complete the project before the deadline”. The simulation targets to discover the RAS that yields the best performance, given the constraints of the initial project schedule. For this purpose, the EPS Proxel-Based Simulator Module runs the proxel-based simulation that collects the statistics that answer our question.

The input file contains all parameters that are listed in Section 3.1. In the example case there are 4 tasks. Each task is specified by the following parameters: Task ID, Task Name, Preceding Tasks, Cancelable, and Floating, specified in the same order. Each team is specified by: Team ID and Team Name. Distributions are specified by: Distribution ID, Distribution Type, and Parameters. Each team-task-distribution mapping contains three values, i.e. the id’s of the team, task and distribution that are connected to form the mapping. As specified in the input file, the values of the parameters of the duration distribution functions are:

4.1. Input File Specifications In the following we describe and explain the input file specification of the example model, which is shown in Figure 7.

522

x x x x

This would create correspondingly:

Duration of Task 1 ~ Uniform (2.0, 10.0) Duration of Task 2 ~ Normal (7.0, 1.0) Duration of Task 3 ~ Uniform (2.0, 8.0) Duration of Task 4, performed by: o Team A ~ Weibull (3.5, 1.5) o Team B ~ Uniform (2.0, 5.0)

3) (((1, 2), ('t, 't), ()), 0, 1-p1-p2). Note that the age intensities for each task that is still being worked on in the next time step are updated accordingly.

0, t  a ­ °t  a , a d t d b. ® b  a ° 1, t ! b ¯

4.3. Experimental Results In the following we present the results of the simulation of our model. The proxel-based simulation provides complete results for any quantity of interest; in this case it is the probability function of the duration of the project (shown in Figure 8). The performance measures are not limited to this and it is provided for illustration only. In general, they can include any quantities that are relevant to project’s goals. Simulation results provide us with an overview to aid the selection of the most suitable remedial action scenario with respect to project’s goals. From the simulation results we can see that the best RAS that yields the highest probability of having the project completed before the deadline is RAS (a), closely followed by (b). Judging from this, we can conclude that the RAS (a) seems most favorable for this enhanced project schedule. Also, we can clearly see that the rigid RAS (c) which does not allow for any changes has the lowest probability of having the project completed before the deadline. 5. CONCLUSIONS The framework presented in this paper demonstrates a vision and its implementation strategy of how to create a better project schedule, one that will provide more information and decision making guidance to project managers. This is what we term as EPS Design Framework. The output of the framework, i.e. the resulting EPSs is planned to be further used to support project managers in the decision making processes throughout project implementation. Decisions made in this way, based on the RAS recommendations that accompany the optimal EPS will not be solely based on human judgment, as it is the case in classical approach, but also based on sound models and their analysis. In this paper we present the details of the framework and, in particular, the details of its core module, i.e. the Proxel-Based Simulator, for which we present the modified proxel-based simulation algorithm. We believe that this is an effective way of designing schedules and it enhances the classical project schedule by allowing all available information to be utilized. Instead of having a static schedule, the remedial action scenarios make the schedule dynamic and evolving. In addition, our framework provides support for managers to incorporate their knowledge within the project schedules by simulating various possible RAS when new uncertainties occur. The

4.2. Proxel-Based Simulation Details In the following we provide some insight in the proxelbased simulation of our example model to illustrate the simulation method. The proxel-based simulation of the EPS commences with the initial state, as specified in the input file. This would create the following initial proxel: (((1, 2), (0, 0), ()), 0, 1.0). In the next time step, one of the three developments could be seen: 1) Team A completes working on Task 1, implying that it can start working on Task 4, as the only possibility

2) Team B completes working on Task 2, and x

implying that it can start working on Task 4, as the only possibility

3) Both teams continue corresponding tasks.

working

on

proxels,

2) (((1, 4), ('t, 0), (2)), 0, p2), and

As specified in the input file, the concrete fuzzy membership function is P (t ,11.25,15.0) . The symbol C stands for “cancel” and the subsequent number specifies the task id of the task to be canceled. The non_avail function evaluates to true/false depending on the availability of team B (specified by its id, i.e. 20 as a parameter). The framework allows custom specification of the fuzzy functions and actions. It is also extendable as to the type of actions that can be taken. Currently it features only “cancel”. Finally the pre-determined deadline of 20 time units (as an important parameter) and the initial state of the EPS are provided. According to the latter one, the project begins by Team A working on Task 1 and Team B working on Task 2. In addition to the EPS model specification, inputs to the framework are the simulation parameters (size of the time step and maximum simulation time) and project goals.

x

following

1) (((4, 2), (0, 't), (1)), 0, p1),

Next, the definition of the RAS is provided. The fuzzy membership function that defines fuzzy1 is defined in the framework as follows:

P (t , a, b)

the

the

523

EPS Design Framework File

Results

About

Results for goal: Minimize project duration

Change simulation parameters: Minimize project duration

Maximum simulation time:

Maximize number of completed tasks

Time step:

Maximize team utilization

Add Custom Goal

Simulate

Figure 8: Probability of having the project completed for the three possible RAS

framework encourages analysis and deep thinking about project plans and, hence, supports the creation of a higher quality initial plan, identified as one of the key success factors for projects. As part of our future work agenda, we plan to extend the capabilities of our simulation approach to handle multi-project resource sharing.

Isensee, C., S. Lazarova-Molnar, et al. (2005). "Combining Proxels and Discrete Phases." Proceedings of ICMSAO. Jing-wen, Z. and S. Hui-fang (2009). Multi-Mode Double Resource-Constrained Time/Cost Trade-Offs Project Scheduling Problems. International Conference on Management and Service Science, 2009. MASS '09. . Lazarova-Molnar, S. (2005). The proxel-based method: Formalisation, analysis and applications, Ottovon-Guericke-Universität Magdeburg, Universitätsbibliothek. Lazarova-Molnar, S. (2005). The Proxel-Based Method: Formalisation, Analysis and Applications. Faculty of Informatics. Magdeburg, University of Magdeburg. Ph.D. Lazarova-Molnar, S. and R. Mizouni (2010). Floating Task: Introducing and Simulating a Higher Degree of Uncertainty in Project Schedules, IEEE. Lazarova-Molnar, S. and R. Mizouni (2010). "Modeling Human Decision Behaviors for Accurate Prediction of Project Schedule Duration." Enterprise and Organizational Modeling and Simulation: 179-195. Matta, N. F. and R. N. Ashkenas (2003). "Why good projects fail anyway." Harvard Business Review 81(9): 109-116. Sobel, M. J., J. G. Szmerekovsky, et al. (2009). "Scheduling projects with stochastic activity duration to maximize expected net present value." European Journal of Operational Research 198(3): 697-705.

REFERENCES Arauzo, J. A., J. M. Galán, et al. (2009). "Multi-agent technology for scheduling and control projects in multi-project environments. An Auction based approach." Inteligencia Artificial 42: 1220. Cox, D. R. (1955). "The analysis of non-Markovian stochastic processes by the inclusion of supplementary variables." Proceedings of the Cambridge Philosophical Society 51(3): 433441. Denning, P. J. and R. D. Riehle (2009). "The profession of IT Is software engineering engineering?" Communications of the ACM 52(3): 24-26. Herroelen, W. and R. Leus (2005). "Project scheduling under uncertainty: Survey and research potentials." European Journal of Operational Research 165(2): 289-306. Horton, G. (2002). "A new paradigm for the numerical simulation of stochastic Petri nets with general firing times." Proceedings of the European Simulation Symposium. Huang, W., L. Ding, et al. (2009). Project Scheduling Problem for Software Development with Random Fuzzy Activity Duration Times, Springer.

524

Stewart, W. J. (1994). Introduction to the Numerical Solution of Markov Chains., Princeton University Press.

AUTHORS BIOGRAPHY SANJA LAZAROVA-MOLNAR is an Assistant Professor in Computer Science at United Arab Emirates University. She received her Diploma in Computer Science from the “Sts. Cyril and Methodius” University in Macedonia and her M.Sc. in Computational Visualistics and Ph.D. in Computer Science from “Ottovon-Guericke” University of Magdeburg in Germany. Her main research interests are simulation and modeling, with the main focus on their methodology for achievement of more accurate and valid results. Her email is [email protected]. RABEB MIZOUNI is an assistant professor in the Department Software engineering at Khalifa University, Abu Dhabi, UAE. She got her PhD and her MSc in Electrical and Computer Engineering from Concordia University, Montreal, Canada in 2007 and 2002 respectively. After her graduation, Rabeb joined SAP research labs in Montreal where she acquired experience in the development of mobile applications. Her research interests include modeling and simulation of software requirements to improve the development of robust project schedules, development of mobile applications, and web services. Her email is [email protected]. NADER KESSERWAN got his Master Degree in Computer Science from McGill University. Montreal, Canada. He joined the college of Information Technology as Instructor in 2001. Before joining the academia, he acquired experience in diverse fields of computer science, simulation and real time systems. He spent two years at CAE Company, Canada working on a Flight Simulator, and on testing the integration of its software. Also, he had worked two years on games development for Disney Studio, Canada and USA. He taught programming courses and software engineering at McGill University, Canada. His experience includes team management, system analysis, software development and project management. His email is [email protected].

525

A NOVEL APPROACH TO REALISTIC MODELING AND SIMULATION OF STATE-VARYING FAILURE RATES Sanja Lazarova-Molnar Faculty of Information Technology, United Arab Emirates University, United Arab Emirates [email protected]

increase with each time it breaks). This is what we term as a state-varying failure. According to a study of medical equipment (Baker 2001) it was shown that there was a decreasing hazard of (first) failure after repair for some types of equipment. The interpretation was that it is a consequence of imperfect or hazardous repair, and also, because of differing failure rates among a population of machines. Likewise, in (Liberopoulos and Tsarouhas 2005) a pizza production line is studied and it was found that most of the failures have a decreasing failure rate because proactive maintenance improves the operating conditions at different parts in the line, and a few failures have an almost constant failure rate. It was also concluded that the longer the time between two failures, the more problems accumulate, and therefore, it takes longer time to fix the latter failure. It also suggests that the more time the technicians spend fixing a failure, the more careful job they do, and therefore, the time period until the next failure is longer. This is a very interesting observation that calls for state-varying failure rates and it can be addressed using our approach. These are some examples that show that failures need to be described more realistically to obtain accurate and useful simulation results. Unfortunately, this has very rarely been the case. Our goal is to provide a deterministic approach to analyze systems that exhibit not only time-, but more importantly, state-varying failure rates. For this we use method of proxel-based simulation, which based on our previous experience, is highly adjustable to treat these complex activities. In (Lazarova-Molnar 2008) we have analyzed and described state-dependent transitions and used proxel-based simulation for their analysis. These are the types of transitions that correspond and can be used to describe state-varying failure rates. Thus, in addition to the simulation approach, this paper provides a concept of how to model this type of failure rates and what changes need to be undertaken in the standard stochastic Petri net (SPN) models to introduce them. The paper is organized as follows. In the subsequent section we describe the state-varying failure rates, along with an introduction to the proxel-based simulation method. Further, we provide a concept for modeling state-varying failures using SPN. Next, we

ABSTRACT There has been a lot of research on time-varying failure rates, which deems constant failure rates as inadequate to model failures accurately. However, besides time, failure rates can also be affected by the state of the system (or its history, in terms of sequences of states and events that it has been through). In our paper we define several classes of state-varying failure rates and extend the formalism of Petri nets to model them. We further use the flexibility of the proxel-based method to accurately analyze behavior of systems that incorporate these kinds of failures. To illustrate our approach and study the effect of such dependencies, we compare simulation results for two models: one that exhibits state-varying failure rates, and another that only contains predefined failure rate functions. Keywords: state-varying failure rates, reliability, proxels, Petri nets 1. INTRODUCTION There has been an intensive research on time-varying failure rates, including their significant impact on reliability (Hassett, Dietrich et al. 1995; Retterath, Venkata et al. 2005; Zhang, Cutright et al. 2010), which have been defined as such almost two decades ago (Billinton and Allan 1992). Recently, Xie developed an analytical model of unavailability due to aging failures too (Xie and Li 2009). Since long time ago it has been shown that constant failure rates are inadequate for describing systems’ failures (Proschan 1963). Nevertheless, they are still widely used due to the fact that the methodology for their analysis is less complex and more accurate. The popular MTTF (meantime to failure) measure is still a widely used one (Coskun, Strong et al. 2009; Sharma, Kahlon et al. 2010), even though it has been deemed many times as inadequate (Schroeder and Gibson 2007). We go one step further as to claim that even time-varying failure rates are not sufficient, as in many cases the rates completely change their functions based on the occurrence of some events or based on the complete state of the system (e.g. a part has been replaced by a new one that is based on a new technology, or if a mechanical part has been physically broken, then it is logical that the failure rate would

526

On the opposite, it allows for a general class of stochastic models to be analyzed regardless of the involved probability distribution functions. In other words, the proxel-based method combines the accuracy of numerical methods with the modeling power of discrete-event simulation. The proxel-based method is based on expanding the definition of a state by including additional parameters which trace the relevant quantities in one model through a previously chosen time step. Typically this includes, but is not limited to, age intensities of the relevant transitions. The expansion implies that all parameters pertinent for calculating probabilities for the future development of a model are identified and included in the state definition of the model. Proxels (stands for probability elements), as basic computational units of the algorithm, follow dynamically all possible expansions of one model. The state-space of the model is built on-the-fly, as illustrated in Figure 1, by observing every possible transiting state and assigning a probability value to it (Pr in the figure stands for the probability value of the proxel). Basically, the state space is built by observing all possible options of what can happen at the next time step. The first option is for the model to transit to another discrete state in the next time step, according to the associated transitions. The second option is that the model stays in the same discrete state, which results in a new proxel too. Zero-probability states are not stored and, as a result, no further investigated. This implies that only the truly reachable (i.e. tangible) states of the model are stored and consequently expanded. At the end of a proxel-based simulation run, a transient solution is obtained which outlines the probability of every state at every point in time, as discretized through the chosen size of the time step. It is important to notice that one source of error of the proxel-based method comes from the assumption that the model makes at most one state change within one time step. This error is elaborated in (Lazarova-Molnar 2005). Each proxel carries the probability of the state that it describes. Probabilities are calculated using the instantaneous rate function (IRF), also known as hazard rate function. The IRF approximates the probability that an event will happen within a predetermined elementary time step, given that it has been pending for a certain amount of time W (indicated as ‘age intensity’). It is calculated from the probability density function (f) and the cumulative distribution function (F) using the following formula:

present an example model which we use to demonstrate our approach and we run experiments based on it. Finally, we present the results of the experiments with a discussion and conclusions. 2. PRELIMINARIES 2.1. State-varying Failures It is a common observation that a failure rate cannot simply be described by one function during its entire lifetime. Even more, failure rates in reality can change not solely based on time (Retterath, Venkata et al. 2005), but also based on the occurrence of certain events in the system (e.g. replacing the service person by another one which fixes them in a different manner, i.e. more thoroughly would influence the failure rate function). We refer to these types of failures as statevarying failure rates. Description of failure rate functions of statevarying failure rates is a complex process and would require an algorithmic description to supplement the graphical model. To illustrate it, one such description may be: If machine is repaired by repairman A Then the failure rate function ~ Normal(a, b) Else if machine is repaired by repairman B Then the failure rate function ~ Normal(c, d) If we add another factor to this, i.e. the age of the machine, and then the description would change to: If machine is repaired by repairman A Then the failure rate function ~ Normal(f(t), b) Else if machine is repaired by repairman B Then the failure rate function ~ Normal(g(t), d) where t is the age of the machine (which can easily be exchanged to represent the number of failures or any other relevant quantity). This observation is more general than the one that uses fixed failure rate functions, and as such, more realistically models the phenomenon of a machine that exhibits failures. Obviously, these models would need more advanced (or extended) modeling formalisms to be described. Thus, we extend stochastic Petri nets to account for the state-varying rates. Finally, to show the difference and compare the effects of such (even very small) dependencies, we compare the simulation results for two models: one that exhibits state-varying failure rates, and another, similar and over-simplified one, that only contains predefined failure rate functions with fixed parameter values.

P (W) =

2.2. Proxel-based Simulation The proxel-based method (Horton 2002; LazarovaMolnar 2005) is a relatively novel simulation method, whose underlying stochastic process is a discrete-time Markov chain (Stewart 1994)and implements the method of supplementary variables (Cox 1955). The method, however, is not limited to Markovian models.

f (W ) 1  F (W )

(1)

As all state-space based methods, this method also suffers from the state-space explosion problem (Lin, Chu et al. 1987), but it can be predicted and controlled by calculating the lifetimes of discrete states in the model. In addition, its efficiency and accuracy can be

527

Typically, they are introduced by extending the basic SPN with additional places and transitions that enable the tracking, as shown in Figure 2. However, to record quantities, such as the duration of the last repair (type (c)), we introduce a novel element which we term as tracking variable (TV), and it is represented by a hexagon in the SPN graphical model. TVs are connected by diamond-shape-ended arrows to the transitions for which they record the last firing time. To summarize, the extension is at both the level of the SPN formalism, and at the Petri net model itself, which is enriched by a number of extra places and transitions to ensure the tracking of relevant rewards. As for the SPN formalism: we extend it by the new element TV, and, in order to account for the statevarying transitions, we allow distributions to have discrete states, i.e. markings, as parameters of the distribution functions that control firing of transitions. In the following, we show by example how a SPN can be extended to allow the tracking of the various relevant quantities.

further improved by employing discrete phases and extrapolation of solutions (Isensee and Horton 2005). More on the proxel-based method can be found in (Lazarova-Molnar 2005). Initial state of the system Pr = 1.0

t=0

What can happen next? iti

ns

e

bl

si

os

s on

tra

lp

fo

t = 't

System can transit to another discrete state If transition is race enable, reset corresp. age variable, else advance it by 't Pr = p1

System can stay in the same discrete state Advance all age variables by 't Pr = 1 - p1

What can happen next?

What can happen next?

...

...

...

t = 2't

l ra

Figure 1: Illustration of the development of the proxelbased simulation algorithm

Last A

TV1

Extension SPN for tracking the#FAILURES relevant quantities Last B

3. MODELING STATE-VARYING FAILURES According to the performed observation, studies and research, we identify several classes of state-varying failure rates, i.e. failure rates that depend on: a) the number of failure occurrences up to the observed point in time, b) the age of the machine up to the observed point in time, c) the duration of the last repair, d) the time between the last two failures, e) the properties of the repair facilities, introduced as additional parameters, and f) the types of failures that have occurred. We allow a combination of a number of these factors to occur in our sample model to illustrate their effects through the proxel-based simulation analysis. Proxelbased simulation can easily be applied to analyze a model that exhibits any combination of them, as well as other types of dependencies on quantities that are part of the model. In the following, we will provide the details of the formal classification of the state-varying failures and our simulation approach. This will be further demonstrated using an example model.

TV2 Who ffail ail

FFAILED AILED

OK

repair repai ir

Basic SPN change change shift shift Technician B Technician A

change change shift shift

Figure 2: Illustration of the extended SPN model

3.2. Petri Net Specifications In the following we provide the formal definition of the extension of the SPN to account for the state-varying failure rates. Each Petri net SPN is defined as:

3.1. Formal Model of State-Varying Failure Rates The underlying discrete stochastic model that exhibits state-varying failure rates is described using a stochastic Petri net (SPN) (Bause and Kritzinger 2002). Nevertheless, we further extend the basic description of SPN to allow the tracking of the relevant rewards. Those are the quantities that are in fact parameters of the distribution functions of the timed transitions, besides the age intensities of the relevant transitions.

SPN = (P, T, A, G, TV, m0) where:

528

x P = {P1, P2, …, Pn}, the set of places, drawn as circles x T = { T1, T2, …, Tm }, the set of transitions along with their distribution functions or probability values, drawn as bars x A = AI ‰ AO ‰ AH‰ AT, the set of arcs, where O A is the set of output arcs, AI is the set of input arcs, AH is the set of inhibitor arcs, and AT is the set of tracking arcs (connect transition to a tracking variable and are ended by a diamond-shape at the tracking variable end); each arc has a multiplicity assigned to it, x G = {g1, g2, …, gr}, the set of guard functions which are associated with different transitions, x TV = { TV1, TV2, …, TVm }, the set of tracking variables that store the last duration of the enabling time of a transition (drawn as hexagons), x m0 – the initial marking of the Petri net.

of the definition of the proxel to introduce the notion of relevant rewards. They incorporate in the state definition all quantities, which in addition to age intensities; can be parameters of probability distribution functions of events in the model. This yields the following definition of a state: State = (Discrete state, Relevant Rewards, State Relevant Age Intensities) Thus, all parameters required for computing transition distribution functions are contained in the state vector (making the model implicitly a non-homogeneous Markov chain). The discrete state typically corresponds to a marking in the SPN model. The relevant rewards are determined by the nature of the events in the model, i.e. what kind of dependencies the failure rates in the model exhibit. In Table 1 we provide the relevant rewards for the six classes of state-varying failures that we have identified.

Each transition is defined as Ti = (F, type), where type  {enabling, age, immediate} is the type of memory policy if it is a timed transition or “immediate” if the corresponding transition is an immediate one. F is a cumulative distribution function if the corresponding transition is a timed one. Immediate transitions have a constant value instead of a distribution function assigned to them, which is used for computing the probability of firing of an immediate transition if more than one are enabled at once. The sets of arcs are defined such that

Table 1: Relevant rewards for the six state-varying failure classes State-varying failure Relevant reward class a) the number of failure Number of failures occurrences up to the observed point in time, b) the age of the Age of machine machine up to the observed point in time, c) the duration of the Duration of last repair last repair, d) the time between the Duration of operation of machine between two last two failures, consecutive failures e) the properties of the Parameters of repair facilities (e.g. repair facilities, quantification of introduced as experience of additional repairman) parameters, f) the types of failures Types of failures that that have occurred. have occurred so far

AO = {ao1, ao2,…, aok}, AI = {ai1, ai2,…, aij}, AH = {ah1, ah2,…, ahi}, and AT = {at1, at2,…, atl}, where A H, A I Ž P u T u ℕ , A O Ž T u P u ℕ , A T Ž T u P u ℝ . The multiplicity of the tracking arcs can be a real number, unlike the others, where it is a non-negative integer number. We denote by M = {m0, m1, m2, … } the set of all reachable markings of the Petri net. Each marking is a vector made up of the number of tokens in each place in the Petri net along with the values of the tracking variables, mi = (#P1, #P2,…, #Pn, val(TV1), val(TV2),…, val(TVm)). The set of all reachable markings is the discrete state space of the Petri net. The changes from one marking to another are consequences of the firing of enabled transitions which move (destroy and create) tokens; thus creating the dynamics in the Petri net. This makes the firing of a transition analogous to an event in a discrete-event system. The markings of a Petri net, viewed as nodes, and the possibilities of movement from one to another, viewed as arcs, form the reachability graph of the Petri net.

4. EXPERIMENTS AND RESULTS In this section we present an example model which we will use to illustrate the simulation of state-varying failure rates. We will describe the proxel-based simulation of this model and run it using various time steps. 4.1. The Model The model that we use to demonstrate our approach is a simple model that describes a machine that incorporates both time- and state- varying failure rates, similar to the scenarios described in Section 2.1.

3.3. Adaptation of the Proxel-based method to Accommodate State-varying Failure Rates The main adjustment of the proxel-based method to accommodate state-varying failure rates is the extension

529

x

Using a Petri net, the model can be described as shown in Figure 2. It represents a machine that exhibits one of two possible states: OK and FAILED. When the machine has failed, one of the two repairmen arrives and repairs it, after what the machine’s state becomes OK. Changing shifts during repair is not allowed. The two repairmen have different lengths of working experience. Thus, when the machine is fixed by the Repairman A, the time to the next failure is on average longer, than when it is repaired by Repairman B. This implies that the proxels will need to record the information of who performed the last repair as well.

ܰstands for the normal distribution parameters, with the standard parameters: mean and variance, correspondingly.

In other words, repairs are normally distributed, where the mean and the variance are functions of the Repairman that completed the last repair, the total number of failures, and the age of the machine. This implies that we observe the dependences (a), (b), and (e), as pointed out in Section 3. This directly implies that, as described in Table 1, we need to add the following relevant rewards: x Number of failures, x Age of machine, and x Parameters of repair facilities.

fail

In order to illustrate the enhancements that the state-varying failures would require, we include the required information in the basic Petri net model, whose extended version is shown in Figure 4.

FAILED

OK

repair Last A

TV1

change shift Technician B

#FAILURES

Last B

TV2

Technician A

Who fail

change shift

Figure 3: Basic Petri net model of the example

FAILED

OK

In addition to the afore-described scenario, the distribution of the time to next failure of the machine is also a function of the number of failures and the age of the machine. For instance, in our example model we use the following formula to describe the distribution of the repair time:

repair change shift Technician B

݂௥௘௣௔௜௥ ൫ܴǡ ܽ݃݁ǡ ݊௙ ൯̱ ቊ

Technician A

ܰ൫ͳͲ ൅ ܽ݃݁ ‫Ͳ כ‬ǤͲͳ ൅ ݊௙ ‫Ͳ כ‬Ǥͳǡ ܽ݃݁ ‫݊ כ‬௙ ൅ ͳǤͲ൯ǡ ܴ ൌ ‫ܣ‬ ܰ൫ͳʹ ൅ ܽ݃݁ ‫Ͳ כ‬ǤͲͳͷ ൅ ݊௙ ‫Ͳ כ‬Ǥͳʹǡ ܽ݃݁ ‫݊ כ‬௙ ൅ ͳǤͷ൯ǡ ܴ ൌ ‫ܤ‬

Figure 4: Extended Petri net model of the example

(2) where: x x x

change shift

In Figure 5, the state-transition diagram of the Petri net model from Figure 4, is shown. Note that the model is an unbounded Petri net, i.e. it is practically impossible to accurately analyze it using numerical approaches. This, however, is not a limitation of the proxel-based method, as it dynamically builds the state space on-the-fly. Besides the repair duration probability distribution function, as shown by the Equation (2), the remaining distribution functions that we used in our experiments are as follows:

age is the age of the machine, i.e. the global simulation time, ݊௙ is the number of failures that have occurred, i.e. #FAILURES in the SPN, ܴ is the repairman that did the serviced the last failure, i.e. can be obtained from the SPN by checking the token is in place Last A or Last B, and

530

x x

݂௙௔௜௟ ൫ܽ݃݁ǡ ݊௙ ൯̱ ‫ܧ‬൫ͳͷͲ ൅ ܽ݃݁ ‫Ͳ כ‬ǤͲͷ ൅ ݊௙ ‫Ͳ כ‬Ǥͳ൯ ݂௖௛௔௡௚௘௦௛௜௙௧ ሺሻ̱‫ܦ‬ሺͳʹሻ

a) Machine fails - ሺሺ‫ܨ‬ǡ ‫ܣ‬ሻǡ ሺͳǡ ‫ܣ‬ሻǡ ሺͲǡ ο‫ݐ‬ሻሻ, b) Repairman shift change -ሺሺ‫ܨ‬ǡ ‫ܤ‬ሻǡ ሺͲǡ ‫ܣ‬ሻǡ ሺο‫ݐ‬ǡ Ͳሻሻ, c) No events - ሺሺܱ‫ܭ‬ǡ ‫ܣ‬ሻǡ ሺͲǡ ‫ܣ‬ሻǡ ሺο‫ݐ‬ǡ ο‫ݐ‬ሻሻ.

where E() stands for the exponential distribution function, with the mean as its only parameter, and D() is the deterministic probability distribution function. As described in the following subsection, for this example we slightly modify the discrete state description to better explain our approach. In general, the proxel-based simulation can be directly performed on the enhanced Petri net model.

where F stands for the machine’s state FAILED. The age of the machine is implicitly recorded by the global simulation time variable. Note that we assume that the repairman on shift that has started to work on the repair also has to complete it, and thus, extend his shift. Further, for illustration purposes, we will develop the proxel for the case (a), i.e. when the machine has failed, which yields the following subsequent events and proxels:

4.2. Insight in the Proxel-based Simulation In the following we provide insight in the proxel-based simulation for the example model. The goal is show what exactly happens at lower level when simulating the state-varying failures. We begin by defining the state of the systems in the concrete example as:

a-1) Machine is repaired - ሺሺܱ‫ܭ‬ǡ ‫ܣ‬ሻǡ ሺͳǡ ‫ܣ‬ሻǡ ሺͲǡ ʹο‫ݐ‬ሻሻ, a-2) No events - ሺሺ‫ܨ‬ǡ ‫ܣ‬ሻǡ ሺͳǡ ‫ܣ‬ሻǡ ሺο‫ݐ‬ǡ ο‫ݐ‬ሻሻ. The model description yields that the “change shift” transition is of race age policy, i.e. it needs to record the time spent on shift and not be restarted it when a failure occurs. During this processing, the required statistics that yield the simulation results are collected.

ሺሺ‫݁ݐܽݐ݄ܵ݁݊݅ܿܽܯ‬ǡ ܴ݁‫݊ܽ݉ݎ݅ܽ݌‬ሻǡ ሺ͓‫ݏ݁ݎݑ݈݅ܽܨ‬ǡ ‫݊ܽ݉ݎ݅ܽ݌ܴ݁ݐݏܽܮ‬ሻǡ ‫ݎ݋ݐܸܿ݁ݕݐ݅ݏ݊݁ݐ݊ܫ݁݃ܣ‬ሻǡ

4.3. Experiments and Results In the following we present the results of our simulation experiments, i.e. the statistics that were collected during the proxel-based simulation of the example model. The qustions that our simulation model provides answers to are the following:

which implies that the discrete state of the system is described by the state of the machine (‫)݁ݐܽݐ݄ܵ݁݊݅ܿܽܯ‬ and the repairman on shift (ܴ݁‫)݊ܽ݉ݎ݅ܽ݌‬. The relevant rewards are the number of failures of the machine (͓‫ )ݏ݁ݎݑ݈݅ܽܨ‬and the repairman that completed the last repair of the machine (‫)݊ܽ݉ݎ݅ܽ݌ܴ݁ݐݏܽܮ‬. Finally, the last element of the state of the system is the ‫ ݎ݋ݐܸܿ݁ݕݐ݅ݏ݊݁ݐ݊ܫ݁݃ܣ‬that keeps track of the time that the machine has spent in the specified state, as well as the time during which the repairman has been on shift. This yields the initial proxel as:

a)

What is the probability of having the machine running?, and b) What is the probability that the machine has 5 or more failures? The question (a) is a classical reliability analysis case, and the most typical question for a model like this one. In Figure 6 and Figure 7 we present the answers to the questions (a) and (b), correspondingly. The simulation parameters that we used were: a maximum simulation time t = 300, and a time step 't = 0.5. Apparently, in Figure 6, we can observe that the model has not reached a steady-state during the simulation time of 300 time units, thus it needs to be simulated for a longer period of time. We did this, i.e. we ran the simulation up to time t =10000 using the same time step, and the obtained solution for the steady state reliability of the system is periodically oscillating, as shown in Figure 8.

ሺሺܱ‫ܭ‬ǡ ‫ܣ‬ሻǡ ሺͲǡ ‫ܣ‬ሻǡ ሺͲǡ Ͳሻሻ. which shows that initially the machine is in state OK, and Repairman A is on shift. The number of failures up to simulation time t = 0 is zero, and the age intensities of both machine state OK and duration of Repairman A on shift are zero as well. Note that initially we assume that the last repair was completed by the more experienced repairman, i.e. Repairman A. The subsequent proxels which originate from the initial one at time ‫ ݐ‬ൌ ο‫ݐ‬, along with the three potential events, are the following:

531

(F, A), (1, A)

...

(OK, B), (1, A)

...

(OK, A), (1, A)

(OK, A), (0, A) (OK, B), (0, A)

(F, A), (2, A)

(F, B), (1, B)

(F, B), (2, B)

...

(OK, A), (1, B)

...

(OK, B), (1, B) (F, A), (1, A)

...

(OK, B), (0, A)

...

(OK, A), (0, A)

Probability

Figure 5: State-transition diagram of the unbounded Petri net model from Figure 4

1.00 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 0

50

100

150

200

250

300

Simulation time OK

FAILED

Figure 6: Transient solution for the 2 discrete states, neglecting the fact of what repairman is on shift

0.80 Probability

Probability

1.00 0.80 0.60 0.40

0.60 0.40 0.20 0.00

0.20

Simulation Time

0.00 0

100

200

300

OK

FAILED

Simulation time Figure 8: Steady-state solution of the discrete stochastic model

Figure 7: The probability of the machine having 5 or more failures

532

Probability

1.00 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 0

50

100

150

200

250

300

Simulation time OK

FAILED

Figure 9: Transient solution for the 2 discrete states for the simplified model, neglecting the fact of what repairman is on shift

x x

Probability

The simulation results were obtained in 0.5 seconds on an Intel Core i5 2.53GHz workstation with a 4GB of RAM. The extended computation for the steady-state solution took longer, i.e. 3.5 minutes, which is still an acceptable running time.

1.00 0.80 0.60 0.40 0.20 0.00

5. SUMMARY AND OUTLOOK We presented an approach to more realistically model and simulate failures that exhibit a wide range of dependencies which are typically neglected. Their neglecting, however, can provide highly misleading results, and thus, it is imperative to avoid their oversimplification. The proxel-based method has shown to be very accurate and highly flexible in describing the complex types of dependencies that typically occur in stochastic models. We anticipate extending of the presented work to provide a tool that would facilitate reliability modeling considering state-varying failure rate functions.

Simulation time OK

݂௥௘௣௔௜௥ ሺ ሻ̱ܰሺͳͳǤͲǡ ͳǤͲሻ, ݂௙௔௜௟ ሺ ሻ̱‫ܧ‬ሺͳͷͲሻ

FAILED REFERENCES Baker, R. D. (2001). "Data-based modeling of the failure rate of repairable equipment." Lifetime Data Analysis 7(1): 65-83. Bause, F. and P. S. Kritzinger (2002). Stochastic Petri Nets, Vieweg. Billinton, R. and R. N. Allan (1992). Reliability evaluation of engineering systems, Plenum Press New York. Coskun, A. K., R. Strong, et al. (2009). Evaluating the impact of job scheduling and power management on processor lifetime for chip multiprocessors, ACM. Cox, D. R. (1955). "The analysis of non-Markovian stochastic processes by the inclusion of supplementary variables." Proceedings of the Cambridge Philosophical Society 51(3): 433-4 Hassett, T. F., D. L. Dietrich, et al. (1995). "Timevarying failure rates in the availability and

Figure 10: Steady-state solution of the simplified discrete stochastic model For comparison, in Figure 9 and Figure 10, we provide the transient and steady state solutions of the same model, where the state-varying distributions are substituted with state-independent ones, i.e. with distributions with fixed parameters. Apparently, the results and their nature are quite different as the oscillating steady-state pattern is not present in the simplified model. More specifically, the distribution functions that we used for the state-independent failure rates were the following:

533

reliability analysis of repairable systems." Reliability, IEEE Transactions on 44(1): 155-160. Horton, G. (2002). "A new paradigm for the numerical simulation of stochastic Petri nets with general firing times." Proceedings of the European Simulation Symposium. Isensee, C. and G. Horton (2005). "Approximation of Discrete Phase-Type Distributions." Proceedings of the 38th annual Symposium on Simulation: 99106. Lazarova-Molnar, S. (2005). The Proxel-Based Method: Formalisation, Analysis and Applications. Faculty of Informatics. Magdeburg, University of Magdeburg. Ph.D. Lazarova-Molnar, S. (2008). State-Dependent Transitions in Discrete Stochastic Models: Deterministic Simulation Approach. Summer Computer Simulation Conference 2008. Liberopoulos, G. and P. Tsarouhas (2005). "Reliability analysis of an automated pizza production line." Journal of Food Engineering 69(1): 79-96. Lin, F. J., P. M. Chu, et al. (1987). "Protocol verification using reachability analysis: the state space explosion problem and relief strategies." ACM SIGCOMM Computer Communication Review 17(5): 126-135. Proschan, F. (1963). "Theoretical explanation of observed decreasing failure rate." Technometrics 5(3): 375-383. Retterath, B., S. S. Venkata, et al. (2005). "Impact of time-varying failure rates on distribution reliability." International Journal of Electrical Power and Energy Systems 27(9-10): 682-688. Schroeder, B. and G. A. Gibson (2007). Disk failures in the real world: What does an MTTF of 1,000,000 hours mean to you?, USENIX Association. Sharma, S., K. S. Kahlon, et al. (2010). "Reliability and path length analysis of irregular fault tolerant multistage interconnection network." ACM SIGARCH Computer Architecture News 37(5): 16-23. Stewart, W. J. (1994). Introduction to the Numerical Solution of Markov Chains., Princeton University Press. Xie, K. and W. Li (2009). "Analytical model for unavailability due to aging failures in power systems." International Journal of Electrical Power and Energy Systems 31(7-8): 345-350. Zhang, H., E. Cutright, et al. (2010). Time-varying failure rate for system reliability analysis in largescale railway risk assessment simulation. Safety and Security in Railway Engineering: 29.

in Macedonia and her M.Sc. in Computational Visualistics and Ph.D. in Computer Science from “Ottovon-Guericke” University of Magdeburg in Germany. Her main research interests are simulation and modeling, with the main focus on their methodology for achievement of more accurate and valid results. Her email is [email protected].

AUTHORS BIOGRAPHY SANJA LAZAROVA-MOLNAR is an Assistant Professor in Computer Science at United Arab Emirates University. She received her Diploma in Computer Science from the “Sts. Cyril and Methodius” University

534

MODELLING AND SIMULATION OF ORDER SORTATION SYSTEMS Fahrettin Eldemir (a), Elif Karakaya(b) (a)

Industrial Engineering Department, Fatih University, Istanbul Industrial Engineering Department, Fatih University, Istanbul

(b)

(a)

[email protected], (b) [email protected]

items from different orders are arranged over again (batch orders) and the same product types collected by the same workers. With this method, order pickers are assigned to a specific zone. In this way, unproductive travel time will disappear. However, although this situation saves time and speed, the items of accumulated orders completely mixed. Therefore the items collected by different pickers arrive to the packing area at different times. To wait the other items from the same order, the ready packages are accumulated in accumulation zone. There is no doubt that these products (items) have to be sorted according to the product type and quantity before shipment. At this point, sortation systems (these are often automated systems) are used. The optimal condition for a given system studied would be one in which the rate of sortation (i.e., throughput rate) is maximized, so minimizing the wave sortation time without increasing the capital and operating costs. There is a trade-off between the rate and cost. Using more resources such as labor and machines can increase the rate of sortation; however, the cost of sortation thus increases. This study focuses on maximizing the throughput rate of a given system and assumes that the other variables, such as cost and operating design parameters, are held within satisfactory limits. There are different sortation strategies available. Fixed Priority Rule, Next Available Rule, and Earliest Completion Rule. In the literature a few analytical models have been developed for these sortation strategies. However the sortation models are limited to the one induction lane and one sortation lane.

ABSTRACT The Order Accumulation and Sortation Systems (OASS) are getting more important as distribution centers try to gain competitive advantages. The parameters that affect the sorting time in OASS are analyzed in this study. The length and speed of conveyors, sending the packages within a wave, the wave size, number of the sorting lanes and the sorting strategy are the main parameter in OASS. The time required to sort mixed items depends on the sortation strategy used as well. Different sorting strategies and different conveyor models are analyzed in this study. Available analytical models assume that all orders are at the same size (quantity). In this study this assumption also is relaxed. Simulation models have been developed to compare different design alternatives and design strategies. For different order combinations and for various design choices, simulation is used to compare sortation strategies. The results have been given in tables that show which strategy should be used under which scenario. AutoMod Software is used as the simulation tool. Keywords: sortation strategies, inventory management, simulation 1. INTRODUCTION In today's competitive world, it is desirable that a distribution center runs at its optimal settings to gain a competitive advantage. More efficient distribution centers are needed to respond to the increasing competition and to an increased emphasis placed on time-based service. In distribution centers, long list of orders are put together in an intensive way. Each customer order can be full of various items at different quantities. In classical order picking procedure, each order is collected by an assigned picker and the products in this list might being kept at different storage addresses. Therefore, picker may end up traveling to far distances in a warehouse in order to complete the list and searching the items all over the warehouse. This situation often causes unnecessary transportation costs and ineffective worker utilizations. To overcome shortages mentioned above, zone picking method widely used in warehouses. In this picking method, the

2. LITERATURE REVIEW Order Accumulation and Sortation System (OASS) related publications are very few. The first example related sortation strategies comes from (Bozer et al., 1988) developed Fixed Priority Rule (FPR) for lane assignment by simulating different wave of orders. Johnson (1998) developed a dynamic sortation strategy which is called Next Available Rule (NAR) and compared it with “FPR". Eldemir (2006) developed an alternative sortation strategy called Earliest Completion Rule (ECR) by using order statistics.

535

Closed-loop simple conveyor design researches contain different number of induction lane and number of sortation lanes. Especially the first studies are related one induction and one sortation lane. However later on, because of the variability in products and order sizes, the conveyor designs seen in the literature adapt into many induction and sortation lanes. Following table summarizes the literature on closed-loop conveyor system analysis according to number of its induction and sortation lane. Table 1: Literature Review about Conveyor Design Sortation Literature Summary Problem Setting Citation

Method

One Many One Many Ind. Ind. Sort. Sort.

Bozer and Sharp (1985)

Simulation



Bozer et al (1988)

Simulation





Johnson and Lofgren(1994)

Simulation





Johnson (1998)

Analytical





Meller (1997)

Analytical





Schmidt and Jackman(2000)

Analytical



Johnson and Meller(2002)

Analytical





Russell and Meller(2003)

Descriptive Mdl





Bozer (2004)

Analytical

Eldemir (2006)

Analytical

3.

4. SORTATION STRATEGIES Sortation strategies can be classified into two families, fixed priority rules (FPR) and dynamic assignment rules. In fixed priority rules, the orders are prioritized before sortation based on a certain rule. Dynamic assignment rules are assignment strategies that consider the item locations on the conveyor. The most common examples of this family are the next available rule (NAR) and the earliest completion rule (ECR). All parameters are determined below:



√ √

Figure 2: One- Many Model Conveyor Design



√ √

SORTATION SYSTEM DESIGN

3.1 One-One Model In this design model, one induction lane and one sortation lane is available. When the literature is evaluated thoroughly, it is observed that this model is the first applied model to the re-circulating conveyor. For instance, Bozer and Sharp (1985) have carried out this model in order to develop sortation strategies.

Table 2 : Notation y Number of items within an order m Number of orders within a wave l Length of the closed-loop conveyor v Speed of the conveyors T The time for an item to circulate around the main sortation line n Number of accumulation lanes i Item index within an order j Order index within a wave q The number of orders sorted thus far 4.1 Fixed Priority Rule (FPR) Sortation time evaluation by using Fixed Priority Rule is given follows. The number of accumulation lane is accepted as one, and the number of items within the order is assumed to be constant. Under FPR, The sorting time for all orders within the specific wave will be the summation of all the gaps and spreads as follows:

Figure 1: One- One Model Conveyor Design

TFPR =

3.2 One-Many Model One –Many Model differs from previous model since it has more than one sortation lanes. When it is compared with others, this model is the most applied one. For instance, Johnson and Lofgren (1994), Johnson (1998), Meller (1997) have used this model in their studies.



536

m.T . y  y +1









 

Table 3: Sorting Times for Numerical Examples Order Sorting time (seconds) Sorting Sequence (Order number) FPR NAR ECR 1 83,33 80,39 57,26 2 83,33 80,43 58,40 3 83,33 80,49 59,70 4 83,33 80,56 61,19 5 83,33 80,65 62,94 6 83,33 80,77 65,04 7 83,33 80,95 67,64 8 83,33 81,25 71,02 9 83,33 81,82 75,76 10 83,33 83,33 83,33

4.2 Next Available Rule (NAR) In this Next Available Rule, the expected sorting time each order depends on the number of orders which stays behind to be sorted. If it is supposed that the location of the items in the remaining orders are independent and uniformly distributed, and q the number of orders sorted thus further. Under NAR, the sorting time for all orders within the specific wave will be as follows: m −1 § m−q · ¸ TNAR = T .¦ ¨¨1 − y ( m − q) + 1 ¸¹ q =0 ©

  



Total

833,30

810,64

662,29

5.1.2 Simulation Model In order to compare ECR, FPR and NAR, a simulation method is used as well. Several assumptions are made to facilitate the simulation analysis. To illustrate the expressions for the three sorting strategies, the time to traverse the re-circulating conveyor is T = 222, 8 seconds and there are m =10 orders in each wave with y= 5 boxes per order. A hundred repetitions are done for each simulation experiment. Then, the average of these repetitions is taken.

4.3 Earliest Completion Rule (ECR) In dynamic assignment category, another sortation strategy model is Earliest Completion Rule (ECR). When sortation of an order is finished, the next order is determined based on the location of the last items. The order with the last item being closest to the accumulation lane is selected as next order to be sorted. Like NAR, the sortation time will be dependent on the number of orders which are going around on the main sortation lane. Assuming that all items are randomly and uniformly distributed and on the closed-loop conveyor and the item locations are independent of each other, from order statistics. In Earliest Completion Rule, the total wave sortation time is given:

[

] (3)

m−1 § y (m − q ) T · TECR = ¦ ¨ y ( m−q ) .³ l y .(T y − l y ) m−q −1 .dl ¸ l =0 T ¹ q =0 ©

where (l) is the location of last item on conveyor with the length of (T). 5. 5.1

EXPERIMENTATION One-One Model

Figure 3: Screenshot

5.1.1 Analytical Model To compare ECR, FPR and NAR, an empirical method is used. In developing the analytical models, several assumptions are made to facilitate the analysis. To illustrate the expressions for the three sorting strategies, the time to traverse the re-circulating conveyor is T = 100 seconds and there are m =10 orders in each wave with y= 5 boxes per order. For analytical model experimentations MAPLE software is used.

One-One

Design

Model

Simulation

For simulation model experimentations, AUTOMOD software is used. Figure 3 is the screenshot of the Automod software for One-One Design Model.

537

Table 6: Sort of Spending Time for Simulation Spending Time Duration

Table 4: Sorting Time Comparison for One-One Model by Using Simulation Model Model FPR NAR ECR One-One Model 2.248,78 2.179,25 1.841,70

Transfer Time 1 Transfer Time 2 Transfer Time 3 Load Creation time Packaging Time Total Time

5.1.3 Simulation Model versus Analytic Model Simulation model and Analytical model outputs, according to different scenarios are illustrated in following Table 5. 5.2 Table 5 : Sorting Time Comparison for One-One Model Both Simulation and Analytical Model Wave Sorting Wave Sorting Time (seconds) Time (seconds) Orders/ Items/ Wave Orders Analytical Model Simulation Model FPR NAR 24 12 8 6 4 3 2 1

1 2 3 4 6 8 12 24

2676 1784 1338 1070 765 595 412 214

214 1478 1246 1033 754 591 411 214

49,5 29,76 35,2 69,7 50,7 234,86

One-Many Model

5.2.1 Simulation Model

ECR FPR NAR ECR 214 992 974 878 695 563 403 214

2915 2033 1574 1298 1003 823 640 442

442 1724 1507 1279 991 829 643 442

442 1484 1268 1120 920 792 634 442

It can be realized above Table 5 that Simulation Model’s results are greater than Analytical Model in every case. The reason of this situation is that in simulation model, there are some additional spent times. The following shape points out spending time locations on the simulation system.

Figure 5: One- Many Design Model Simulation Screenshot As it can be seen clearly, the best one is ECR model as One-Many Model. Since, the lowest value which emphasizes the average of the total sorting time is for ECR model. Table 7: Sorting Time Comparison for One-Many Many by Using Simulation Model Model

FPR

NAR

ECR

One-Many Model

690,61

678,72

651,21

5.3 Random and Equal Number of Items in the Order Before studies assumed that number of item in an order are same. For Example, In Johnson (1998)’s article an accepted item number is y=5 for any event. In practice, it is known that it cannot be provided for every wave. Item number varies from one order to another order. Table 8 : Sorting Time Comparison of Sorting Strategies According to Number of y

Figure 4: Extra Times Spending for Simulation In Table 5 there are averages of extra time for each spending point which are shown in preceding shape. Besides, if subtraction is taken from simulation model to analytical model, the average difference is approximately 239 seconds. Also, summation of the extra spending time is 234.86 second. Thus, we can say that these two numbers are too close to each other.

538



Number of y

FPR

NAR

ECR

KŶĞͲKŶĞ DŽĚĞů

Random Equal

2142,65 2248,78

2.116,97 2.179,25

1.818,30 1.841,70

KŶĞͲDĂŶLJ DŽĚĞů

Random Equal

668,95 690,61

650,04 678,72

636,46 651,21

From above shapes, random item size provides more time saving than equal item size in addition to, it does not reflect reality. 5.4 Number of Orders versus Number of Items Different numbers of items and orders combinations are designed in order to comprehend the sortation strategies behavior for various situations. After preparing 8 combinations, for example, 24-1 means that there are 24 different orders within a wave and all orders have only one item, Table 5 represents the strategies’ results: Table 9: Total Sortation Time for Different Sortation Lane in O-OM Orders/ Items/ Wave Sorting Time (seconds) Wave Orders FPR NAR ECR 24 12 8 6 4 3 2 1

1 2 3 4 6 8 12 24

2.915,41 2.032,74 1.574,19 1.297,91 1.003,05 822,74 639,51 441,8

441,8 1.723,56 1.506,52 1.279,31 990,58 828,92 642,73 441,8

Figure 7: Effect of the Distance between Sortation Lanes

441,8 1.484,25 1.268,18 1.119,80 919,74 792,45 634,12 441,8

Simulation models are developed for all designs. Therefore, the results of simulation models are compared with analytical models and in this way, the validation of simulation is provided. Different scenarios are simulated by varying design and operational parameters. For instance, despite of the literature, random item size in an order supports better results. Besides, it is more appropriate for a real case.

As can be seen from Figure 5, great savings can be accomplished in total sortation time for every experiment by using ECR

REFERENCES Bozer, Y. A., and Hsieh, Y. J., 2004, Expected waiting times at loading stations in discrete-space closedloop conveyors, European Journal of Operational Research, vol. 155(2), pp. 516-532 Bozer, Y.A., and Quiroz, M. and Sharp, G.P., 1988, An evaluation of alternate control strategies and design issues for automated order accumulation and sortation system, Material Flow, vol. 4, pp. 265282 Demongodin, I. And Prunet, F., 1993, Simulation modeling for accumulation conveyors in transient behavior, COMPEURO 93, Paris, France, pp. 29-37 Dotoli M., Iacobellis, G, Stecco, G., Ukovich W., 2009, Performance Analysis and Management of an Automated Distribution Center” IEEE Gagliardi, J, Angel R, Renauld, J. 2010, A simulation modeling framework of multiple-Aisles Automated Storage and Retrieval System” CIRRELT -57 Harit, S., Taylor G. D., 1995, Framework for the Design and Analysis of Large ScaleMaterial Handling Systems” Winter Simulation Conference Jayaraman, A., R. Narayanaswamy, et al. 1997, A sortation system model, Winter Simulation Conference Jing, G. and Kelton, W. D. and Arantes, J.C., 1998, Modeling a controlled conveyor network with merging configuration, Proceeding of the Winter Simulation Conference, pp. 1041-1048 Johnson, M. E., 1998, The impact of sorting strategies on automated sortation system performance”, IIE Transactions, vol. 30(1), pp. 67-77

Figure 6: Total Sortation Time for Different Sortation Lane in O-OM 6. CONCLUSION Available sortation strategies are compared and a set of modeling approach in simulation and in analytical is developed for the design and analysis of conveyor sortation system. Consequently, the following contributions are made: Based on simulation models, FPR, NAR and ECR sortation strategies are compared. Overall outputs are represented as fallows in Figure 7.

539

Johnson, M. E. and Russell, M., 2002, Performance analysis of split-case sorting systems, Manufacturing & Service Operations Management, vol. 4(4), pp 258-274 Johnson, M. E. and Lofgren, T., 1994 Model Decomposition speeds distribution center design”, Interfaces, vol. 24(5), pp. 95-106 Kale, N., Zottolo, M., Ülgen, O.M., &Williams, E.J. 2007, Simulation improves end-of line sortation and material handling pickup scheduling at appliance manufacturer. Proceedings of the 2007 winter simulation conference pp.1863–1868, Washington, D.C., USA Koster, R., Le-Duc., T., Roordbergen, K., J., 2007 Design and control of warehouse order picking: A literature review, European Journal of Operational Research 182, 481–501 Le-Duc, T., and de Koster, R. 2005,Determining Number of Zones in a Pick-and-pack Order picking System, ERIM Report Series Research in Management, Rotterdam Maxwell, W. and Wilson, R., 1981, Deterministic models of accumulation conveyor dynamics, International Journal of Production Research, vol. 19(6), pp. 645-655 Meller, R. D. 1997, Optimal order-to-lane assignments in an order accumulation/sortation system, IIE Transactions 29: 293-301 Roodbergen, K.J., and Vis, I.F.A., 2009, A survey of literature on automated storage and retrieval system. European Journal of Operational Research, Vol.194, pp.343, 362 Russell, M. L. and Meller, R. D., 2003, Cost and throughput modeling of manual and automated order fulfillment systems, IIE Transactions, 35, 589-603 Schmidt, L. C. and Jackman, J., 2000, Modeling recirculating conveyors with blocking, European Journal of Operational Research, vol. 124, pp. 422-436 Sonderman, D., 1982, An analytic model for recirculating conveyors with stochastic inputs and outputs, Internal Journal of Production Research, vol. 20(5), pp.591-605 AUTHORS BIOGRAPHY Fahrettin Eldemir is an Assistant Professor in the Department of Industrial Engineering at Fatih University, Istanbul, Turkey. He has a Ph.D. in Decision Sciences and Engineering Systems, an M.E. in Operations Research and Statistics B.S in Industrial and Management Engineering from Rensselaer Polytechnic Institute. Before joining to Fatih University, he served as an engineer and a consultant at Omega Advanced Solutions.

540

DEVELOPMENT OF THE SURFACE TO AIR MISSILE SIMULATOR THROUGH THE PROCESS OF COMPONENT COMPOSITION AND DYNAMIC RECONFIGURATION OF WEAPON SYSTEM J.B. Suk(a), J.O. Lee(b), Y.H. Seo(c) (a)

Department of Industrial Management Engineering, Korea University, Seoul, Korea Department of Industrial Management Engineering, Korea University, Seoul, Korea (c) Department of Industrial Management Engineering, Korea University, Seoul, Korea

(b)

(a)

[email protected], (b)[email protected], (c)[email protected]

hardness, we require developing a basic component model that can give flexibility, scalability, reusability in modeling and simulation. Also, we require assembling and composing component model and need a way of component based development for new simulation model. This way is to assemble developed components and create new applications. It has advantages such cost and time of development and maintenance of software. Ultimate goals of this development are maximizing reusability of components and increasing productivity of the software. New software model of the weapon system is developed by applying the technology of software product line in the field of modeling and simulation of weapon system that has a lot of similarities for each product. Therefore, in order to effectively reuse, prebuilt, core assets of the software, studies are needed for this processes; management, selection, assembly and composition of component models. The component reconfiguration of weapon system and related works are handled in second chapter. In third chapter, the development of surface to air missile simulator and its application example are organized. In the last chapter, it ends the study with conclusion and future work.

ABSTRACT The concerns of the technology of the reusable component are increasing for producing the software effectively and developing the products quickly on demand of customers. Technologies of composition and dynamic reconfiguration of the component are needed to develop the simulator of the weapon system in field of national defense. It is needed to develop the surface to air missile simulator as the example of the weapon system through component-based development. In order to develop the simulator, creation of components of the simulator is required and the process of reconfiguration is defined and realized. The process of reconfiguration consists of four parts: management, semantic test, composition and performance eˆluation. In first part, the developer can easily create lots of component models including characteristics of the product of the weapon system and quickly manage them by using the management tool. In second part, semantic tester tests that whether the reconfiguration and import in the simulator are possible or not. In next part, the composer constructs the product models configuring surface to air simulator by composing existing component and newly created component. Then the results of construction can be entities of detecting radar, the surface to air missile, the launcher of the missile and aircrafts of the simulator of the weapon system. In last part, the existing 3Dbased simulator for evaluating the performance of the component of the weapon system confirms the effects of reconfigured components by importing it. Therefore, this study provides basic framework to simulate common weapon systems through the technology of composition and the process of dynamic reconfiguration.

2.

COMPONENT RECONFIGURATION OF WEAPON SYSTEM Through software product line based development, it designs the dynamic component reconfiguration framework which can reuse and compose quickly the component of weapon system. Software product line based development is the way of keeping common elements of the product and only changing the distinguishing characteristics of it for reusing components. In order to realize the framework, components will be defined and developed as the unit of reusable model. The component model consists of physical part and behavior part. It is reconfigured through considering characteristics of two parts of it and being selected and composed. The component model can be created, modified and deleted by graphic-based management tool of the

Keywords: surface to air missile, modeling and simulation, reusability, component 1.

INTRODUCTION Developing a simulation model will require a lot of time and money in the field of modeling and simulation. Developed model is hard to be used as other applications by its closed architecture. To overcome this

541

component. Also, it can be aproved to reuse for constructing the reconfiguration framework of weapon system by semantic tester of the component after different developers created it. In order to develop the simulator of suface to air missile of weapon system, it is reconfigured as composing the reuable components. For composing existing component and newly created component by the tool, the composers of the component model are developed. They also consist of physical composer and behavior composer. Physical composer constructs physical model of the product by composing components grouped by physical characteristic of component model of weapon system like the performance of the component. Furthermore, behavior composer constructs behavior model of product by including logics and rules in product. Then we can reconfigure product model of the simulator of weapon system through use of two composers.

components as specification of the variability from the basic framework.

Figure 1: The Structure of Software product Line Engineering 2.1.2. Development of Basic Component Model By applying software product line engineering, it is needed to develop basic component model of weapon system that is possible to reconfigure as unit of the component. This study develops basic component model that is possible to create various products efficiently by considering scalability of weapon system. Basic component model consists of physical part and behavior part. Physical component model contains representing physical characteristics of the weapon model. In order to develop physical component, it is needed to define criteria for description of weapon system. It is decided by requirements of the domain. This study decided criteria limits at entity level. Next, this study considers fidelity of physical component model that represents correspondence with real entity model like the missile. High fidelity of entity model causes low reality. So it is needed to decide at optimum level. As example, the missile is composed 4 parts by analyzing its functionality and structure in figure 2. Behavior component model also represents tactical behavior and decision. And it is developed by dividing 3 parts; basic behavior, composition behavior and judgment.

2.1. The Process of Component Composition 2.1.1. Software Product Line Engineering The recent interest in the field of modeling and simulation is development of reusable and configurable simulation model. It is generalized small quantity batch production from rise of importance of personalization of the customers and issued software product line engineering for corresponding the demand of customers and environment of the market. Software product line engineering is paradigm that reusing the core asset from similar product and inevitable choice for development of new product that satisfying time and economic restriction of the development (Chen, Yu, Gannold, Gerald C., Collofello and James 20006). The definition of software product line engineering is paradigm for developing the applications of software sing platform and mass customization (Pohl, Klaus and Bockle 2005). The objective of software product line is increasing efficiency of the development by reusing the product strategically as analysis of similarities and differences between the products of software set. So it is to choice the option of product according to the user’s intention and to produce new software on one of basic platforms along the choice. The basic concepts of it are understanding the similarity among characteristics of the product and supporting the variability among the application programs by distinguishing that of the product complicity (Clements, Paul and Northrop 2002). The structure of software product line engineering is as below Figure 1. It is classified domain engineering and application engineering. Domain engineering is process of development by setting up core assets from common function through analysis of similarity and the variability. And it specifies requirements and characteristic of components and grafts them onto the basic framework. On the other hand, application engineering classifies the variability from characteristics of components and develops the target application. These two engineering help developing target application by reconfiguring pre-modeling

x x x

Basic behavior: The behavior that performs alone fundamentally designed Composition behavior: Combined behavior except basic and original function of entity Judgment: It represents the state of the behavior by deciding changing behavior or not.

Figure 2: Structure of Missile Considering Fidelity

542

various and differenced product needed to customers as other composition of products.

2.2. The Process of Dynamic Reconfiguration For dynamic reconfiguration of component model, developed basic component models are configured onto product model that means the entity of weapon system. Next, the product models are also reconfigured to system model as application program by following the designed DCRF (Dynamic Component Reconfiguration Framework).

3.

DEVELOMENT OF THE SURFACE TO AIR MISSILE SIMULATOR By applying designed the framework, the surface to air missile simulator is developed as one of software application. By using this simulator, reusability of developed component model is confirmed and performance evaluation of it is also possible. The user develops surface to air missile simulator by importing composed physical part and behavior part of the product model in existing 3D-based performance evaluation simulator of the weapon system. In addition, the user reconfigures various products and evaluates performace of it through the use of management tool and semantic tester.

Figure 3: The Process of Dynamic Reconfiguration 3.1. Framework of Surface to Air Missile Simulator As application from reconfiguration framework, SAM simulator is one of systems in system layer. Basically, it needs 4 product models such as missile, radar, launcher, and aircraft. With these as the center, there exist various components in these. For completing the product, core assets are existed and distinguishing components are composed with relative function independently.

2.2.1. Description of Dynamic Component Reconfiguration Framework The framework of dynamic component reconfiguration is proposed based on requirement of customers and reusable basic components. It consists of system layer, product layer, component layer and supporting tools. At component layer, the components are reconfigured to form product models and they also are reconfigured to develop the target application system.

Figure 5: Framework of SAM Simulator Figure 4: Dynamic Framework

Component

3.2. Application example In this study, it develops SAM simulator that comprised of 4 product models based on the framework. Objective of this simulator is to bring down aircraft of enemy that has the mission of destroying our core facilities by shooting our missiles. The roles of each product model are as below in SAM simulator.

Reconfiguration

2.3. Related works Typical national defense simulation system by reusing components and configuring is OneSAF (One Semi-Automated Forces) model. OneSAF is comprehensive model that used in ACR (Advanced Concepts and Requirements), RDA (Research, Development and Acquisition) and TEMO (training, exercise and military operation) (Giampapa JA, Sycara K, Owens S, Glinton R, Seo YW, Yu B 2004). OneSAF model has the concept of assembled software product line. This concept is based on completed system that consists of many products, and they also consist of many components. OneSAF is possible to comprise

1. 2. 3. 4.

543

Radar: Detect aircraft and delivery information and order. Launcher: Receive information and order of aircraft and fire the missile. Missile: Hit the aircraft of enemy. Aircraft: Move and destroy the core facilities.

After the execution of simulation, the simulator provides accuracy rate and information of product models in real time. Also the process of reconfiguration as application is realized the developments of management tool, semantic, physical composer and behavior composer. x

x x

components that developed by the management tool and semantic tester of the component. Product models are entities of missiles, detecting radar, launchers and aircrafts for constructing surface to air missile simulator through the use of the composers. The entities include not only the physical model, but also include the behavior model to express developer’s logical and specific intentions. It is possible to evaluate the characteristics of the component of the system quickly and easily through realization of the surface to air simulator. In addition, more studies are needed to store and manage the components

Management tool: It manages so many various the basic components that have distinguishing performance models such that add, delete and modify them. Semantic Tester: It tests whether newly added or modified component model is possible to be used at simulator of weapon system. Physical Composer and Behavior Composer: By composing the basic component models, both composer creates new product models that having characteristics that are choiced by user

ACKNOWLEDGMENTS This work was supported by Defense Acquisition Program Administration and Agency for Defense Development under the contract UD110006MD, Korea. REFERENCES Chou SC, Chen YC. 2006. Retrieving reusable components with variation points from software product lines, Information Processing Letters, 99 (3), 106-110. Giampapa JA, Sycara K, Owens S, Glinton R, Seo YW, Yu B, 2004. Extending the OneSAF Testbed into a C4ISR Testbed. Simulation, 80(12), 681-691. Henderson C, Rodriguez A, 2002. Modeling in OneSAF. Computer generated forces and behavioral representation, 337-348. Chen Y, Gerald C, Gannod, Collofello JS, 2006. A software product line process simulator. Software process improvement and practice, 11(4),385-409. Clements, Paul and Northrop, Linda, 2002. Software Product Lines: Practices and Patterns, 5-50. Pohl, Klaus and Bockle, Gunter, 2005. Software Product line engineering : Foundation, Principles, and Techniques, Springer, 159-370.

Figure 6: Application of Reconfiguration

Figure 7: Surface to Air Missile Simulator 4.

CONCLUSION In this study, dynamic component reconfiguration framework is designed and realized through software product line-based development to produce the simulator easy and fast. Developed components of modeling and simulation of the weapon system increase productivity of simulation model and reduce the development cost and time. It is possible to reconfigure products as the user intended by composing reusable

544

MODELLING AND SIMULATING A BENCHMARK ON DYNAMIC RELIABILITY AS A STOCHASTIC ACTIVITY NETWORK Daniele Codetta-Raiteri(a) (a)

Dipartimento di Informatica, Università del Piemonte Orientale, Alessandria, Italy (a)

[email protected]

behaviour which are relevant to Dependability evaluation. The mechanisms that lead to the failure of a technological item are very complex and depend on physical, technical, human and environmental factors which may not obey deterministic laws; so, the modelbased method follows the probabilistic approach. The concept of Dependability is quite general; in order to quantify the Dependability, we need particular measures: the Reliability R(t) of an item (component or system) is the probability that the item performs the required function in the time interval (0, t), given the stress and environmental conditions in which the item operates; the Unreliability U(t)=1-R(t) is the probability that the item is in the failed state at time t (Sahner, Trivedi, and Puliafito 1996). We talk about dynamic reliability (Marseguerra, Zio, Devooght, and Labeau 1998) when the reliability parameters of the system change according to the current configuration of the system. For instance, the failure rate of a component may be expressed as a function of one or more variables describing the current behaviour or state of the system. In dynamic reliability, considering only the combinations of failure events is not sufficient to evaluate the system (un)reliability, but we actually have to take into account the complete behaviour of the system. This means modelling the normal functioning of the system, the occurrence of failure events and their effect on the system functions. For this reason, combinatorial models such as Fault Trees and Reliability Block Diagrams (Sahner, Trivedi, and Puliafito 1996) are not suitable to deal with cases of dynamic reliability because such kinds of model can only represent combinations of component failure events. Their extensions such as Dynamic Fault Trees (Dugan, Bavuso, and Boyd 1992) and Dynamic Reliability Block Diagrams (Distefano and Xing 2006) introduce the possibility to represent dependencies among the events, but they still only focus on the failure propagation ignoring the other aspects of the system behaviour. Such aspects could be represented instead by means of state space based models, such as Markov Chains and Stochastic Petri Nets (Sahner, Trivedi, and Puliafito 1996), but their use typically leads to the state space explosion because the complete dynamics of the system has to be modelled. Therefore the model analysis

ABSTRACT Several versions of a benchmark on dynamic reliability taken from the literature are examined: each version deals with particular aspects such as state dependent failure rates, failures on demand, and the repair of components. The benchmark was modelled in the past, using two types of Petri Nets; in this paper, we exploit another Petri Net based modelling formalism called Stochastic Activity Network (SAN). This allows a more compact model of the system by exploiting input and output gates, together with the possibility to represent float variables by means of extended places. The SAN model of the system undergoes simulation in order to compute the system unreliability: the results are coherent with those obtained with other methods, and this confirms that Petri Net based models can be a valid approach to dynamic reliability evaluation. Keywords: dynamic reliability, benchmark, modelling, simulation, Stochastic Activity Network, Petri Net. 1. INTRODUCTION We talk about safety critical systems when their incorrect behaviour may cause undesirable consequences to the system itself, the operators, the population or the environment. This definition fits categories of systems such as industrial production plants, electric power plants, and transportation systems. Dependability is a fundamental requirement for this class of systems. The Dependability level of a system is the degree of confidence that the system will provide its service correctly during its life cycle. There are two main methods to evaluate the Dependability: the Measurement-based method and the Model-based method. The first one requires the observation of the behaviour of the physical objects composing the system, in the operational environment. This method is more believable, but it may be impractical or too expensive. Therefore the modelbased method is preferable and consists of the construction of a model representing the behaviour of the system in terms of modelling primitives defined in a formalism. The model of the system must be a convenient abstraction of the system; this means that the level of accuracy of the model must be high enough to represent correctly the aspects of the system

545

In Codetta and Bobbio (2005a), Versions from 1 to 4 have been modelled as Generalized Stochastic Petri Nets (GSPN) (Ajmone, Balbo, Conte, Donatelli, Franceschinis 1995) by means of the GreatSPN tool (Chiola, Franceschinis, Gaeta, and Ribaudo 1995). The GSPN model can undergo analysis, but this requires the liquid level to be discretized in several intermediate integer levels. This is because in GSPN models, only discrete variables can be represented as the number of tokens (marking) inside places. The number of such intermediate levels must not be high; otherwise the state space dimensions may explode, with the consequent increase of the computing cost. Moreover, in the GSPN model, some deterministic timed events such as the action of the pumps or the valve on the liquid level, have to be approximated by stochastic events, in order to allow the model analysis. So, despite of the advantages given by the model analysis instead of simulation, the GSPN model suffers from some approximation about the liquid level and its variations during the time. In Codetta and Bobbio (2005a), Versions from 1 to 4 are modelled also as Fluid Stochastic Petri Net (FSPN) (Gribaudo, Sereno, Horvath, and Bobbio 2001), a particular form of Petri Net including fluid places containing a continuous amount of fluid instead of a discrete number of tokens. Fluid places directly represent continuous variables, such as the liquid level or temperature. FSPN models are typically simulated. This can be done by means of the FSPNedit tool (Gribaudo 2001). Version 5 of the benchmark could not be modelled as a GSPN because the temperature would have been approximated in several intermediate integer values leading to an unacceptable approximation of the current temperature, and consequently to the approximation of the current failure rates. Besides this, the expression of the failure rate as a function of the liquid temperature (Eq. 3 in Sec. 3) cannot be represented in the GSPN model. Moreover, the combination of the possible temperature values together with the possible values of the other parameters describing the system state, would have led to the explosion of the underlying state space dimensions. For this reason, in Codetta and Bobbio (2005b), Version 5 of the system has been modelled and simulated only as a FSPN. In this paper, the benchmark is modelled and simulated using the SAN formalism. The SAN models can undergo analysis as well, but to this aim, the deterministic activities (transitions) have to be replaced by stochastic activities reducing the accuracy of the model with respect to the real behaviour of the system.

becomes impractical because of the high computing cost (and time). For these reasons, dynamic reliability cases are typically evaluated by means of simulation. In this paper, we take into account a benchmark on dynamic reliability taken from the literature (Marseguerra and Zio 1996). The system consists of a tank containing some liquid whose level is influenced by two pumps and one valve managed by a controller, with the aim of avoiding the failure of the system occurring in case of the dry out or the overflow of the liquid. Such events are consequences of the pumps or valve failure because in such condition the components ignore the orders coming from the controller. The dry out or the overflow does not occur as soon as a particular combination of component failures occurs, but such basic failures may influence the liquid level, possibly leading the system to the failure after that some time has elapsed or another event has happened. Because of this, not only the component failure combinations have to be modelled, but also any variation in the liquid level caused by the components action or failure. In Marseguerra and Zio (1996), several versions of the benchmark are proposed: the initial case of state independent failure rates (that we call Version 1), the case of state dependent failure rates (Version 2), the case with possible failure on demand of the controller (Version 3), the case with repairable components (Version 4), and finally the case with temperature dependent failure rates (Version 5). All the versions are described in Sec. 3. In Sec. 5, each version of the system is modelled as a Stochastic Activity Network (SAN) (Sanders and Meyer 2001), a particular form of Stochastic Petri Net; the SAN formalism is described in Sec. 4. The SAN models are designed and simulated by means of the Möbius tool (Deavours, Clark, Courtney, Daly, Derisavi, Doyle, Sanders, and Webster 2002); the aim is to compute the system unreliability in each version of the benchmark (Sec. 6). The advantages of SAN with respect to other forms of Petri Nets (Sec. 2) are presented in Sec. 7. 2. RELATED WORK In Marseguerra and Zio (1996) the unreliability of the system in Versions 1, 2, and 3 is evaluated first in an analytical way by computing the probabilities of the minimal cut sets of component failure events leading to the dry out or the overflow. Then, the system unreliability is evaluated by means of Monte Carlo simulation. The cut set analysis only considers the combinations of events, while the Monte Carlo simulation deals with the complete dynamics of the system: therefore there is a relevant difference between the unreliability values returned by the two approaches, in particular in Versions 2 and 3. Such difference highlights the necessity to take into account the complete behaviour in order to evaluate the system in an accurate way. Versions 4 and 5 are only evaluated by means of Monte Carlo simulation in Marseguerra and Zio (1996).

3. THE BENCHMARK The system (Fig. 1.a) is composed by a tank containing some liquid, two pumps (P1 and P2) to fill the tank, one valve (V) to remove liquid from the tank, and a controller (C) monitoring the liquid level (H) and acting on P1, P2 and V.

546

Table 1: Failure Rates in Versions 1, 3, 4 component failure rate (λ) P1 0.004566 h-1 P2 0.005714 h-1 V 0.003125 h-1

Initially H is equal to 0, with P1 and V in state ON, and P2 in state OFF; since P1, P2 and V have the same level variation rate (Q=0.6 m/h), the liquid level does not change while the initial configuration holds. The cause of a variation of H is the occurrence of a failure of one of the components consisting of turning to the state Stuck-ON or Stuck-OFF. The time to failure is random and obeys the negative exponential distribution; the failure rate (Tab. 1) does not depend on the current state of the component, so the effect of the failure is the stuck condition, while the state transitions toward the StuckON or the Stuck-OFF state are uniformly distributed (Fig. 2.a).

Table 2: The Configuration configuration 1 2 3 4 5 6 7 8

Level P1 ON ON ON ON OFF OFF OFF OFF

Variation P2 OFF ON OFF ON OFF ON OFF ON

in

V OFF OFF ON ON OFF OFF ON ON

each

State

effect on L n nn = n = n p =

Table 3: Control Boundaries and Laws boundary P1 P2 V H < HLA ON ON OFF H > HLB OFF OFF ON

Figure 1. The System Schemes in Versions 1, 2, 3 (a), in Version 4 (b), in Version 5 (c)

3.1. Version 2: state dependent failure rates In this version, the failure rate of a component changes according to its current state and the state reached as a consequence of the failure (Tab. 4). Table 4: Failure Rates for each Component in each State, in Version 2 component from to failure rate (λ) P1 ON Stuck-ON 0.004566/2 h-1 ON Stuck-OFF P1 0.004566/2 h-1 OFF P1 Stuck-ON 0.045662 h-1 OFF Stuck-OFF P1 0.456621 h-1 P2 ON Stuck-ON 0.057142 h-1 ON Stuck-OFF P2 0.571429 h-1 OFF P2 Stuck-ON 0.005714/2 h-1 OFF Stuck-OFF P2 0.005714/2 h-1 V ON Stuck-ON 0.003125/2 h-1 ON Stuck-OFF V 0.003125/2 h-1 OFF 0.031250 h-1 V Stuck-ON OFF Stuck-OFF V 0.312500 h-1

Figure 2: The States of P1, P2 and V in Versions 1, 3, 5 (a). The States of P1, P2 and V in Version 4 (b) Tab. 2 shows how H changes with respect to the current configuration of the component states; the controller believes that the system is correctly functioning while H is inside the region between the levels HLA (-1 m) and HLB (+1 m). If H exceeds HLB, the controller orders both pumps to switch OFF, and the valve to switch ON, with the aim of decreasing H (Tab. 3) and avoiding the liquid overflow; this event occurs when H exceeds the level HLP (+3 m). If a component is stuck, it does not obey the controller order and maintains its current state. The other undesired situation is the tank dry out; this happens when H is below HLV (-3 m); if H goes below HLA, the controller orders both pumps to switch ON, and the valve to switch OFF, with the aim of increasing H (Tab. 3) and avoiding the dry out. The failure of the whole system happens when the dry out or the overflow occurs. We denote such configuration of the system as Version 1. In this paper, we deal with several versions of the benchmark, still proposed in Marseguerra and Zio (1996).

3.2. Version 3: controller failure on demand Here, the controller has a probability of failure on demand equal to 0.1. This means that each time H exceeds the region of correct functioning (HLA t0. In Eq. 1, L0 is the liquid level at time t0; in Eq. 2, T0 is the liquid temperature at time t0. The failure rates of P1, P2 and V are temperature dependent according to Eq. 3 where λ0 is the failure rate of the component for a temperature equal to 20°C (Tab. 5). Besides the dry out and the overflow, another condition determines the failure of the system: T reaches 100°C. The initial level of the liquid in the tank is 7 m; HLA and HLB are set to 6 m and 8 m respectively; HLV and HLP are equal to 4 m and 10 m respectively (Fig. 1.c).

x

x

A SAN model can contain also output gates. The role of an output gate is specifying only the effect of the activity completion on the marking of the places. Therefore an output gate is characterized only by a function. The marking enabling the same activity can be expressed by means of oriented arcs, or by means of an input gate. Gates graphically appear as triangles (input gate: ◄ - output gate: ►). In a SAN model, it is possible to set several completion cases for an activity; each case corresponds to a certain effect of the completion and has a certain probability: when the activity completes, one of the cases happens. A case graphically appears as a small circle close to the activity; from the case an arc is directed to a gate or a place. The completion of an activity can be immediate or timed. In the second case, the completion time can be constant or random. A random completion time has to be ruled by a probability distribution; in this paper, we always resort to the negative exponential one, but several other distributions are available in the SAN formalism. In this paper (and in Codetta (2011)), we call “immediate activity” an activity completing as soon as it is enabled; we call “deterministic activity” an activity whose completion time is deterministic and not immediate; finally, we call “stochastic activity” an activity whose completion time is random.

Table 5: Failure Rates for T = 20°C in Version 5 component O0 P1 0.004566 h-1 P2 0.005714 h-1 V 0.003125 h-1 L(t) = L0 + Q ˜ (t – t0) T(t) = T0 ˜ L0/L(t) + Tin ˜ Q ˜ (t – t0) / L(t) O(T) = O0 ˜ (0.2e0.005756(T-20) + 0.8e-0.2301(T-20))

(1) (2) (3)

Three versions of the benchmark are characterized by the aspects described above: x x x

a predicate consists of a Boolean condition expressed in terms of the marking of the places; if this condition holds, then the activity is enabled to complete. A function expresses the effect of the activity completion on the marking of the places.

Version 5.1: the controller cannot fail. Version 5.2: the controller has a probability of failure on demand equal to 0.2. Version 5.3: initially the controller has a probability of failure on demand equal to 0.2; due to the wear of the controller, such probability is increased of 50% every time that the controller has to act (at each demand).

5. MODELING THE SYSTEM Each version of the benchmark (Sec. 3) has been modelled as a SAN where each aspect of the system behaviour is represented: the state of components, the

548

failure events, the current liquid level and its variations, the orders by the controller, the liquid dry out or overflow, etc.

gate I_V_fail. The initial ON state of V is modelled by the presence of one token inside V_on and no tokens inside V_stuck. The current level (H) of the liquid in the tank expressed in meters, is represented by the extended place Level whose marking is a float variable initially set to 0 corresponding to the initial level of the liquid (Sec. 3). In the SAN model, we model any variation of H by 0.01 m; this is done by increasing or decreasing the marking of Level by 0.01. The action of P1, P2 and V on H are modelled by the deterministic activity Level_variation and in particular by the corresponding input gate I_Level_variation. Such gate enables Level_variation to complete only when the state configuration 1, 2, 4, 6, or 7 holds (Tab. 2). The action of P1, P2 or V on H is ruled by a level variation rate equal to 0.6 m/h (Sec. 3); this means that the action of a pump (valve) increases (decreases) the liquid level by 0.01 m every 0.016667 h. Since we are interested in representing any variation of H by 0.01 m, Level_variation completes every 0.016667 h in state configurations 1, 4, 6, 7, or every 0.016667/2 h in state configuration 2 (Tab. 2). The gate I_Level_variation specifies also the effect of the completion of Level_variation: each time such activity completes, the marking of the place Level is increased by 0.01 in the state configurations 1, 2, 4, 6, or is decreased by 0.01 in the state configuration 7 (Tab. 2). The place normal is initially marked with one token in order to represent that H is inside the region of correct functioning (Sec. 3). The completion of the immediate activity n2a removes the token inside the place normal; according to the input gate I_n2a, this happens if the marking of the extended place Level is less than HLA or more than HLB (Tab. 3). The same gate sets the marking of the place alert to 1. In this way, we model that the liquid level in the tank is outside the correct region. The presence of one token inside alert enables the immediate activity control to complete. The effect of its completion is ruled by the output gate O_control_law executing the control laws in Tab. 3: such gate acts on the marking of the places P1_on, P2_on and V_on, and consequently on the state of P1, P2 and V. So, control together with O_control_law, models the orders given by the controller. If the place P1_stuck, P2_stuck or V_stuck is marked, then the output gate O_control_law has no effect on the place P1_on, P2_on or V_on respectively. In this way we model that the controller cannot act on the state of a stuck component. The controller action on the component states may lead H back to the region of correct functioning. In this case, the immediate activity a2n is enabled to complete by the input gate I_a2n checking that the marking of the extended place Level is equal or greater than HLA and less or equal to HLB. The effect of the completion of a2n is the presence of one token inside the place normal in order to represent that H in inside the region of correct functioning.

5.1. Modelling Version 1 The SAN model of Version 1 is depicted in Fig. 3 where the current state of the pump P1 is represented by means of the places P1_on and P1_stuck: if P1_on is empty, this means that P1 is off; if instead the place P1_on is marked with one token, then P1 is on. The place P1_stuck is used to represent the stuck condition of P1: if such place is empty, then P1 is not stuck; if instead the place P1_stuck contains one token, this means that P1 is currently stuck. According to the marking combinations of the places P1_on and P1_stuck, we can model all the possible states of P1: ON, OFF, Stuck-ON, Stuck-OFF (Sec. 3). Initially the place P1_on is marked with one token, and the place P1_stuck is empty, in order to model that P1 is initially in state ON. The state transitions of P1 caused by its failure are modelled by the stochastic activity P1_fail whose completion rate is equal to the failure rate of P1 (Tab. 1). The completion of such activity is ruled by the input gate I_P1_fail: P1_fail may complete only if the place P1_stuck is empty (P1 is not stuck), while there is no condition about the place P1_on (the failure may occur during both the ON state and the OFF state). The same gate partially specifies the effect of the completion of P1_fail: the gate sets the marking of the place P1_stuck to 1 (P1 becomes stuck), and sets the marking of P1_on to 0. The effect of the activity P1_fail is ruled also by two completion cases: in one case the marking of the place P1_on is not changed, and in this way we model the state transition toward the state Stuck-OFF; in the other case, one token appears in P1_on, in order to represent the state transition toward the state Stuck-ON. These two completion cases have the same probability to occur: 0.5.

Figure 3: The SAN Model of Version 1 The current state of the pump P2 and the state transitions due to a failure of P2 are modelled in the same way by the places P2_on and P2_stuck, the stochastic activity P2_fail and the input gate I_P2_fail. Initially both P2_on and P2_stuck are empty in order to model that P2 is initially in state OFF. The state evolution of the valve V is modelled by the places V_on and V_stuck, the stochastic activity V_fail and the input

549

The dry out and the overflow condition determining the system failure, are detected by the immediate activity fail and in particular by the corresponding input gate I_fail: if the marking of the extended place Level is less than HLV, then one token appears in the place dry_out in order to model the occurrence of the dry out. If instead the marking of Level is greater than HLP, then the effect of the completion of fail is the presence of one token inside the place overflow modelling the occurrence of the overflow. Further details of the SAN model in Fig. 3 are reported in Codetta (2011).

other case, the failure on demand occurs and the only effect is the addition of one token to the marking of the new place controller_ko counting the number of failures of the controller.

Figure 6: The SAN Model of Version 4 5.4. Modelling Version 4 Version 4 is modelled in Fig. 6 where the immediate activity n2a still completes when H exceeds the region of correct functioning, but now n2a inserts also one token inside the new place grace in order to represent that the grace period (Sec. 3.3) has begun. Such marking enables the new stochastic activities P1_repair, P2_repair and V_repair ruled by the input gates I_P1_repair, I_P2_repair and I_V_repair, and modelling the repair of P1, P2 and V respectively. The effect of the completion of such activities is the removal of the token inside the place representing the stuck condition of the component (P1_stuck, P2_stuck and V_stuck respectively). The immediate activity fail still models the system failure, but now it also removes the token inside the place grace, in order to represent the end of the grace period.

Figure 4: The SAN Model of Version 2 5.2. Modelling Version 2 In Version 2 (Sec 3.1), the failure rates of P1, P2 and V are state dependent (Tab. 4). The SAN model of Version 2 appears in Fig. 4 where the current state of P1 is still modelled by the marking of the places P1_on and P1_stuck, but the state transitions caused by the failure are now modelled by three stochastic activities: P1_on_fail models the failure of P1 during the state ON; P1_off_fail_off represents the failure during the state OFF and leading to the state Stuck-OFF; P1_off_fail_on models the failure during the state OFF, but leading to the state Stuck-ON. The state evolution of P2 and V is modelled in a similar way. The other parts of the SAN model in Fig. 5 are the same as in the model of Version 1 (Fig. 3).

Figure 7: The SAN Model of Version 5.1 5.5. Modelling Version 5 The SAN models representing Versions 5.1, 5.2, 5.3 are depicted in Figures 7, 8, 9, respectively. Such models are characterized by the presence of a new extended place called Temperature representing the current liquid temperature (T). The deterministic activity Level_variation, together with the input gate I_level_variation, models the variation of H and T, as a consequence of the heat source (Sec. 3.4) and the injection of new liquid in the tank by the pumps (Eq. 2). The rates of the stochastic activities P1_fail, P2_fail

Figure 5: The SAN Model of Version 3 5.3. Modelling Version 3 In Version 3, the controller failure on demand is introduced (Sec. 3.2); this aspect is represented in the SAN model in Fig. 5 by the presence of two completion cases for the immediate activity control modelling the action of the controller. In one case, the effect of the completion of control is ruled by the output gate O_control executing the control laws (Tab. 3). In the

550

500 h in Version 4, as in Marseguerra and Zio (2006)). The cdf provides the system unreliability (Sec. 1) according to a specific failure condition. For instance, the value of the dry out cdf at time t > 0 is the probability that the system has failed because of the dry out, during the time period (0, t). The cdf of the dry out probability is computed as the mean value over the 100’000 simulation batches, of the marking of the place dry_out present in all the SAN models (Sec. 5). In each simulation batch and at a certain time, the number of tokens inside the place dry_out is equal to 0 if the dry out has not occurred, or it is equal to 1 if the dry out condition holds (Sec. 5). So, the mean value of its marking at a certain time, over the 100’000 simulation batches, provides the probability that the dry out condition holds at that time. The cdf of the overflow probability is computed in the same way, but with reference to the place overflow present in all the SAN models (Sec. 5). In Versions 5.1, 5.2 and 5.3, another system failure condition is taken into account: the temperature of the liquid reaching 100°C (Sec. 3.4). The cdf of such condition is computed as the mean number of tokens inside the place High_Temp present in the SAN models in Fig. 7, 8 and 9 (Sec. 5.5).

and V_fail modelling the failure of P1, P2 and V respectively, are expressed as a function of the marking of Temperature according to Eq. 3. In Versions 5.1, 5.2, 5.3, the system failure condition due to the high temperature of the liquid is introduced (Sec. 3.4). In the SAN models in Fig. 7, 8, 9, such condition is detected by the new immediate activity fail2 ruled by the input gate I_fail2: when the marking of Temperature reaches the value of 100, one token appears inside the new place High_Temp.

Figure 8: The SAN Model of Version 5.2 The Version 5.2 is characterized by the possible failure on demand of the controller (Sec. 3.4). In the SAN model in Fig. 8, such aspect is represented in the same way as in the SAN model of Version 3 (Fig. 5).

6.1. Results for Versions 1, 2, 3 The values of the cdf of the dry out in Versions 1, 2 and 3 (SAN model in Fig. 4, 5 and 6 respectively) are reported in Tab. 6 and are graphically compared in Fig. 10. The values of the cdf of the overflow are reported in Tab. 7 and are graphically compared in Fig. 11. The results returned by the SAN model simulation are similar to the values returned by Monte Carlo simulation, GSPN analysis and FSPN simulation. Figure 9: The SAN Model of Version 5.3

Table 6: The cdf of the Dry Out in Versions 1, 2, 3 time Version 1 Version 2 Version 3 100 h 4.5900E-03 2.0240E-02 4.9090E-02 200 h 2.2390E-02 4.0400E-02 8.6710E-02 300 h 4.4890E-02 5.4090E-02 1.0952E-01 400 h 6.5990E-02 6.3360E-02 1.2664E-01 500 h 8.2600E-02 6.9870E-02 1.3844E-01 600 h 9.5290E-02 7.3750E-02 1.4707E-01 700 h 1.0393E-01 7.6650E-02 1.5313E-01 800 h 1.1003E-01 7.8340E-02 1.5739E-01 900 h 1.1435E-01 7.9440E-02 1.6024E-01 1000 h 1.1747E-01 8.0240E-02 1.6220E-01

In Version 5.3, the probability of failure on demand of the controller is increased by 50% after each demand (Sec. 3.4). In Fig. 9, the marking of the new place demands indicates the number of demands: every time that the immediate activity n2a completes (H reaches the control boundaries), the marking of demands is increased by one. The immediate activity control still has two completion cases, but now their probabilities are a function of the marking of demands. The full details of all the SAN models can be found in Codetta (2011). 6. SIMULATION RESULTS The SAN models presented in the previous section have been simulated. In particular, for each model, 100’000 simulation batches have been performed by means of the Möbius tool, requiring a confidence level equal to 0.95, and a relative confidence interval equal to 0.1. The measures computed by the simulation are the cumulative distribution function (cdf) of the probability of each system failure condition (Sec. 3) for a mission time varying between 0 and 1000 h (or between 0 and

Figure 10: The cdf of the Dry Out in Versions 1, 2, 3

551

Table 7: The cdf of the Overflow in Versions 1, 2, 3 time Version 1 Version 2 Version 3 100 h 7.8880E-02 7.9370E-02 1.3517E-01 200 h 1.9914E-01 1.6852E-01 2.7244E-01 300 h 2.9386E-01 2.3411E-01 3.6541E-01 400 h 3.6207E-01 2.7882E-01 4.2492E-01 500 h 4.0667E-01 3.0943E-01 4.6332E-01 600 h 4.3665E-01 3.2938E-01 4.8808E-01 700 h 4.5683E-01 3.4310E-01 5.0444E-01 800 h 4.7063E-01 3.5284E-01 5.1537E-01 900 h 4.7929E-01 3.6009E-01 5.2298E-01 1000 h 4.8572E-01 3.6500E-01 5.2797E-01

Figure 12: The cdf of the Dry Out in Version 4

Figure 11: The cdf of the Overflow in Versions 1, 2, 3

Figure 13: The cdf of the Overflow in Version 4

6.2. Results for Version 4 The results obtained by simulating the model in Fig. 6, are reported in Tab. 8, in Fig. 12 (dry out) and in Fig. 13 (overflow). They differ from the results returned by Monte Carlo simulation in Marseguerra and Zio (1996), even though they are in the same order of magnitude. Moreover, they differ from the results obtained by means of GSPN analysis and FSPN simulation in Codetta and Bobbio (2005a) where the repair is erroneously assumed to be allowed only while the level is outside the region of correct functioning, instead of during the grace period (Sec. 3.3).

6.3. Results for Version 5 The results for Versions 5.1, 5.2 and 5.3 are reported in Tables 9, 10 and 11 respectively. In particular, the results in Version 5.1 (Fig. 14) and Version 5.2 (Fig. 15) are similar to the values returned by Monte Carlo simulation in Marseguerra and Zio (1996) and FSPN simulation in Codetta and Bobbio (2005b). Version 5.3 was not modelled as a FSPN in the past. According to the results for such version (Fig. 16), the wear of the controller (Sec. 3.4) does not seem to have a relevant impact on the cdf values, with respect to Version 5.2. In Marseguerra and Zio (1996) instead, the controller wear determines a slight increase of the dry out cdf values.

Table 8: The cdf Version 4 time 50 h 100 h 150 h 200 h 250 h 300 h 350 h 400 h 450 h 500 h

of the Dry Out and the Overflow in dry out 0.000E+00 6.000E-05 1.500E-04 2.200E-04 3.300E-04 3.700E-04 4.500E-04 4.500E-04 4.700E-04 5.100E-04

Table 9: The cdf of the Dry Out, the Overflow and the High Temperature in Version 5.1 time dry out overflow high temp. 100 h 3.1650E-02 2.4007E-01 0.0000E+00 7.9330E-02 3.9531E-01 0.0000E+00 200 h 300 h 1.0517E-01 4.5631E-01 0.0000E+00 400 h 1.1706E-01 4.8133E-01 0.0000E+00 500 h 1.2200E-01 4.9161E-01 1.3000E-04 600 h 1.2376E-01 4.9588E-01 3.8200E-02 700 h 1.2424E-01 4.9750E-01 7.3850E-02 800 h 1.2436E-01 4.9826E-01 1.1855E-01 900 h 1.2438E-01 4.9864E-01 1.2614E-01 1.2438E-01 4.9884E-01 1.2724E-01 1000 h

overflow 8.000E-04 2.430E-03 4.230E-03 6.090E-03 7.920E-03 9.460E-03 1.069E-02 1.197E-02 1.298E-02 1.363E-02

552

Figure 14: The cdf of the Dry Out, the Overflow and the High Temperature in Version 5.1

Figure 16: The cdf of the Dry Out, the Overflow and the High Temperature in Version 5.3

Table 10: The cdf of the Dry Out, the Overflow and the High Temperature in Version 5.2 time dry out overflow high temp. 100 h 1.1102E-01 3.3740E-01 0.0000E+00 1.4934E-01 4.7526E-01 0.0000E+00 200 h 300 h 1.6601E-01 5.2228E-01 0.0000E+00 400 h 1.7271E-01 5.4072E-01 2.0000E-05 500 h 1.7559E-01 5.4828E-01 2.2000E-04 600 h 1.7650E-01 5.5178E-01 2.0370E-02 700 h 1.7677E-01 5.5325E-01 4.0250E-02 800 h 1.7685E-01 5.5387E-01 6.7460E-02 900 h 1.7687E-01 5.5416E-01 7.1930E-02 1.7687E-01 5.5428E-01 7.2550E-02 1000 h

7. CONCLUSIONS A benchmark on dynamic reliability taken from the literature has been examined. Each version focuses on a particular aspect of the dynamic behaviour of the system, such as state or temperature dependent failure rates, repairable components, failures on demand. The benchmark was originally evaluated in terms of system unreliability by means of Monte Carlo simulation. In this paper, the benchmark versions have been modelled and simulated using SAN, a particular form of Petri Net. The results in this paper are in general coherent to the original ones and those obtained by means of other Petri Net based formalisms such as GSPN and FSPN. This confirms that Petri Net models are a valid approach to deal with dynamic reliability cases because of the possibility to model the stochastic, timed or immediate events characterizing the complete behaviour of the system. In particular, the use of SAN has several advantages. Gates make the SAN model more compact: many predicates and functions (Sec. 4) that are incorporated into the input or output gates, would have required more transitions (activities) and arcs in order to be represented in a GSPN or FSPN. For instance, in the GSPN models, the orders by the controller are represented by 6 transitions, while in the SAN model, only the activity (transition) control, together with its output gate, is necessary. In the GSPN model, the variations to the liquid level are represented by 5 transitions, while the activity Level_variation is enough in the SAN. The failure of a pump or valve is represented by four transitions in the GSPN; in the SAN instead, one activity is necessary (Sec. 5). The SAN formalism can represent float variables, as in FSPN, by means of extended places. An example is the place Level modelling the liquid level. This avoids the discretization into integer values of float variables, required in GSPN. The negative values of variables can be directly mapped into the marking of SAN places. For instance, the liquid level in Versions 1, 2, 3 varies between -3 m and +3 m (Sec. 3), just like the marking of the place Level (Sec. 5). In the GSPN and FSPN model instead, the liquid thresholds had to be redefined in order to avoid negative values.

Figure 15: The cdf of the Dry Out, the Overflow and the High Temperature in Version 5.2 Table 11: The cdf of the Dry Out, the Overflow and the High Temperature in Version 5.3 time dry out overflow high temp. 100 h 1.1427E-01 3.4165E-01 0.0000E+00 1.5201E-01 4.7762E-01 0.0000E+00 200 h 300 h 1.6799E-01 5.2503E-01 0.0000E00 400 h 1.7470E-01 5.4360E-01 1.0000E-05 500 h 1.7748E-01 5.5109E-01 2.0000E-04 600 h 1.7849E-01 5.5462E-01 1.9230E-02 700 h 1.7878E-01 5.5604E-01 3.7920E-02 800 h 1.7886E-01 5.5663E-01 6.5320E-02 900 h 1.7889E-01 5.5694E-01 6.9780E-02 1.7889E-01 5.5708E-01 7.0460E-02 1000 h

553

flush-out arcs: Modelling and analysis. Discrete Event Dynamic Systems, 11(1&2):97–117. Marseguerra, M., Zio, E., 1996. Monte Carlo Approach to PSA for dynamic process system. Reliability Engineering and System Safety, 52:227–241. Marseguerra, M., Zio, E., Devooght, J., Labeau, P.E., 1998. A concept paper on dynamic reliability via Monte Carlo simulation. Mathematics and Computers in Simulation, 47:371-382. Sahner, R.A., Trivedi, K.S., Puliafito, A. 1996. Performance and Reliability Analysis of Computer Systems; An Example based Approach Using the SHARPE Software Package. Kluwer Academic Publisher. Sanders, W.H., Meyer, J.F., 2001. Stochastic activity networks: Formal definitions and concepts. Lecture Notes in Computer Science, 2090:315– 343.

AUTHOR BIOGRAPHY Daniele Codetta-Raiteri received the Ph.D. in Computer Science from the University of Turin, Italy, in 2006. Now he is a researcher at the University of Eastern Piedmont, Italy. His research focuses on stochastic models for Reliability evaluation, with a particular experience in Fault Trees, Petri Nets and Bayesian Networks. He is the (co-)author of more than thirty papers published in proceedings or journals. REFERENCES Ajmone-Marsan, M., Balbo, G., Conte, G., Donatelli, S., Franceschinis, G., 1995. Modelling with Generalized Stochastic Petri Nets. Wiley Series in Parallel Computing. Chiola, G., Franceschinis, G.,Gaeta, R., Ribaudo, M., 1995. GreatSPN 1.7: Graphical Editor and Analyzer for Timed and Stochastic Petri Nets. Performance Evaluation, special issue on Performance Modeling Tools, 24(1&2):47–68. Codetta-Raiteri, D., 2011, SAN models of a benchmark on dynamic reliability, Università del Piemonte Orientale. Available from: http://people.unipmn.it/dcr Codetta-Raiteri, D., Bobbio, A., 2005a. Solving Dynamic Reliability Problems by means of Ordinary and Fluid Stochastic Petri Nets. Proceedings of the European Safety and Reliability Conference (ESREL), pp. 381–389. June, Gdansk (Poland). Codetta-Raiteri, D., Bobbio, A., 2005b. Evaluation of a benchmark on dynamic reliability via Fluid Stochastic Petri Nets. Proceedings of the International Workshop on Performability Modeling of Computer and Communication Systems (PMCCS), pp. 52–55. September, Turin (Italy). Deavours, D., Clark, G., Courtney, T., Daly, D., Derisavi, S., Doyle, J., Sanders, W., Webster, P.G., 2002. The Möbius Framework and its Implementation. IEEE Transactions on Software Engineering, 28(10):956–969. Distefano, S., Xing, L., 2006. A New Approach to Model the System Reliability: Dynamic Reliability Block Diagrams. Proceedings of the Annual Reliability and Maintainability Symposium (RAMS), pp. 189–195. January, Newport Beach (California, USA). Dugan, J.B., Bavuso, S.J., Boyd, M.A., 1992. Dynamic Fault-Tree Models for Fault-Tolerant Computer Systems. IEEE Transactions on Reliability, 41:363–377. Gribaudo, M., 2001. FSPNEdit: A fluid stochastic Petri net modeling and analysis tool. Proceedings of Tools of International Multiconfernce on Measurements Modelling and Evaluation of computer Communication Systems, pp. 24–28. September, Aachen (Germany). Gribaudo, M., Sereno, M., Horvath, A., Bobbio, A., 2001. Fluid Stochastic Petri Nets augmented with

554

A STOCHASTIC APPROACH TO RISK MODELING FOR SOLVENCY II Vojo Bubevski

Bubevski Systems & Consulting™ TATA Consultancy Services™ [email protected]

Solvency II offers two options for calculating VAR/SCR, i.e. by applying either: i) a standard model, which will be provided by the regulator; or ii) an internal model developed by the (re)insurer’s risk department. The standard models are non-stochastic risk models. They are rather deterministic, scenario-based or covariance models. They are also conservative by nature and generic across the EU so they cannot consider the company’s specific factors. Moreover, they do not use optimisation to determine the minimal variance investment portfolios in order to minimise the financial risk for the (re)insurers. Thus the calculated VAR/SCR will be higher. These are apparent most important limitations of the standard models. For example, the deterministic model applies analytically calculated or estimated input parameters to calculate the results. However, the likelihood of the outcome is not considered at all. Also, the scenariobased models consider the worst, most likely and best case scenarios. However, they fail to answer the two basic questions: i) how likely are the worst and best case scenarios? And more importantly, ii) how likely is the most likely case? Solvency II offers capital-reduction incentives to insurers that invest in developing advanced internal models, which apply a stochastic approach for risk management and control. Thus, insurers will benefit from using internal models. A very good explanation of developing the Enterprise Risk Management (ERM) frameworks in (re)insurance companies for Solvency II is presented in a handbook edited by Cruz (2009). There are a number of published examples of recommended internal model, which could be used for Solvency II (Cruz 2009). These suggested internal model examples don’t consider optimisation to determine the minimal variance investment portfolios in order to minimise the financial risk, which is a significant deficiency. The stochastic models usually apply the Monte Carlo Simulation method, which assigns distributions of random variables to the input parameters and the calculated results are presented in the form of a histogram. This allows statistical and probabilistic tools to be used to analyse the results. A comprehensive

ABSTRACT Solvency II establishes EU-wide capital requirements and risk management standards for (re)insurers. The capital requirements are defined by the Solvency Capital Requirement (SCR), which should deliver a level of capital that enables the (re)insurer to absorb significant unforeseen losses over a specified time horizon. It should cover insurance, market, credit and operational risks, corresponding to the Value-at-Risk (VAR) subject to a confidence level of 99.95% over one year. Standard models are deterministic, scenario-based or covariance-based, i.e. non-stochastic. They don’t optimise the investment portfolios. These are two major deficiencies. A stochastic approach is proposed, which combines Monte Carlo Simulation and Optimisation. This method determines minimal variance portfolios and calculates VAR/SCR using the optimal portfolios’ simulation distributions, which ultimately eliminates the standard models’ deficiencies. It offers (re)insures internal model options, which can help them to reduce their VAR/SCR providing higher underwriting capabilities and increasing their competitive position, which is their ultimate objective. Keywords: Solvency II stochastic model, VAR/SCR reduction, portfolio optimisation – minimal variance, Monte Carlo simulation 1. INTRODUCTION The Solvency II regulations are fundamentally redesigning the capital adequacy regime for European (re)insurers and will be effective from 1st January 2013. Solvency II establishes two levels of capital requirements: i) Minimal Capital Requirement (MCR), i.e. the threshold below which the authorization of the (re)insurer shall be withdrawn; and ii) Solvency Capital Requirement (SCR), i.e. the threshold below which the (re)insurer will be subject to a much higher supervision. The SCR should deliver a level of capital that enables the (re)insurer to absorb significant unforeseen losses over a specified time horizon. It should cover, at a minimum, insurance, market, credit and operational risks, corresponding to the VAR of the (re)insurer’s own basic funds, subject to a confidence level of 99.95% over a one-year period.

555

elaboration of Monte Carlo Simulation in Finance is given by Glasserman (2004). An investment portfolio is defined by the fraction of the capital put in each investment. The problem of determining the minimum variance portfolio that yields a desired expected return was solved by Markowitz in the 1950’s. He received the 1991 Nobel Prize for his work in Economics (Markowitz 1987). Mostly, the Optimisation methodology is used to find the minimum variance portfolio in order to minimise the financial risk. VAR is a widely used financial risk measure. The approach to calculate VAR is well summarised by Jorion (2011). This approach includes the VAR Parametric method and VAR Monte Carlo Simulation method. This paper proposes a stochastic approach to risk modelling for Solvency II. This method applies combined Monte Carlo Simulation and Optimisation methodologies. The method uses Optimisation to calculate the minimal variance portfolios that yield desired expected returns to determine the Efficient Frontier of optimal portfolios. Monte Carlo Simulation is used to calculate VAR/SCR for every portfolio on the Efficient Frontier by using the respective portfolios’ simulation distributions. Therefore, by using the synergy of Monte Carlo Simulation and Optimisation, the method eliminates the deficiencies and limitations, which are identified above. This approach can help (re)insures to develop and improve their internal risk models in order to reduce their VAR/SCR. Consequently, this will provide insurers with higher underwriting capabilities and increase their competitive position, which is their ultimate objective. According to research by Mercer Oliver Wyman, the impact of the four quantifiable risks on the economic capital of insurance companies is: i) 64% Investment Asset Liability Management (ALM) Risk, i.e. Market Risk; ii) 27% Operational Risk; iii) 5% Credit Risk; and iv) 4% Insurance Risk. Considering that the Market (ALM) Risk is the top contributing risk factor, the method is demonstrated by using an example of Market (ALM) Risk Management. Also, in order to facilitate the presentation, a simple Market (ALM) Risk model is demonstrated. Only the practical aspects of the Market (ALM) Risk modelling are discussed. Microsoft™ Excel® and Palisade™ @RISK® and RISKOptimizer® were used in these experiments.

Factor deterministic model, where the risk capital calculation is based on linear combinations of static risk factors. Actually, the model is mostly a Covariance (or VAR) Model, which is a very simplified version of Stochastic Risk Models. 1.1.2. Market Risk in the Swiss Solvency Test (SST) Model This is the standard model of Swiss Federal Office of Private Insurance. The Market Risk in the SST model is handled by the ALM model. The SST ALM model is a Risk Factor Covariance model complemented with Scenario-Based models (SST 2004). 1.1.3. Bourdeau’s Example of Internal Market Risk Model Michele Bourdeau published an example of an internal model for Market Risk. This model calculates VAR using Monte Carlo Simulation. This is an example of a true Stochastic Risk Model (Bourdeau 2009). 2. ALM RISK MODELLING PROCEDURE The following sections demonstrate the ALM Risk modelling procedure for Solvency II step-by-step. Actual financial market data are used in the presentation. 2.1. Problem Statement The following is a simplified problem statement for the demonstrated investment ALM risk model under Solvency II. Determine the minimum variance investment portfolio that yields a desired expected annual return to cover the liabilities of the insurance company. Calculate the VAR considering all the company’s specific factors including their risk appetite. The model should allow the insurer to reduce their VAR (SCR) providing for higher underwriting capabilities and increasing their competitive position. The model should help the company to achieve their ultimate objective. 2.2. Calculating Compounded Monthly Return The monthly returns of four investment funds are available for a period of seven years, i.e. 1990-1996 (Table 1). Note that the data for the period July/1990June/1996 are not shown.

Month Jan/1990 Feb/1990 Mar/1990 Apr/1990 May/1990 Jun/1990 Jul/1996 Aug/1996 Sep/1996 Oct/1996

1.1. Related Work The following is a summary of some published works related to Market (ALM) Risk modelling for Solvency II. 1.1.1. Market Risk in the GDV Model The GDV (Gesamtverband der Deutschen Versicherungswirtschaft) Model is the standard model of the German Insurance Association for Solvency II (GDV 2005). This model is to some extent a Static

556

Table 1: Monthly Return Fund 1 Fund 2 Fund 3

0.048 0.066 0.022 0.027 0.112 Ͳ0.02 0.086 0.067 0.089 0.036

Ͳ0.01 0.096 0.022 Ͳ0.04 0.116 Ͳ0.02 Ͳ0.07 0.026 Ͳ0.03 0.117

Ͳ0.06 0.037 0.12 Ͳ0.02 0.123 Ͳ0.04 Ͳ0.12 0.146 Ͳ0.04 0.049

Fund 4

Ͳ0.01 0.038 0.015 Ͳ0.04 0.075 Ͳ0.01 Ͳ0.02 0.018 0.092 0.039

The Compounded Monthly Return (CMR) is calculated for each month and each investment fund from the given Monthly Return (MR) fund using the following formula (Table 2):

Table 3: Correlation Matrix CMR1 CMR2 CMR3 CRM 1 CRM2 CRM3 CRM4

CMR = ln (1 + MR) Table 2: Compounded Monthly Return (CMR) Month CMR1 CMR2 CMR3 CMR4 Jan/1990 0.047 Ͳ0.01 Ͳ0.06 Ͳ0.01 Feb/1990 0.063 0.092 0.036 0.038 Mar/1990 0.021 0.022 0.113 0.015 Apr/1990 0.027 Ͳ0.04 Ͳ0.02 Ͳ0.04 May/1990 0.106 0.11 0.116 0.073 Jun/1990 Ͳ0.02 Ͳ0.02 Ͳ0.04 Ͳ0.01 Jul/1996 0.082 Ͳ0.07 Ͳ0.13 Ͳ0.02 Aug/1996 0.065 0.026 0.136 0.018 Sep/1996 0.085 Ͳ0.03 Ͳ0.04 0.088 Oct/1996 0.036 0.111 0.048 0.038

1 0.263 0.038 0.087

0.263 1 0.244 0.089

0.038 0.244 1 0.095

CMR4

0.0868 0.0895 0.095 1

2.5. Generating Compounded Monthly Return The Compounded Monthly Return (CAR) is randomly generated for each investment fund from the best fit distribution considering the correlations. The following distribution functions of the Palisade™ @RISK® are used: CMR1=RiskLogistic(0.0091429,0.044596)) CMR2=RiskLognorm(1.1261,0.077433,Shift(-1.1203)) CMR3= RiskWeibull(6.9531,0.46395, Shift(-0.42581))

2.3. Fitting Distributions to Compounded Monthly Return For the Monte Carlo method, we need the distribution of the compounded monthly return for each investment fund. Thus, for each investment fund, we determine the best fit distribution based on the Chi-Square measure. For example, the best fit distribution for the compounded monthly return of Fund 4 (i.e. CMR4) is the normal distribution presented in Figure 1.

CMR4= RiskNormal(0.0060531,0.047225) The correlation is applied by using the “RiskCorrmat” function of the Palisade™ @RISK®. 2.6. Calculating Compounded Annual Return by Fund The Compounded Annual Return (CAR) is calculated for each investment fund from the respective Compounded Monthly Return (CMR), using the following formula: CAR = 12*CMR 2.7. Calculating Expected Annual Mean Return on the Portfolio The expected annual mean return on the portfolio (EAPR-Mean) is calculated from the asset allocation weights vector (Weights-V) and the vector of compounded annual returns of funds (CAR-V) by using the following Excel® formula: EAR-Mean = SumProduct(Weights-V, CAR-V) 2.8. Calculating Variance, Standard Deviation and VAR of the Portfolio The variance, standard deviation and VAR of the portfolio are calculated from the distribution of the expected annual mean return on the portfolio (EARMean) by using the following Palisade™ @RISK® functions:

Figure 1: Fund 4 Best Fit Distribution 2.4. Finding Compounded Monthly Return Correlations The compounded monthly returns of the investment funds are correlated. We need to find the correlation to allow the Monte Carlo method to generate correlated random values for the compounded monthly returns. The correlation matrix is presented in Table 3.

Variance = RiskVariance(EAR-Mean) Standard-Deviation = RiskStdDev(EAR-Mean)

557

VAR = RiskPercentile(EAR-Mean,0.005) 2.8.1. Portfolio Simulation and Optimisation #1 Palisade™ RISKOptimizer® is used to solve the portfolio simulation and optimisation problem. That is to find the minimal variance portfolio of investments, which yields sufficient return to cover the liabilities. Thus, the aim of the simulation and optimisation model is to minimise the variance of the portfolio subject to the following specific constraints: x The expected portfolio return is at least 8.2%, which is sufficient to cover the liabilities; x All the money is invested, i.e. 100% of the available funds is invested; and x No short selling is allowed so all the fractions of the capital placed in each investment fund should be non-negative.

Figure 2: Probability Distribution #1 The probability that the portfolio return is below 7.5% is 49.2%. There is a 33.5% probability that the return is in the range of 7.5%-50%. From the simulation statistics we also find that there is a 43.2% probability that the portfolio return is negative, and 51.4% probability that the return is below 10%. From the correlation graph (Figure 3), we can conclude that the portfolio return is most dependent on the return of Fund 4 with a correlation coefficient of 77%. The other three funds, i.e. Fund 3, Fund 2 and Fund 1, are less influential with correlation coefficients of 48%, 46% and 44% respectively.

The model should also calculate the Standard Deviation and VAR of the portfolio. 2.8.2. Finding the Efficient Frontier of Portfolios Palisade™ RISKOptimizer® is used repetitively to solve the portfolio simulation and optimisation problem in order to find the Efficient Frontier of investment portfolios. That is to find the minimal variance portfolios of investments, which yield expected portfolio returns of at least 8.4%, 8.6%, …, 10% and 10.2%. Thus, the aim of the simulation and optimisation models is to find in ten iterations, the ten minimal variance portfolios subject to the following specific constrains: x The expected portfolio return is at least 8.4%, 8.6%, …, 10% and 10.2% respectively; x All the money is invested, i.e. 100% of the available funds is invested; and x No short selling is allowed so all the fractions of the capital placed in each investment fund should be non-negative. The model should also calculate the Standard Deviation and VAR of these ten portfolios. 3.

RESULTS AND DISCUSSION

3.1. Simulation and Optimisation #1 The optimal portfolio found by this model has the following investment fractions: 14.6% in Fund 1; 11.6% in Fund 2; 18.6% in Fund 3; and 55.2% in Fund 4. The Portfolio Return is 8.2% with Variance of 19.9%, Standard Deviation of 44.6% and VAR of -7%. The probability distribution of this optimal portfolio is given in Figure 2. From the graph, we can read the confidence levels as follows.

Figure 3: Correlation Sensitivity The regression sensitivity graph is given in Figure 5. This graph shows how the portfolio mean return is changed in terms of Standard Deviation, if the return of a particular fund is changed by one Standard Deviation.

558

Figure 5: Efficient Frontier of Optimal Portfolios

Figure 4: Regression Mapped Values Sensitivity

The Efficient Frontier shows that an increase in expected return of the portfolio causes an increase in portfolio Standard Deviation. Also, the Efficient Frontier gets flatter as expected. This tells us that each additional unit of Standard Deviation allowed, increases the portfolio mean return by less and less.

Therefore, we can read from the graph for example, that if Fund 4 return is changed by one Standard Deviation, the portfolio return will be changed by 0.3142 Standard Deviations (as shown by the regression coefficient of 0.3142). Again, the other three funds, i.e. Fund 3, Fund 2 and Fund 1, are less influential as their regression coefficients are 0.1639, 0.1408 and 0.1080 respectively.

3.4. Portfolio Expected Mean Return versus VAR Figure 6 shows the dependency of the expected portfolio returns against VAR. From the graph we can see that an increase in expected return of the portfolio causes an increase in portfolio VAR in terms of money. (It should be noted that mathematically, VAR is a negative number, which actually decreases when the return increases.) Also, the curve on the graph gets flatter, again as expected. This tells us that each additional unit of VAR allowed, increases the portfolio mean return by less and less.

3.2. Overall Simulation & Optimisation Results The overall results of all the eleven simulation and optimisations are presented in Table 4 showing the Mean Return, Variance, Standard Deviation and VAR of the optimal portfolios. Portfolio No. 1 2 3 4 5 6 7 8 9 10 11

Table 4: The overall results Mean Variance Standard Return Deviation

0.082 0.084 0.086 0.088 0.09 0.092 0.094 0.096 0.098 0.1 0.102

0.199 0.202 0.204 0.215 0.222 0.246 0.266 0.295 0.33 0.37 0.42

0.446 0.45 0.452 0.464 0.471 0.496 0.516 0.543 0.574 0.608 0.648

VAR

Ͳ0.067 Ͳ0.088 Ͳ0.11 Ͳ0.147 Ͳ0.223 Ͳ0.257 Ͳ0.308 Ͳ0.403 Ͳ0.498 Ͳ0.608 Ͳ0.753

3.3. Efficient Frontier of Portfolios Efficient Frontier of the optimal portfolios is presented in Figure 5. Figure 6: Portfolios Return versus VAR 3.5. Portfolio Standard Deviations versus VAR Figure 7 shows the dependency of the portfolio Standard Deviation against VAR.

559

3.7.2. The Simulation & Optimisation Method versus the Swiss Solvency Test (SST) Model The SST ALM model is a Risk Factor Covariance model complemented with Scenario-Based models. Therefore, The SST ALM Model has the same deficiencies as the Scenario-Based models and the Covariance Models, which are not true Stochastic Models. In addition, the SST ALM Model does not apply optimisation to minimise the variance of the investment portfolios, which is another major deficiency. The proposed method has eliminated these deficiencies by using the synergy of the Mote Carlo Simulation and Optimisation methodologies. Therefore, the proposed approach is also superior to the SST ALM Model. 3.7.3. The Simulation & Optimisation Model versus Bourdeau’s Internal Market Risk Model Example The Market Risk internal model proposed by Michele Bourdeau is a true Stochastic Risk Model. However, it does not use optimisation to minimise the variance of the investment portfolio, which is a main limitation. In this sense, the proposed method has a significant advantage versus this example because it minimises the variance of the investment portfolio, which ultimately minimises the risk and VAR.

Figure 7: Standard Deviation versus VAR From the graph we can see that the portfolio VAR is almost linearly proportional to the Standard Deviation. This is also as expected because a higher Standard Deviation translates to a higher risk, thus the VAR also increases in money terms (decreases mathematically). 3.6. Decision Support The results presented above provide for comprehensive and reliable decision support for the decision makers, i.e. the financial risk executives of the insurance company. In particular, considering the Efficient Frontier of portfolios and the dependencies between portfolio expected return, Standard Deviation and VAR (shown in Figure 5, Figure 6 and Figure 7), the decision maker can decide in which assets to invest according to the desired expected return, risk appetite (i.e. standard deviation) and VAR. These results can help to reduce the SCR as required.

4. CONCLUSION This paper proposed a stochastic method for risk modelling under Solvency II. The method combines Monte Carlo Simulation and Optimisation methodologies in order to manage financial risk. The Optimisation methodology is used to calculate the minimal variance portfolios that yield desired expected returns in order to determine the Efficient Frontier of portfolios. The Monte Carlo methodology is used in order to calculate VAR/SCR for every portfolio on the Efficient Frontier by using the respective portfolios’ simulation distributions. Consequently, the synergy of the Monte Carlo Simulation and Optimisation methodologies, which are used by the method, eliminates the identified significant limitations of the standard models. Also, the method has a significant advantage against the internal models, which do not use simulation and optimisation methodologies. This stochastic approach can help the insurance and reinsurance companies to develop or improve their Solvency II internal risk models in order to reduce their VAR/SCR. Reducing the VAR and SCR will ultimately provide the insurance and reinsurance firms with higher underwriting capabilities, which will increase their competitive position on the market. Moreover, the proposed method can significantly assist the insurance and reinsurance companies to achieve their business objectives.

3.7. The Simulation & Optimisation Approach Comparison with the Related Work Examples A comparison of the Simulation and Optimisation method proposed in this paper with the related work examples summarized in Sec. 1.1 is given below. 3.7.1. The Simulation & Optimisation Method versus the GDV Model The GDV Model inherits its limitations from the Static Factor deterministic model. Also, this model is a very simplified Stochastic Model, which is an additional limitation. Moreover, the model doesn’t use Optimisation to minimise the variance of the investment portfolios of the insurer, which is another major limitation. In contrast, the proposed method does not have these two major limitations because they are resolved by using the Monte Carlo Simulation and Optimisation methodologies. Thus, the proposed approach is superior to the GDV Model.

ACKNOWLEDGMENTS I would like to thank my daughter, Ivana Bubevska, for reviewing the paper and suggesting relevant

560

improvements. She has also substantially helped with the editing and formatting of the paper. Her contribution has been essential to the successful publication of this work.

is also a specialist in Business Systems Analysis & Design (Banking & Insurance) and has delivered major business solutions across several organizations. He has received several formal awards and published a number of written works, including a couple of textbooks. Vojo has also been featured as a guest speaker at several prominent conferences internationally. .

REFERENCES Bourdeau, M., 2009. Market Risk Measurement Under Solvency II. In: Cruz, M., ed. The Solvency II Handbook, London, Risk Books – Incisive Media, 193–226. Cruz, M., 2009. The Solvency II Handbook, London, Risk Books – Incisive Media. GDV, 2005. Discussion Paper for a Solvency II Compatible Standard Approach, Gesamtverband der Deutschen Versicherungswirtschaft. Available from: http://www.gdv.de/Downloads/English/Documenta tion_Sol_II.pdf [Accessed 20 June 2011] Glasserman, P., 2004. Monte Carlo Methods in Financial Engineering, New York, Springer Science. Jorion, P., 2011. Financial Risk Manager Handbook, New Jersey, John Wiley & Sons. Markowitz, H.M., 1987. Mean-Variance Analysis in Portfolio Choice and Capital Markets, Oxford, UK, Basil Blackwell. SST, 2004. White Paper of the Swiss Solvency Test, Swiss Federal Office of Private Insurance. Available from: http://www.naic.org/documents/committees_smi_i nt_solvency_switzerland_sst_wp.pdf [Accessed 20 June 2011] AUTHORS BIOGRAPHY Vojo Bubevski comes from Berovo, Macedonia. He graduated from the University of Zagreb, Croatia in 1977, with a degree in Electrical Engineering Computer Science. He started his professional career in 1978 as an Analyst Programmer in Alkaloid Pharmaceuticals, Skopje, Macedonia. At Alkaloid, he worked on applying Operations Research methods to solve commercial and pharmaceutical technology problems from 1982 to 1986. In 1987 Vojo immigrated to Australia. He worked for IBM™ Australia from 1988 to 1997. For the first five years he worked in IBM™ Australia Programming Center developing systems software. The rest of his IBM™ career was spent working in IBM™ Core Banking Solution Centre. In 1997, he immigrated to the United Kingdom where his IT consulting career started. As an IT consultant, Vojo has worked for Lloyds TSB Bank in London, Svenska Handelsbanken in Stockholm, and Legal & General Insurance in London. In June 2008, he joined TATA Consultancy Services Ltd. Vojo has a very strong background in Mathematics, Operations Research, Modeling and Simulation, Risk & Decision Analysis, Six Sigma and Software Engineering, and a proven track record of delivered solutions applying these methodologies in practice. He

561

DESIGN AND IMPLEMENTATION OF A FUZZY COGNITIVE MAPS EXPERT SYSTEM FOR OIL PRICE ESTIMATION

A. Azadeh (a), Z.S. Ghaemmohamadi (b) (a)

Department of Industrial Engineering and Center of Excellence for Intelligent Based Experimental Mechanics, College of Engineering, University of Tehran, Tehran, Iran (b) Department of Industrial Engineering, College of Engineering, University of Tehran, Iran (a) [email protected], (b) [email protected]

the forecasts (Azadeh et al. 2010). For crude oil price forecasting, Mirmirani and Li (2004) applied VRA and ANN techniques to make ex-post forecast of US oil price movement. Lagged oil price, lagged oil supply, and lagged energy consumption were used as three endogenous variables for VAR-based forecast. Ye et al. (2006) provided a model to forecast crude oil spot prices in the short-run using high- and low-inventory variables. They showed that the non-linear-term model better captures price responses at very high- or very low-inventory levels and improves forecasting capability. Wang et al. (2005) proposed a new integrated methodology-TEI@I methodology and showed a good performance in crude oil price forecasting with back propagation neural network (BPNN) as the integrated technique. Xie et al. (2006) proposed a support vector regression (SVR) model to predict crude oil price. Similarly, Shambora and Rossiter (2007) and Yu et al. (2007) also used the ANN model to predict crude oil price. Yousefi et al. (2005) introduces a wavelet-based prediction procedure and market data on crude oil is used to provide forecasts over different forecasting horizons. Sadorsky (2006) uses several different univariate and multivariate statistical models such as TGARCH and GARCH to estimate forecasts of daily volatility in petroleum futures price returns. Amin-Naseri and Gharacheh (2007) proposed a hybrid AI approach integrating feedforward neural networks, genetic algorithm, and kmeans clustering, to predict the monthly crude oil price and obtain better results. In this paper, we develop a fuzzy cognitive maps expert system model for forecasting monthly crude oil spot prices using readily available data. The objective of this model is to provide a forecast of monthly West Texas Intermediate (WTI) prices using readily available data. In addition, this paper examines the feasibility of applying fuzzy cognitive maps expert system in crude oil price forecasting through the contrast with ANN, ANFIS and GA models. The rest of the paper is organized as follows: Section 2 describes fuzzy cognitive maps expert system method for crude oil price prediction. To evaluate the fuzzy cognitive maps expert system, a main crude oil

ABSTRACT The objective of this study is to design a fuzzy cognitive maps expert system for estimation of monthly oil price based on intelligent approaches and meta heuristics. Oil price is influenced by several elements, such as politic and social factors. In this paper a fuzzy cognitive maps (FCMs) approach is presented in order to explore the importance of these factors in oil price estimation. To this purpose, causal relationship between affective factors and oil price are depicted and relationship values between them are computed. The proposed expert system utilizes Genetic Algorithm (GA), Artificial Neural Network (ANN) and Adaptive Neural Fuzzy Inference System (ANFIS). The system is coded in .net environment by C# and Matlab and Excel are also used linked for data processing and evaluation. The expert system identifies the preferred method (from GA, ANN, ANFIS) through mean absolute percentage error (MAPE). Keywords: Expert system, Fuzzy cognitive maps, oil price, Estimation 1. INTRODUCTION Crude oil, sometimes called the blood of industries, plays an important role in any economies (Fan et al. 2008). The role of oil in the world economy becomes more and more significant because nearly two-thirds of the world’s energy consumption comes from the crude oil and natural gas (Alvarez-Ramirez et al. 2003). The crude oil price is basically determined by its supply and demand, and is strongly influenced by many events like the weather, inventory, GDP growth, refinery operable capacity, political aspects and people’s expectation. Sharp oil price movements are likely to disturb aggregate economic activity, volatile oil prices have been considerable interest to many researchers and institutions. Therefore, forecasting oil prices is an important and very hard topic due to its intrinsic difficulty and practical applications (Wang et al. 2004). There is an array of methods that are available today for forecasting energy price. An appropriate method is chosen based on the nature of the data available and the desired nature and level of detail of

562

price series, West Texas Intermediate (WTI) crude oil spot price is used to test the effectiveness of the proposed methodology, and its comparable results with ANN, ANFIS and GA methods. Some concluding remarks are made in section 4.

(Kosko 1986). Three paths connect C1 to C5, so there are three indirect effects of C1 on C5: along path P1 ( C1 , C 2 , C 4 , C 5 ) : I1 (C 1 , C 5 )

min

^e 12

I 2 (C 1 , C 5 )

min

^e 13

, e 24 , e 45 along path P2 ( C 1 , C 3 , C 5 ) :

2. FUZZY COGNITIVE MAPS EXPERT SYSTEM FOR CRUDE OIL PRICE In this section, a fuzzy cognitive maps expert system method for time series forecasting and its application in crude oil price prediction are presented. We apply ANN, ANFIS and GA in this fuzzy cognitive maps expert system model. Then present the fuzzy cognitive maps expert system method for oil price forecasting.

, e 35

`

3

(C

1

,C

)

5

min

^e 13

some

much

along path P3 ( C 1 , C 3 , C 4 , C 5 ) : I

`

, e 34 , e 45

`

some

Thus, the total effect of C1 to C5 is:

max ^I 1 ( C 1 , C 5 ), I 2 ( C 1 , C 5 ), I 3 ( C 1 , C 5 )`

T (C1 , C 5 )

2.1. Fuzzy Cognitive Maps Expert System Expert system technology has proven to benefit decision making process in businesses and accounting management of corporations. Most applications are developed in production/operations management area with lowest number of applications in the human resources area (Mearns et al. 2003). There are several applications in the area of diagnosis. They include defects diagnostic system for tire production and service (Prez-Carretero et al. 2002). Benefits of an expert system approach to productivity analysis include cost reductions due to the reduced need for manpower, faster analysis of pressing productivity problems, and more consistent appraisals and interpretation of productivity performance (Azadeh et al. 2008).

much

Figure 1: Example of FCM 2.1.2. Data Fuzzification Fuzzification is a process in which the input data, precise or imprecise is converted into linguistic formation, which is easily perceptible by the human minds (Wagner et al. 2001). All relationship between concepts (indicators of the proposed oil price volatility estimation) are linguistic variables. The most typical fuzzy set membership function has the graph of a triangle. The fuzzy set membership function of our model is also a triangle. This approach translates the point ( x1*,...,x*n ) in set A to a fuzzy set A' as shown in (1). Fuzzy sets for relationship between concepts are shown in Figure 2.

2.1.1. Fuzzy Cognitive Maps (FCM) Cognitive maps (CMs) were introduced by Axelrod (1976) in the 1970s. CMs are signed diagraphs designed to represent the causal assertions and belief system of a person (or group of experts) with respect to a specific domain, and use that statement in order to analyze the effects of a certain choice on particular objectives. Two elements are used when realizing CMs: concepts and causal belief. Concepts represent the variables that describe the belief system of a person, while the causal belief consists in the causal dependencies between variables. Such variables can be continuous, ordinal or dichotomous (Kardaras and Karakostas, 1999). In signed cognitive maps, each relationship is linked to a sign that represents the sense of casual influence of the cause variable on the effect variable. Fuzzy cognitive map is a well-established artificial intelligence technique that incorporates ideas from artificial neural networks and fuzzy logic. FCMs were introduced by Kosko (1986) to extend the idea of cognitive maps by allowing the concepts to be represented linguistically with an associated fuzzy set rather than requiring them to be precise. In order to describe the degree of the relationship between concepts it is possible to use a number between [0,1] and [-1, 1], or use linguistic terms, such as “often”, “always”, “some”, “a lot”, etc. Figure 1 shows an example of FCM used by Kosko to define the indirect and the total effects for an FCM

P A' ( x)

­ x  x1* xn  x n* °1  1 ,...,1  , ® b1 bn ° 0, ¯

-High 1

-Middle

-Low

(1)

xi  xi* d bi else

zero

+Low

+Middle

+High

0.8

0.6 Degree of membership 0.4

0.2

0 -1

-0.8

-0.6

-0.4

-0.2

0 Influence

0.2

0.4

0.6

0.8

1

Figure 2: The fuzzy set for relationship between concepts

563

are premier than other factors, according to experts' ideas. Eight experts draw casual graphs between these factors and volatility oil price. The experts use linguistic terms for expressing relationships between factors. Taber (1991) suggested a relation to unify different judgments of experts, based on the credibility weight of each expert. In our applications, each expert had the same credibility. At the end of the process the result of different factors effects on volatility oil price appears as in Table 1.

2.1.3. Data Defuzzification Defuzzification is the process of producing a quantifiable result in fuzzy logic (Wilson et al. 1992). There are some methods for defuzzification such as centroid average (CA), center of gravity (CG), maximum center average (MCA), mean of maximum (MOM), smallest of maximum (SOM), largest of maximum (LOM) (Wong et al. 1995).

Our fuzzy cognitive maps expert system uses center of gravity (CG). This is because the approach provides better solution than other methods via data engine. We have the center of gravity (CG) as shown in (2):

y'

³A yP A ( y)dy ³A P A ( y)dy '

Table 1: The effects of different factors on volatility oil price The effect on Effective factors volatility oil price World oil demand 0.78 Reduce excess capacity 0.67 Agiotage 0.42 Devaluation dollar 0.83 Financial Crisis 0.25 Government changes and - 0.16 internal turmoil Environmental policy 0.29 OPEC cuts oil production 0.4 Natural events 0.25

(2)

'

2.2. The Proposed Fuzzy Cognitive Maps Expert System An excellent approach is fuzzy cognitive maps expert system that can implement crude oil price forecasting in the volatile crude oil market. The flow chart of the fuzzy cognitive maps expert system is shown in Figure 3. From Figure 3, the fuzzy cognitive maps expert system for crude oil price forecasting consists of some main components, i.e., graphical user interface module (GUI), oil price forecasting with ANN, ANFIS, GA module, oil price volatility correction with fuzzy cognitive maps module and integration module. GUI: It is a graphical window through which users can exchange information with the fuzzy cognitive maps expert system and also users enter necessary data in system. In details, it handles all input/output between users and the fuzzy cognitive maps expert system. Oil price forecasting with ANN, ANFIS and GA module: in this study ANN, ANFIS and GA predict the future value of oil price using the historical data. The crude oil prices data are used in this paper are monthly spot prices of West Texas Intermediate (WTI) crude oil. For a univariate time-series forecasting problem, the inputs of the network are the past lagged observations of the data series and the outputs are the future values. According to Pierson coefficient of correlation, oil price of a month before and oil price of two months before are chosen for oil price forecasting. Oil price forecasting accomplish with three methods, ANN, ANFIS and GA, then according to MAPE best forecasting is chosen between three methods. The parameters in these methods are chosen based on previous studies and also using Trial and error. Oil price volatility correction with fuzzy cognitive maps module: Crude oil market is an unstable market with high volatility and oil price is often affected by many related factors (Wang et al. 2004). In this paper we used from nine factors that they

Integration module: Crude oil price forecasting obtains by implementing, oil price forecasting with ANN, ANFIS, GA module and oil price volatility correction with fuzzy cognitive maps module. Indeed crude oil price forecasting obtains by adding two values of two modules.

Figure 3: The overall computational flow chart of the fuzzy cognitive maps expert system

564

log( xt )  log( xt 1 )

1

( xt  xtc ) 2 n

¦t

n 1

n

MAPE

1

,

(5)

( xt  xtc ) 2 n

¦t

,

xt  x c xt n

All methods, except MAPE have scaled output. MAPE method is the most suitable method to estimate the relative error because input data used for the model estimation, preprocessed data and raw data have different scales (Azadeh et al. 2011). 3.4. Results and Analysis Result of Each of the forecasting method described in the last section is presented in Table 2 in which a comparison among them is performed. Each of methods is estimated and validated by train data. The model estimation selection process is then followed by an empirical evaluation which is based on the test data. Table 2 shows the detailed results of the simulated experiment via the four methods. It can be seen that the fuzzy cognitive maps expert system method outperforms other models in term of MAPE. Focusing on the MAPE indicators, the values of fuzzy cognitive maps expert system model are explicitly lower than those of ANN, ANFIS and GA except in the third subperiod. The main reasons for the above conclusions are as follows. As Panas et al. (2000) reported, the crude oil market is one of the most volatile markets in the world and shows strong evidence of chaos. All methods can in principle describe the nonlinear dynamics of crude oil price. Best method between ANN, ANFIS and GA for different periods is different, so, a hybrid method that it uses from all methods is necessary and useful. Fuzzy cognitive maps expert system is resistant to the overfitting problem and can model nonlinear relations in an efficient and stable way.

(3)

However, for the first difference of the logarithm method the transformation is adjusted as follows:

yt

¦t

n

RMSE

3.2. Data Preprocessing As in time-series methods making the process, covariance stationary is one of the basic assumptions and also using preprocessed data is more useful in most heuristic methods (Zhang et al. 2005), and so the stationary assumption should be studied for the models. In time series forecasting, the appropriate preprocessing method should have two main properties. It should make the process stationary and have post processing capability. The most useful preprocessed methods are presented in the sections. The first difference method: The difference method was proposed by Box et al. (1994) In this method, transformation should be applied:

xt  xt 1

MSE

xt  xtc

1

n

3.1. Data The crude oil price data used in this study are monthly spot prices of West Texas Intermediate (WTI) crude oil from January 1991 to December 2010 with a total of n = 240 observations. These data include train data and test data. The train data are 216 (90 percent of total data as usual) observations and the test data are 24 (10 percent of total data as usual) observations. The main reason of selecting this oil price indicator is that this crude oil price is the most famous benchmark price, which is used widely as the basis of many crude oil price formulae (Yu et al. 2008). The crude oil price data used in this study are obtainable from the energy information administration (EIA) website of Department of Energy of USA (http://www.eia.doe.gov).

yt

MAE

¦t

n

3. A CASE STUDY In this section, we first describe the data, and then define some evaluation criteria for prediction purposes. Finally, the empirical results are presented.

(4)

Table 2: Crude oil forecast results according to mape SubSubSubSubFull period period period period period 4 3 2 1 Method (1991(1991- (1996- (2001- (20062010) 2010) 2005) 2000) 1995)

Normalization: There are different normalization algorithms which are Min-Max Normalization, Z-Score Normalization and Sigmoid Normalization. We used these methods to estimate time series functions and finally according to mean absolute percentage error (MAPE) we didn't apply any methods for preprocessing. 3.3. Evaluation Criteria There are four basic error estimation methods which are listed: Mean absolute error (MAE), Mean square error (MSE), Root mean square error (RMSE) and Mean absolute percentage error (MAPE). They can be calculated by the following equations, respectively:

565

ANN

0.0594

0.0256

0.0588

0.0525

0.0222

ANFIS

0.1534

0.0392

0.0658

0.334

0.0328

GA

0.0654

0.047

0.0799

0.0457

0.0285

FCM expert system

0.0474

0.0245

0.0423

0.0386

0.0216

planning process. Information and Software Technology, 41, 197–210. Kosko, B., 1986. Fuzzy cognitive maps. International Journal of Man–Machine Studies, 24, 65–75. Taber, R., 1991. Knowledge processing with fuzzy cognitive maps. Expert Systems with Applications, 2, 83–87. Wang, Sh., Yu, L., Lai, K. K., 2004. A Novel Hybrid AI System Framework for Crude Oil Price Forecasting. CASDMKM, LNAI 3327, 233–242. Mearns, M., Whitaker, S. M., Flin, R., 2003. Safety climate, safety management practice and safety performance in offshore environments. Safety Science, 41, 641–680. Mirmirani, S., Li, H.C., 2004. A comparison of VAR and neural networks with genetic algorithm in forecasting price of oil. Advances in Econometrics, 19, 203–223. Panas, E., Ninni, V., 2000. Are oil markets chaotic? A non-linear dynamic analysis. Energy Economics, 22, 549-568. Prez-Carretero, C., Laita, L. M., Roanes-Lozano, E., Lazaro, L., Gonzalez-Cajal, J., Laita, L., 2002. Logic and computer algebra based expert system for diagnosis of anorexia. Mathematics and Computers in Simulation, 58, 183–202. Sadorsky, P., 2006. Modeling and forecasting petroleum futures volatility. Energy Economics, 28, 467–48. Shambora, W. E., Rossiter, R., 2007. Are there exploitable inefficiencies in the futures market for oil?. Energy Economics, 29, 18–27. Wang, S. Y., Yu, L.A., Lai, K. K., 2005. Crude oil price forecasting with TEI@I methodology. Journal of Systems Science and Complexity, 18, 145-166. Wagner, W.P., Najdawi, M.K., Chung, Q.B., 2001. Selection of knowledge acquisition techniques based upon the problem domain characteristics of production and operations management expert systems. Expert Systems, 18, 76–87. Wilson, J.R., Corlett, E.N., 1992. Evaluation of Human Work: A Practical Ergonomics Methodology. Taylor and Francis, USA, 141–169. Wong, B.K., Monaco, J.A., 1995. A bibliography of expert system applications for business (1984– 1992). European Journal of Operational Research 85, 416–432. Xie, W., Yu, L., Xu, S.Y., Wang, S.Y., 2006. A new method for crude oil price forecasting based on support vector machines. Lecture Notes in Computer Science, 3994, 441 451. Ye, M., Zyren, J., Shore, J., 2006. Forecasting short-run crude oil price using high- and low-inventory variables. Energy Policy, 34, 2736–2743. Yousefi, S., Weinreich, I., Reinarz, D., 2005. Waveletbased prediction of oil prices. Chaos, Solitons and Fractals, 25, 265–275. Yu, L., Lai, K.K., Wang, S.Y., He, K.J., 2007. Oil price forecasting with an EMD-based multi scale neural

Actual values and forecasting values with fuzzy cognitive maps expert system for test data are shown in Figure 3. 90 Predicted

80

Actual

70 Dollar per barrel 60

50 40 30 0

5

10

15

20

25

Data Number

Figure 3: WTI crude oil price forecast based on fuzzy cognitive maps expert system for test data REFERENCES Alvarez-Ramirez, J., Soriano, A., Cisneros, M., Suarez, R., 2003. Symmetry/anti-symmetry phase transitions in crude oil markets, Physica A, 322, 583-596. Amin-Naseri, M. R., Gharacheh, E. A., 2007. A hybrid artificial intelligence approach to monthly forecasting of crude oil price time series. The Proceedings of the 10th International Conference on Engineering Applications of Neural Networks, CEUR-WS 284, 160–167. Axelrod, 1976. Structure of Decision. Princeton University Press, Princeton, NJ. Azadeh, A., Asadzadeh, S. M, Ghanbari, A., 2010. An adaptive network-based fuzzy inference system for short-term natural gas demand estimation: Uncertain and complex environments. Energy Policy, 38, 1529–1536. Azadeh, A., Fam, I. M., Khoshnoud, M., Nikafrouz, M., 2008. Design and implementation of a fuzzy expert system for performance assessment of an integrated health, safety, environment (HSE) and ergonomics system The case of a gas refinery. Information Sciences, 178, 4280–4300. Azadeh, A., Saberi, M., Asadzadeh, S.M., 2011. An adaptive network based fuzzy inference system– auto regression–analysis of variance algorithm for improvement of oil conjsumption estimation and policy making The cases of Canada, United Kingdom, and South Korea. Applied Mathematical Modelling, 35, 581–593. Box, G.E.P., Jenkins, G.M., Reinsel, G.C., 1994. Time Series Analysis: Forecasting and Control. Prentice-Hall, Englewood Cliffs, NJ. Fan, Y., Liang, Q., Wei, Y. M., 2008. A generalized pattern matching approach for multi-step prediction of crude oil price. Energy Economics, 30, 889-904. Kardaras, D., Karakostas, B., 1999.The use of cognitive maps to simulate the information systems strategic

566

network learning paradigm. Lecture Notes in Computer Science, 4489, 925–932. Yu, L., Wang, Sh., Lai, K. K., 2008. Forecasting crude oil price with an EMD-based neural network ensemble learning paradigm. Energy Economics, 30, 2623–2635. Zhang, G.P., Oi, M., 2005. Neural network forecasting for seasonal and trend time series. European Journal of Operational Research, 160, 501–514. AUTHORS BIOGRAPHY Ali Azadeh is an Associate Professor and the founder Department of Industrial Engineering and co-founder of RIEMP at the University of Tehran. He graduated with first class honor Degree (BS) in Applied Mathematics from the University of San Francisco and obtained his MS and PhD in Industrial and Systems Engineering from the San Jose State University and the University of Southern California. He received the 1992 Phi Beta Kappa Alumni Award for excellence in research and innovation of doctoral dissertation in USA. He is the recipient of the 1999–2000 Applied Research Award and has published more than 300 academic papers. Zeinab Sadat Ghaemmohamadi is currently a Graduate student in Industrial Engineering at the University of Tehran. He earned his BS in statistics from Isfahan University of Technology and his MS in Industrial Engineering from Tehran University, Iran. His current research interests include systems modeling and forecasting of supply, demand and price of energy via artificial intelligence tools.

567

HOW TO BENEFIT MORE FROM INUITIVE POWER AND EXPERIENCE OF THE HUMAN SIMULATION KNOWLEDGE STAKEHOLDER Gaby Neumann Faculty of Engineering / Industrial Engineering Technical University of Applied Sciences Wildau [email protected]

ABSTRACT Generally, it is pretty clear and widely accepted that the human actor plays a significant role in any simulation project – although in recent years some authors proclaimed a revival of human-free simulation at least related to distinct parts of a simulation study. Therefore, the paper aims to provide an overview on needs and challenges in model-user interaction as well as on approaches, methods and tools to support the user in bringing in his/her knowledge in all phases of a simulation project from model building via understanding a model and using it for experimentation to correctly interpreting simulation outcome. Furthermore, barriers and problems hindering a simulation stakeholder in sharing his/her knowledge are identified and approaches to access and extract such knowledge are discussed in order to avoid inefficiency and failure in future projects.

situation just because of their individual knowledge and experience. Consequently, the individual background significantly impacts the whole range of a simulation project from model building till interpretation of results (see Figure 1).

Keywords: knowledge-based simulation, simulation knowledge, discrete event simulation, knowledge management

Figure 1: Impact of the simulation user on the outcome of a simulation project

aim of the simulation questions addressed to the simulation

simulation model

simulation output

simulation system

findings from simulation

methodical support

knowledge/experience of the simulation user

Neumann and Ziems (1997) went into detail with identifying human simulation knowledge stakeholders’ impact on certain simulation project stages. According to this, simulation experts are primarily responsible for model building and implementation steps, whereas domain experts mainly provide application-specific knowledge for problem description, identification of input data and evaluation of results. This corresponds to the type of knowledge and experience brought into a simulation project and gained from a simulation project by the different actors. Therefore, simulation needs to be understood in its entire characterization as complex problem-solving, knowledge-generation and learning process at the same time. This view is in line with literature characterizing modeling and simulation in general as both, knowledgeprocessing activity and goal-directed knowledgegeneration activity (Ören 1990). Based upon this, advanced methodologists and technologists were expected to be allowed to integrate simulation with several other knowledge techniques. But looking at today’s situation in simulation projects it still has to be considered that a sound application of a knowledge

1. INTRODUCTION AND MOTIVATION The impact of a person’s knowledge and background on the design, level-of-detail and focus of the simulation model, i.e. on the way a simulation model appears and functions, was demonstrated, for example, by Neumann and Page (2006). Here, two groups of students with different background (computing vs. logistics) but the same level of simulation knowledge and experience were assigned with the same problem to be investigated. In the end both student projects produced valid and usable simulation models, but efforts for model implementation, model modification in the course of experimentation and visualization of results were quite different. Results achieved from either model equally allowed responding to the initial questions addressed to the simulation project; from this it was possible to conclude that despite of different modeling approaches simulation results are comparable and of similar quality. This way, the case study gave proof of the fact that different persons with different background might produce different but in the same way correct and usable simulation models of the same problem and

568

management perspective to modeling and simulation is still missing. Instead, the term ‘knowledge-based simulation’ is typically used for applying AI approaches to automatically create simulation models from expert knowledge. Research focuses, for example, on developing efficient and robust models and formats to capture, represent and organize the knowledge for developing conceptual simulation models that can be generalized and interfaced with different applications and implementation tools (Zhou, Son and Chen 2004). Other work aims to develop concepts for modeling human decisions, e. g. in manufacturing systems (Zülch 2006), or to model and simulate human behavior to support workplace design by use of digital human models (Monteil et al. 2010) and especially by incorporating human characteristics like fear, aggressiveness, fatigue and stress in particularly challenging situations (Bruzzone et al. 2010). In contrast to this, the fact that non-formalized expert knowledge finds its way into the simulation model on one hand or is created throughout the simulation lifecycle and needs to be externalized on the other is not in the focus of research in this field. That is why, information about decisions taken when building the model or running experiments as well as really new knowledge about the particular application or even about the simulation methodology gained in the course of a simulation project quite often stays in the heads of the people involved in the project. Furthermore, the simulation model itself also forms a kind of dynamic repository containing knowledge about parameters, causal relations and decision rules gathered through purposeful experiments. This knowledge is being somewhat hidden as long as not being discovered, understood and interpreted by another person. Against this background, research on implementing a knowledge management perspective in simulation projects should address the following questions: x x x x

x

quality and with the right costs at the right place and time. In other words, it is essential not to focus on the introduction of knowledge management technology and integration of software tools for storing and retrieving knowledge and information only, but to put the human resources running model building and simulation projects into the centre of gravity and to try to give them that kind and amount of support which is needed in a particular situation. Therefore, the paper aims to provide an overview on needs and challenges in model-user interaction (Section 2) as well as approaches, methods and tools to support the user in bringing in his/her knowledge in all phases of a simulation project from model building via understanding a model and using it for experimentation to correctly interpreting simulation outcome (Section 3). Barriers and problems hindering a simulation stakeholder in sharing his/her knowledge are identified and approaches to access and extract such knowledge are discussed (Section 4). Findings are summarized and conclusions are drawn in Section 5. 2.

NEEDS AND CHALLENGES IN MODELUSER INTERACTION Once a valid simulation model is available it serves as tool for different types of studies: x

x

Which information and knowledge is needed by whom at what stage of a simulation project? Which knowledge and information is provided by whom in which step of a simulation project? Which knowledge is generated with whom in which step of the simulation project? Which knowledge is “stored” in the conceptual and simulation models, evolves from simulation experiments, and is “hidden” in the input/output data of simulation runs? How simulation knowledge with the different stakeholders or repositories can be accessed, extracted, externalized and distributed, shared, applied?

x

In a what-if analysis the user discovers how a system reacts on changing conditions or performance requirements, i.e. system loads. During experimentation a particular type of changes is introduced to the model in a systematic way in order to understand sensitivity of a certain parameter, design or strategy. A what-to-do-to-achieve investigation aims to answer questions like how to set system parameters or how to improve process control in order to reach a certain behavior or performance level. Experimentation might be multidimensional including different types of changes to the model; it is strongly oriented towards identifying modification strategies for reaching a particular performance objective or target behavior. Performance optimization experiments serve to solve a particular target function such as minimizing job orders’ time in system or stock level, maximizing service level or resources’ utilization, etc. Here, the limits of typical performance characteristics are to be identified with the respective limit value itself forming the goal of the investigation.

No matter which type of investigation is on the agenda the user always needs to interact with the model in order to implement the intended experimentation strategy and to gain simulation results. Interaction prior to the simulation run (or a batch of simulation runs) might consist in adjusting the structure

To generalize research needs, the biggest challenge for properly handling modeling and simulation knowledge by applying knowledge management methods and tools consists in providing the right knowledge of the right

569

of the simulation model, in purposefully changing one or more model or simulation parameters, or even simply in starting the simulation in order to produce and collect simulation output data that are expected to be of use for the investigation. Post-run interaction focuses on accessing and dealing with simulation output data in the form of dynamic visualization (i.e. watching animations) or statistical analysis (i.e. checking original or condensed data, viewing diagrams or other types of graphical representation) in order to achieve findings with regard to the focus and aim of the investigation. Consequently, the entire interaction cycle can be characterized as a user-model dialogue: any pre-run interaction with the model corresponds to the concept of asking questions; post-run interaction is adequate to the concept of responding to questions. Pre-condition for a successful user-model dialogue is true understanding in both directions. The simulation model needs to “understand” what the user is interested in and looking for. This requires the ability to ask the right questions from the user. Those questions might either be very specific and clearly matching “the language of the model” (i.e. directly addressing input/output data of a simulation) or they are of more principle, general, eventually even fuzzy nature requiring a kind of translation for being understandable to the model. When it comes to the responding part of the dialogue the user needs to understand the simulation output for getting the answers s/he was looking for.

principle has to be transferred to the data level by explaining it in detail and putting it in terms of concrete data (see Figure 2). As result of this process of interpretation a set of specific questions is defined with each of them providing a specific part of the overall answer in which the user is interested. Questions at data level correspond to results that can be delivered directly by the simulation even if minor modifications to the simulation model should be required (Tolujew 1997). This is the kind of study also current approaches for automatic trace file analysis in order to better cope with large amounts of simulation output data support (Kemper and Tepper 2009, Wustmann et al. 2009). Those approaches mainly focus on formalizing simulation outcome in the context of a certain application area. With this they remain at data level, whereas deriving answers of principle to questions of principle requires processing further the respective set of specific answers. These steps of additional analysis and condensing can be understood as a process of reinterpretation to transfer results from data to user level. All steps of interpretation and re-interpretation aim to link the user’s point of view to that of the simulation model. They not only require an appropriate procedure, but, even more importantly, an interpretative model representing the application area in which simulation takes place. This model needs to be based on knowledge and rules expressed in the user’s individual expertise, but also in generalized knowledge of the problem environment regarding design constraints or system behavior and the experience of the model building expert derived from prior simulations. This knowledge might not only be of explicit nature, i.e. existing independent of a person and suitable to be articulated, codified, stored, and accessed by other persons, but also comprises implicit or tacit knowledge carried by a person in his or her mind often not being aware of it. Whereas explicit knowledge might be transferred into rules and algorithms, tacit knowledge cannot be separated from its owner and therefore requires direct involvement of the knowledge holder in the interpretation process. In the end, knowledge stored in the simulation model can be considered proven, independently of whether it was developed by the domain expert him- or herself or by a consultant simulation expert (Neumann and Ziems 1997). Unfortunately, this knowledge is usually not very well documented and therefore does exist implicitly only inside the simulation model. To be used when the results of the simulation project are put into practice, it needs to be explained in such a way as to be accessible to the domain expert in the subjectspecific terminology and to be applicable without any loss of information or misrepresentation. Otherwise the technical or organizational solution in the real world cannot be expected to work in the way demonstrated by the respective simulation model or knowledge important for the realization of simulated functionality needs to be re-developed by renewed implementation and testing.

user of the simulation model

?

!

principle answer

interpretation

re-interpretation

data level

???

specific specific question specific question question

interpretation layer

user level

principle question

! ! !

specific specific specific answer answer answer

test for data availability

proposal of watch dogs

filter for data

analysis of data

model level modification of model simulation run

trace file

converting

data

Figure 2: User-data interaction for simulation output analysis When the potential interests a simulation user might have in a simulation study are compared, one significant difference emerges: specific questions formulated by the user might directly be answered with concrete simulation output at data level; those usually fuzzy questions of principle from the more global user’s point of view require interpretation and re-interpretation steps before being answered. Here, any question of

570

creation and acquisition at the same time without too much additional effort for all involved (see Figure 3).

3.

METHODS AND TOOLS FOR BRINGING IN SIMULATION KNOWLEDGE Human resources involved in a simulation project are the key factors for its success and efficiency. As discussed in the previous section it is always up to the simulation user to define objectives of any simulation and target functions of any experimentation. For this detailed knowledge and understanding on the particular system/process to be investigated and problem to be solved is needed as well as sound background knowledge on the domain and experiences in simulation-based problem solving. As this individual knowledge and experience belongs to the person carrying it and continuously develops and grows over time with each new simulation project, it can be separated from the person, i.e. externalized, to some extent only. Therefore, a mix of methods and tools for bringing in a user’s knowledge and experience into the simulation project is needed: x

x x

x x x

contribution to

knowledge management

problem solving benefit from

structured documentation

knowledge logistics

identification of knowledge generation of knowledge use of knowledge demand for knowledge

model building running experiments interpreting results

simulation model

specifying the simulation problem

supply of knowledge

evaluating the simulation outcome

simulation portal

Figure 3: Problem solving and knowledge acquisition in the course of a simulation project The clue to the successful implementation of those knowledge management procedures is often an appropriate (supporting) environment and climate in the organization. Concerning this, there is a greater need for a cultural shift than for additional software tools and IT solutions. Adopting a statement on human needs for computer technology by Shneiderman (2002) the link between knowledge management and (simulationbased) problem solving can generally be described as follows: the old discussion about how to support problem solving is about what (software) tools can do; the new discussion about how to support problem solving is (and must be) about what kind of problemsolving support people really need.

Formalize what can be formalized and incorporate this into simulation tools completed by a rule-based supporting system and an interface for its continuous improvement. Apply algorithms to routine problem-solving (Kemper and Tepper 2009, Wustmann et al. 2009). Enable a structured dialogue between the user and the tool by applying the concept of oraclebased simulation model validation (Helms and Strothotte 1992). Provide support in structured documentation of problem, model, experiments, solution/findings and lessons learned (Neumann 2006). Use human intuition and tacit knowledge for all that cannot be formalized (yet). Allow the user to bring in his/her ability of flexible thinking for problems and questions that unexpectedly pop-up in the course of a simulation study (Tolujew et al. 2007).

4.

PROBLEMS AND BARRIERS IN SIMULATION KNOWLEDGE SHARING In the course of a simulation project there are bidirectional links between activities for problem solving and knowledge management. On one hand knowledge available with persons, inside organizations and in the form of technology is (re-)used to build a model, plan and run experiments, analyze and understand simulation output. On the other hand knowledge about the problem’s final solution and the chosen mode of action for its generation characterizes the increased scientific basis and additional experience of the problem-solving person, team or organization. Usually these links are based upon the persons directly involved in the simulation project. It’s quite common to make use of own experience, but to benefit from knowledge, experience and lessons learned from other parts of the organization that is still not the usual procedure yet. To overcome this and to make knowledge of a successful or even unsuccessful problem solving process available to future simulation projects that is the challenge for knowledge management and its integration into personalized problem solving. Being aware of this, organizations invest a large amount of money in technology to better leverage information, but often the deeper knowledge and

Here, it is crucial to initiate an ongoing learning and improvement process as basis of structured knowledge explication and gathering of experiences similar to what has been proposed by Brandt et al. (2001) for software engineering projects. Applying this approach to learning from simulation projects, a welldefined and well-structured documentation of both simulation model and simulation runs and the simulation project with all its assumptions, agreements, and decisions has to be established (seamlessly and continuously). Procedures help to identify who knows what about the system and process, but also about the simulation project behind it, why something was decided in which way, which system configuration and which set of parameters work well together, what is in the simulation model and what the limitations of its validity and usability are. With this the process of a simulation project becomes a process of knowledge

571

expertise that exists within the organization remains untapped. The sharing of knowledge remains limited in most respects, and at least, strained. APQC (2004) sees major reasons for this in technology that is too complicated and the human nature that poses barriers to knowledge sharing. Cultural aspects can enhance an open knowledge transfer or inhibit a positive attitude towards knowledge sharing. Taking cultural aspects into consideration requires letting the knowledge management approach – and with this the knowledge sharing process in particular – fit the culture, instead of making an organization’s culture fitting the knowledge management approach (McDermott and O’Dell 2000). In a perfect world the benefits of accessing and contributing knowledge would be intrinsic: people who share knowledge are better able to achieve their work objectives, can do their jobs more quickly and thoroughly, and receive recognition from their peers and mentors as key contributors and experts. Nevertheless, knowledge is often not shared. O'Dell and Grayson (1998) identified four common reasons for this: x

x

x

x

1989): in general an apprentice is a learner who is coached by a master to perform a specific task. Based on this, the theory transfers the traditional apprenticeship model as known from crafts, trade and industry to the cognitive domain. More precise, cognitive apprenticeship aims at externalizing processes that are usually carried out internally. This approach works with methods like modeling, coaching, scaffolding, articulation, reflection and exploration. Coaching, for example, is to be understood as helping a person in actively creating and successfully passing individual learning processes through guidance-ondemand. In the end, the coach (i.e. the expert) offers support in case of difficulties (i.e. scaffolding), provides hints, feedback and recommendations, and eventually takes over certain steps for solving the given problem. However, the coach only appears when explicitly being called by the person to be coached (i.e. like a help system) and the scaffolding is gradually fading as the learning novice proceeds. So, coaching seems to be a very useful concept for sharing and developing simulation knowledge in practice as it aims to develop heuristic strategies through establishing a culture of expertise and with this goes far beyond pure learning as typically provided in workplace learning environments.

Ignorance. Those who have knowledge don't realize others may find it useful and at the same time someone who could benefit from the knowledge may not know another person in the company already has it. No absorptive capacity. Many times, an employee lacks the money, time, and management resources to seek out information they need. Lack of pre-existing relationship. People often absorb knowledge from other people they know, respect, and like. If two managers don't know each other, they are less likely to incorporate each other's experiences into their own work. Lack of motivation. People do not see a clear business reason for pursuing the transfer of knowledge.

5. SUMMARY AND CONCLUSIONS To the same extent as a new simulation project provides another challenge to model building, experimentation and interpretation of results it rarely can be planned comprehensively and in all details. Therefore, the simulation knowledge stakeholder cannot be fully replaced by algorithms in a simulation project. Instead his/her intuitive power and experience is needed to appropriately and creatively cope with the unexpected. Here, challenges typically consist in enabling or strengthening purposeful interaction between the simulation model and its user, supporting the user in bringing in his/her simulation knowledge, and overcoming barriers hindering in distributing and sharing knowledge and experience for extending the organizational simulation knowledge base and speeding up the learning curve in human resource development. The paper presents approaches for dealing with those challenges from a knowledge management perspective. Here, the focus is clearly put on the methodological aspect, whereas implementation into simulation tools or supportive systems remains an open task. Against this background the main message of the paper consists in underlining the key role a human resources play in simulation projects – no matter if we talk about simulation experts, experts from the application area or even novices to those fields. Despite of this, there are many useful methods, concepts and algorithms even coming from other areas that should be applied to simulation-based investigations in order to support the simulation knowledge stakeholder in more efficient and effective problem-solving and sustainable knowledge explication.

To meet these challenges, the discipline of knowledge sharing should continuously be reinforced. For this, there are two different approaches: the organization might host visible knowledge-sharing events to reward people directly for contributing to knowledge or the organization might rely on the link between knowledge sharing and everyday work processes by embedding knowledge sharing into "routine" work processes. Here, initiating of a close, interpersonal link between a mentor or coach (the expert) and the novice is a promising way not to rely on enthusiasm only, but to bring in a personal commitment to the process of developing another person’s simulation competence. Those expert-novice links might also be part of learning processes to improve an individual’s simulation competence in a learning-by-doing scenario. The pedagogical framework for this is formulated by the cognitive apprenticeship theory (see Collins et al.

572

ergonomic evaluation for the operational improvement of the slate splitters work. Proc. of European Modeling and Simulation Symposium, pp. 191-200. October 13-15, Fes (Morocco). Shneiderman, B., 2002. Leonardo's Laptop: Human Needs and the New Computing Technologies. Cambridge: The MIT Press. Tolujew, J., 1997. Werkzeuge des Simulationsexperten von morgen. Proc. of Simulation and Animation, pp. 201-210. March 6-7, Magdeburg (Germany). (Tools of the simulation expert of tomorrow, in German) Tolujew, J., Reggelin, T., Sermpetzoglou, C., 2007. Simulation und Interpretation von Datenströmen in logistischen Echtzeitsystemen. EngelhardtNowitzki, C., Nowitzki, O., Krenn, B., eds. Management komplexer Materialflüsse mittels Simulation. State-of-the-Art und innovative Konzepte. Wiesbaden: Deutscher Universitäts-Verlag, pp. 215-232. (Simulation and interpretation of data streams in logistics real-time systems, in German) Wustmann, D., Vasyutynskyy, V., Schmidt, T., 2009. Ansätze zur automatischen Analyse und Diagnose von komplexen Materialflusssystemen. Proc. Expert Colloquium of WGTL, pp. 1-19. October 12, Ilmenau (Germany). (Approaches for automatic analysis and diagnosis of complex material flow systems, in German). Zhou, M., Son, Y.J., Chen, Z., 2004. Knowledge representation for conceptual simulation modeling. Proc. of Winter Simulation Conference, pp. 450458. December 5-8, Washington (D.C., USA). Zülch, G., 2006. Modelling and simulation of human decision-making in manufacturing systems. Proc. of Winter Simulation Conference, pp. 947-953. December 3-6, Monterey (California, USA).

REFERENCES APQC, 2004. Failures in knowledge sharing. APQC American Productivity and Quality Center. Brandt, M., Ehrenberg, D., Althoff, K., Nick, M., 2001. Ein fallbasierter Ansatz für die computergestützte Nutzung von Erfahrungswissen bei der Projektarbeit. Proceedings 5. Internationale Tagung Wirtschaftsinformatik. September 19-21, Augsburg (Germany). (A case-based approach for computer-based use of experience-based knowledge in projects, in German) Bruzzone, A., Madeo, F., Tarone, F., 2010. Modelling country reconstruction based on civil military cooperation. Proc. of European Modeling and Simulation Symposium, pp. 315-322. October 1315, Fes (Morocco). Collins, A., Brown, J. S., Newman, S. E., 1989. Cognitive apprenticeship: Teaching the crafts of reading, writing, and mathematics. Resnick, L. B., ed. Knowing, learning, and instruction. Hillsdale, NJ: Lawrence Erlbaum Associates, pp. 453-494. Helms, C., Strothotte, T., 1992. Oracles and Viewpoint Descriptions for Object Flow Investigation. Proc. of EUROSIM Congress on Modelling and Simulation, pp. 47-53, September 30 – October 3, Capri (Italy). Kemper, P., Tepper, C., 2009. Automated trace analysis of discrete-event system models. IEEE Transactions on Software Engineering, 2 (35), pp. 195-208. McDermott, R., O’Dell, C., 2000. Overcoming the ‚Cultural Barriers’ to Sharing Knowledge. APQC - American Productivity and Quality Center. Neumann, G., 2006. Projektwissen in der Logistiksimulation erschließen und bewahren: Auf dem Weg zu einer neuen Dokumentationskultur. Proc. of Simulation in Production and Logistics, pp. 341-350. September 26-27, Kassel (Germany). (Gaining and storing project knowledge in logistics simulation: on the way towards a new documentation culture, in German). Neumann, G., Page, B., 2006. Case study to compare modelling and simulation approaches of different domain experts. Proc. of the International Mediterranean Modelling Multiconference, pp. 517-522. October 4-6, Barcelona (Spain). Neumann, G., Ziems, D., 1997. Transparente Modelldokumentation and Resultatpräsentation schafft Vertrauen. Proc. of Simulation and Animation, pp. 237-250. March 6-7, Magdeburg (Germany). (Transparent model documentation and presentation of results increases trust, in German) O'Dell, C., Grayson, J., 1998. If only we knew what we know. Free Press. Ören, T.I., 1990. A paradigm for artificial intelligence in software engineering. Advances in Artificial Intelligence in Software Engineering, vol. 1, pp. 155. Rego Monteil, N., del Rio Vilas, D., Crespo Pereira, D., Rios Prado, R., 2010. A simulation-based

AUTHOR BIOGRAPHY Gaby Neumann received a Diploma in Materials Handling Technology from the Otto-von-GuerickeUniversity of Technology in Magdeburg and a PhD in Logistics from the University of Magdeburg for her dissertation on “Knowledge-Based Support for the Planner of Crane-Operated Materials Flow Solutions”. Between December 2002 and June 2009 she was Junior Professor in Logistics Knowledge Management at the Faculty of Mechanical Engineering there. In December 2009 she became Professor on Engineering Logistics at the Technical University of Applied Sciences Wildau. Since 1991 she also has been working as part-time consultant in material handling simulation, logistics planning and specification of professional competences in certain fields of logistics. Her current activities and research interests are linked amongst others to fields like problem solving and knowledge management in logistics simulation. She has been or is being involved in a couple of research projects in these fields. Gaby Neumann has widely published and regularly presents related research papers at national and international conferences.

573

OBJECT-ORIENTED MODELLING AND VERIFICATION AIDED BY MODEL SIMPLIFICATION TECHNIQUES Anton Sodja

Borut Zupanˇciˇc

Faculty of Electrical Engineering, University of Ljubljana Tržaška 25, 1000 Ljubljana, Slovenia [email protected] [email protected]

ABSTRACT

general design situations. The error bounds on model prediction are relatively large and validation with experimental data is usually not possible due to the lack of measurements. Therefore, the model is only verified with general design experience and comparison with earlier models of systems with similar characteristics (MurraySmith, 2009). Later in the model building process, as more data becomes available, more complex description of the system are integrated into the model. However, more thorough validation may even indicate that some parts of the model are unnecessarily detailed and model is consequentially simplified. The model building process may be also carried out in reverse: all known details about the systems are included into a model and in the further modeling phases the unnecessarily parts of the model are identified and removed (Murray-Smith, 2009).

Object-oriented modeling approach brought efficient model reuse and thus possibility to create rich model libraries which enable rapid development of large heterogeneous models. However, verification and debugging of large complex models is becoming and increasingly challenging task.Furthermore, model should not be more complicated as needed for a given purpose. A suitable component describing a subsystem in sufficient detail should be selected from a library which might contain several components describing the same system but with different level of detail. Benefits of model simplification techniques for objectoriented model development are discussed in this paper. A modeler may help himself with them in a decision making process of how detailed components should be used (e.g., how complicated model should be) and also as an assistance for verifying model by some informal verification methods. Simplified model should be represented in the same way as original, therefore, two simplification techniques are discussed, simplification of object-diagrams and simplification of equations, which are the usual representation of models in Modelica, one of the commonly used object-oriented modeling language today.

The emergence of the object-oriented modeling exacerbated the difficulty of assuring proper model complexity and model-verification in some respects. Objectoriented models are built of inter-connected components (models of subsystems). Large collections of various preprepared components are available in designated model libraries (Tummescheit, 2002; Andres et al., 2009; Cellier, 1991). Therefore, a very complex model can be built easily. However, this kind of modeling paradigm has some pitfalls, for example, when the modeler is not familiar with the assumptions made at formulation of the components, they can be used incorrectly, e.g., resulting model is invalid due to incompatible components (especially when they come from different libraries). Furthermore, at different modeling stages, different level of model complexity is required and hence a set of components describing the same subsystem but with different level of detail must be present. Another option is to include switches into the components which enables to turn off unnecessary level of details (Casella et al., 2006). When a model consists of many components (as it is the case with large models) it can be quite intricate to determine how complex each component should be. The complexity of the components with low impact on overall

Keywords: nonlinear model simplification, model verification, Modelica 1.

INTRODUCTION

Dynamic models are an important part of many engineering applications. Various purposes which models have in engineering applications include assistance for system and control design, explaining complex-system behavior and help in operators training. The purpose of the model also determines a desirable complexity of the model. Modeling is an iterative process and the complexity of the model usually increases during the process as well as the model purpose tends to change. In the early stages of model development relatively simple conceptual models are used which help examine some 574

model behavior should be of course kept low to bound overall model complexity. 2.

Prior to simulation, model must be translated. First, model is flattened (i.e., hierarchical structure of a model is mapped into a set of differential, algebraic and discrete equations together with the corresponding variable declarations (Modelica Association)), then some further modifications (e.g., tearing of algebraic loops and DAE index reduction) are performed so that model can be brought into the form required by numerical solvers. There are no tools known to the authors which support simplification of Modelica models directly. Model must be thus exported to external tools using either Functional Mockup Interface (Blochwitz et al., 2011) or by reparsing flattened model (if this feature is supported by modeling environment). This kind of an approach have many downsides, because model is flattened or even other translation steps are performed (algebraic-loop tearing and index reduction) before the export. Therefore much information about the organization of the model (for example, hierarchical structure) is lost and the simplified models may not be convenient for verification purposes. Furthermore, it may be very intricate and laborious to simplify only certain submodel and evaluate it together with whole model. We believe that model simplification tool should be closely integrated into modeling environment and representation of a simplified model should represented in the same way as original model. Models are in Modelica generally represented graphically (i.e., object diagrams) on a higher levels and textually (i.e., equations) on the lowest. Therefore, there are needed two classes of simplification techniques, simplification of object-diagrams and simplification of equation sets. In many cases only rankings of elements might be sufficient and simplification of the model could be done manually if needed. Simplification of submodels should be supported as only subset of model might be under consideration.

MODEL SIMPLIFICATION TECHNIQUES

Engineers use experience and intuition to determine important parts of the model which have the highest impact on system’s dominant dynamics or on the model’s simulation response in specific scenario. In an attempt to diminish reliance on subjective factors such as experience, numerous model simplification and order-reduction methods have been developed (Schwarz et al., 2007; Sommer et al., 2008; Lall et al., 2003; Louca, 1998; Chang et al., 2001). Model simplification techniques consist of running a series of simulations, ranking the individual coordinates or elements by the appropriate metric and removing those that fall below a certain threshold (Chang et al., 2001). The choice of the ranking metric and simplification steps depends on modeling formalism used and may be limited by modeling domains. A special class of model simplification methods are those that produce proper models, i.e., simplified models have physically meaningful variables and parameters. It is obvious how model simplification tools can help us determine if the model is too complicated: if model could be simplified a lot without loosing too much accuracy, it is clearly too complicated. Model simplification methods can also facilitate model verification when less formal approaches such as deskchecking are used (Sodja and Zupanˇciˇc, 2010). Even rather small models and their simulation results can be to large to be human interpretable and understandable. Hence, physically interpretable simplified models could be used instead to provide a deeper insight into the system’s behavior needed for model verification. In some cases when only a part of the model (a submodel) is under consideration, it is desired that only this submodel could be simplified. 3.

4.

SIMPLIFICATION OF MODELICA MODELS

SIMPLIFICATION OF MODELICA OBJECTDIAGRAMS

Object diagrams consist of connected components (submodels). A connection defines interactions which are determined by types of connectors (i.e., ports) which are used in components. Connector are rather loosely defined in Modelica. In general, a list of variables with some qualifications (e.g., causality, type of variable: intensive – extensive, etc.) is defined, but it can also have a hierarchical structure (Modelica Association).

Modelica modeling language was designed for efficient modeling of large, complex and heterogeneous physical systems (Modelica Association). Model is usually decomposed into several hierarchical levels. On the bottom of the hierarchy there are submodels of basic physical phenomena which are most commonly stated as a set of (acausal) differential-algebraic equations. It is thus most conveniently that these equations can be entered directly (e.g., without a need for any kind of manipulation or even transformation to some other description formalism). On higher hierarchical levels, the model is described graphically by schematics (i.e., object diagrams) and the obtained scheme efficiently reflects the topology of the system. Such model representation in Modelica is thus understandable also to domain specialists who do not have a profound knowledge about computer simulation.

4.1.

Choice of ranking metric

Although different domains are modeled with rather different schemes and connections, acausal connections for modeling physical interactions are of special interest. Each (dynamic) interaction between physical systems results in an energy exchange between the systems. So it is very intuitive to chose the energy-based metrics for the simplification of a physical systems models. 575

Metric we chose Louca (1998) is defined by Eq. 1. It is the integral of absolute net energy flow that element exchanges with environment in time interval [t1 ,t2 ]. Activity =

 t2 t1

| ∑ pi (t)| · dt

account also a suitable selection of a frame of reference (forces, torques and orientation are expressed in local, while position is in global frame of reference). A definition of the connector is the following:

(1)

i

c o n n e c t o r Frame SI . P o s i t i o n r_0 [ 3 ] ; Frames . O r i e n t a t i o n R ; flow SI . Force f [ 3 ] ; flow SI . Torque t [ 3 ] ; end Frame ;

In Eq. 1, pi (t) designates an energy flow through the boundary of the element (in Modelica usually modeled with connection). 4.2.

Determination of energy-flows in object diagrams

The position is determined with the variable r_0, while the orientation R is a structure containing the transformation matrix T from global to local frame of the reference and the vector of angular velocities ω in the local frame of reference. Forces and torques are given by vectors f and t respectively. The power of the connection can be calculated by the expression: p = dtd (T·ro )· f +ω ·t, where again, there is a need to differentiate the position after transformation to local frame.

Modelica object diagrams, when modeling physical systems, share many similarities with bond graphs, which can be efficiently used for object-oriented acausal modeling. Therefore it is possible to adapt most of bond-graph simplification techniques to Modelica’s object diagrams. Of course the energy concept in bond graphs is much more unified in comparison with different Modelica libraries. So we analyzed the energy interactions between components in different Modelica libraries at the beginning. Connectors usually contain a pair of effort and flow variables. However, their product is not necessarily an energy flow like in bond graph formalisms. This can be seen by inspecting Modelica Standard Library (Mod, 2010) where elementary connector definitions for almost all physical domains are gathered:

• Connector for modeling the heat transfer in 1-D consists of the effort variable temperature T and the flow variable for heat-flow rate Q f low . The energy transfer is in this case equal to flow variable p = Q f low . • Library Modelica.Fluid deals with modeling of heat and mass transfer. The connector used in library’s components which covers also mass transfer is implemented as following:

• Interaction between components in analog circuits (Modelica.electric) is determined by voltage v and current i, the latter is a flow variable, and the power of the interaction is the product of both variables: p = v · i.

connector FluidPort r e p l a c e a b l e p a c k a g e Medium = M o d e l i c a . Media . I n t e r f a c e s . P a r t i a l M e d i u m ; f l o w Medium . MassFlowRate m_flow ; Medium . A b s o l u t e P r e s s u r e p ; s t r e a m Medium . S p e c i f i c E n t h a l p y h_outflow ; s t r e a m Medium . M a s s F r a c t i o n X i _ o u t f l o w [ Medium . nXi ] ; end F l u i d P o r t ;

• Similar features has also the connector in Modelica.Magnetic, which is composed of variables for magnetic potential difference Vm and magnetic flux Φ, an effort and flow variables respectively. The power of the connection is the product of both variables: p = Vm · Φ.

Besides effort and flow variables, pressure p and mass-flow rate m f low respectively, the connector includes also additional information about properties of the substance which is being exchanged in the interaction modeled by a connection of type FluidPort: specific enthalpy h and composition of substance (vector of mass fractions Xi if substance is a mixture). The thermodynamic state of the substance is uniquely determined by the variables of connector and all the other (thermodynamic) properties can be calculated by using functions provided by package Medium which is a parameter of the connector. However, thermal diffusion is not covered by this connector (it is neglected).

• Connectors used for modeling of 1-D translational and rotational mechanics consist of position s and angle φ respectively, and force f and torque τ respectively. However the product of connector’s effort and flow variables is no longer the power. For determination of the power of the connection, displacement variable has to be differentiated: p = d d dt s · f and p = dt φ · τ for translational and rotational mechanics respectively. • In Modelica Multibody library, which deals with 3D mechanics, effort and flow variables are no longer scalars, they are 6-dimensional vectors, so a state of a free-body (having 6 degree-of-freedom) can be determined. Furthermore, due to computational restrictions, implementation of connector takes into

Energy flow associated with the connector is composed of thermal, hydraulic and chemical term and could be calculated as following (in Modelica By 576



 



 





 Figure 1: Scheme of a car suspension system. Element gravityForce_s spring_s ground mass_s damper_s spring_t gravityForce_t mass_t damper_t displacement_s displacement_t

Activity [J] 2,270.06 1,763.33 795.02 787.65 198.82 192.57 92.98 24.53 0.53 0.00 0.00

Relative [%] 37.06 28.79 12.98 12.86 3.25 3.14 1.52 0.40 0.01 0.00 0.00

Accumulated [%] 37.06 65.85 78.82 91.68 94.93 98.07 99.59 99.99 100.00 100.00 100.00

Figure 2: Car suspension system: model represented by a Modelica object diagram.

Table 1: Ranking of components when model from Fg. 2 is fed by input shown in Fg. 3.

4.4. Use of Chemo-bonds, 2009): p = m˙ · s · T + m˙ · p/ρ + ∑ μi · N˙ i . Quantities specific entropy s, temperature T , density ρ, chemical potential μi and molar flow N˙ i can be calculated from thermodynamical state equations provided by package Medium.

4.3.

Example

The model from Fg. 2 is excited by signal depicted in Fg. 3. Components of the model are ranked with activity metrics (Fg. 1) and results are shown in table 1. The second column of table 1 consists of activities of all components calculated with Eq. 1. The third column contains relative activities components (the sum equals 100%) and the last column shows the accumulated relative activities. This column very illustratively shows how many components has to be taken into account for a reasonable accuracy. The aim of the ranking tables is to simplify the model in Fg. 2 by removing components from the bottom of the tables 1. However, a high accumulated relative activity of the remaining components (e.g., components not removed) do not necessary imply high similarity of original and simplified model responses. It is necessary to validate the simplification comparing the original and simplified responses.

Ranking of elements

Although it is possible to calculate energy flow of the connector from the variables of the connector, this is not always possible to do from simulation results, because some variables can not be available. For example, the derivative of a position or an angle in the connector of the library for 1-D mechanics may not be available if this variable is not chosen as a state variable. This implies instrumentation of the model, i.e., additional auxiliary variables and equations are inserted into the model. Ranking is done as post-processing of simulation results. Activity of each element required for ranking is calculated with Eq. 1. Each hierarchical level is considered separately. After the ranking of the elements is available, model can be simplified by removing all the elements that fall below certain threshold (value of activity in our case). However, our current implementation provides only results of ranking in a printed form (a table). The ranking table can be then used to simplify the model manually. Automatic simplification is the matter of future investigations.

5.

SIMPLIFICATION OF MODEL’S EQUATIONS

From a mathematical point of view, models in Modelica are systems of hybrid differential-algebraic equations (DAE). Therefore, it some cases we might want to investigate these equations directly. In all modeling environments supporting Modelica, models can be printed in flat form (i.e., as a system of equations). However, this can yield complex expression even for relatively small models, so symbolic model reduction techniques are applicable to achieve favorable representation. 577

A set of nonlinear algebraic equations (3) is obtained for each design point by solving original equation system (2) with x˙ = 0 and given u.

30

displacement [cm]

25

F(x, x˙ , y, u) = 0

20

Values of variables at steady-state of the modified system ˜ x˙ , y, u,t) (where a single term in one of equations F(x, is changed) are estimated by performing single NewtonRalphson iteration (4) and the solution of the original system is used as a guess value.

15 10

input signal mass_t displacement mass_s displacement

5 00

1

2

time [s]

3

4

 5

εi

(4)

= || [x∗ , y∗ ]T − [x(1) , y(1) ]T || ˜ ∗ , x˙ ∗ , y∗ ) || = || J−1 (x∗ , x˙ ∗ , y∗ ) · F(x

(5)



To further reduced computational costs, inverse Jacobian matrix is computed only once for original system at each design point and inverse Jacobians of the modified system are obtained by Sherman-Morisson formula (6).

Simplification strategies

There are many mature simplification methods available for linear systems (Eitelberg, 1981; Fodor, 2002; Innis and Rexstad, 1983), while there is a lack of more general simplification methods that could be applied effectively on multi-domain models described as a set of nonlinear DAE. Most commonly employed simplification strategies for nonlinear DAE combine model order reduction techniques (i.e., deletion of variables and variables’ time derivatives or substitution of variables with constant values) and an approximation of single terms, for example, linearization, deletion or substitution of the term with constant value, etc. (Sommer et al., 2008; Wichmann et al., 1999). The order of the applied simplification steps is determined by ranking. The most apparent ranking metric is estimation of reduced-model error for selected variables (variables of interest). One possible approach is to repeat simulation after for each possible simulation step which would yield a perfect ranking. However, this method is too time consuming to have any practical meaning. Another option is to use energy-based ranking metric as in case of object-diagram simplification. In method suggested by (Chang et al., 2001) it is required that Lyapunov function of the system is known which is a rather harsh restriction. The metric (Wichmann et al., 1999) suggest for simplification of nonlinear DAE systems obtained in analog circuit design estimates the error caused by simplification step (e.g., term deletion, substitution of the term with a constant value, term linearization) and it is done in two parts: the DC analysis and AC analysis. The former requires calculation of several design points, i.e. steadystates of the system at different inputs. F(x, 0, y, u,t) = 0

  ∗ x x(1) ˜ ∗ , x˙ ∗ , y∗ ) = − J−1 (x∗ , x˙ ∗ , y∗ ) · F(x F˜ y∗ y(1)

The error estimation εi is then calculated by equation (5).

Figure 3: A car hits a smooth curb: low-frequency excitation signal is given as an input to the model in Fg.2. Also responses – displacement of an unsprung (mass_t) and sprung mass (mass_s) are depicted. 5.1.

(3)

T −1 −1 −1 T −1 = J−1 J−1 F − (1 + v JF el ) JF el v JF F˜

(6)

Finally, error estimations εi are combined in equation (7). ε = || [ε1 , . . . , εn ]T ||

(7)

The latter part of the metric, the AC analysis, requires linearization of the DAE system at selected design points and then transfer functions are computed. The resulting transfer functions are then simplified using methods for linear-systems simplification. However, the simplification method of (Wichmann et al., 1999) is limited on DAE systems representing analog electrical circuits and analogue systems. The most impractical limitation preventing its use for more general multi-domain models in Modelica is that it requires calculation of design points, i.e. solving usually large nonlinear system of equations. Could be this method extended to handle simplify transients of nonlinear DAE system directly (i.e., without linearization at selected design points)? Influence of a simplification on the equations’ transient solution could be predicted by performing single Newton-Ralphson iteration of equation system (2) at different time instants of the transients and then combine the obtained error estimates as suggested by the (Wichmann, 2003). However, this kind of error estimation is much more difficult as in case of purely algebraic nonlinear system (system at steady-state), because only local integration error is estimated and the elimination of low-ranked terms often results in an unstable system. (Wichmann, 2003) does not report how this problem was solved.

(2) 578

6.

CONCLUSION

G. Innis and E. Rexstad. Simulation model simplification techniques. Simulation, 41(1):7–15, 1983.

As the complexity of the models continuously increases, it is important than contemporary modeling environments include tools which help understand and cope with these models effectively. One class of those tools are also model simplification techniques. Models which can be built in Modelica can be very heterogeneous and include submodels from different physical domains. On contrary, most model simplification techniques require strict modeling formalism and are limited on certain physical domains. They are thus not easily applicable on most models in Modelica. As it was shown in the paper, simplification methods developed for bond-graphs can be easily adapted for simplification of Modelica’s object-diagrams. However, the simplification of the model on equation level is much more difficult and there are no publicly published simplification methods known to the author which could be efficiently used for a simplification of wide variety of Modelica models.

Sanjay Lall, Petr Krysl, et al. Structure-preserving model reduction for mechanical systems. Physica D, 284: 304–318, 2003. Loucas Sotiri Louca. An Energy-based Model Reduction Methodology for Automated Modeling. PhD thesis, University of Michigan, 1998. Modelica Standard Library 3.1, User’s Guide. Modelica Association, 2010. https://www.modelica. org/libraries/Modelica. Modelica Association. Modelica Specification, version 3.2, 2010. http://www.modelica.org/ documents/ModelicaSpec32.pdf. J. D. Murray-Smith. Simulation model quality issues in product engineering: A review. Simulation News Europe, 19(2):47–57, 2009. P. Schwarz et al. A tool-box approach to computer-aided generation of reduced-order models. In Proceedings EUROSIM 2007, Ljubljana, Slovenia, 2007.

REFERENCES M. Andres, D. Zimmer, and F. E. Cellier. Object-oriented decomposition of tire characteristics based on semiempirical models. In Proceedings of the 7th Modelica Conference, pages 9–18, Como, Italy, 2009.

A. Sodja and B. Zupanˇciˇc. Model verification and debugging of eoo models aided by model reduction techniques. In Proceedings of the EOOLT 2010, pages 117–120, Oslo, Norway, 2010.

T. Blochwitz et al. The functional mockup interface for tool independent exchange of simulation models. In Proceedings of 8th International Modelica Conference, Dresden, Germany, 2011.

Ralf Sommer, Thomas Halfmann, and Jochen Broz. Automated behavioral modeling and analytical modelorder reduction by application of symbolic circuit analysis for multi-physical systems. Simulation Modelling Practice and Theory, 16:1024–1039, 2008.

F. Casella et al. The modelica fluid and media library for modeling of incompressible and compressible thermofluid pipe networks. In Proceedings of the 5th Modelica Conference, pages 631–640, Vienna, Austria, 2006.

Hubertus Tummescheit. Design and Implementation of Object-Oriented Model Libraries using Modelica. PhD thesis, Lund Institute of Technology, 2002.

F. E. Cellier. Continous System Modeling. Springer Verlag, New York, 1991.

T. Wichmann. Transient ranking methods for the simplfication of nonlinear dae systems in analog circuit design. Proceedings in Applied Mathematics and Mechanics, 2:448–449, 2003.

Samuel Y. Chang, Christopher R. Carlson Carlson, and J. Christian Gerdes. A lyapunov function approach to energy based model reduction. In Proceedings of the ASME Dynamic Systems and Control Division – 2001 IMECE, pages 363–370, New York, USA, 2001.

T. Wichmann et al. On the simplification of nonlinear dae systems in analog circuit design. In Proceedings of CASC’99, Munich, Germany, 1999.

E. Eitelberg. Model reduction by minimizing the weighted equation error. International Journal of Control, 34(6):1113–1123, 1981. I. K. Fodor. A survey of dimension reduction techniques. Technical report, Lawrence Livermore National Laboratory, 2002. Technical Report UCRL-ID-148494. Modeling Chemical Reactions in Modelica By Use of Chemo-bonds. Cellier, f. e. and greifeneder, j. In Proceedings of the 7th Modelica Conference, pages 142– 150, Como, Italy, 2009. 579

THE EXCLUSIVE ENTITIES IN THE FORMALIZATION OF A DECISION PROBLEM BASED ON A DISCRETE EVENT SYSTEM BY MEANS OF PETRI NETS Juan Ignacio Latorre-Biel(a), Emilio Jiménez-Macías(b) (a)

Public University of Navarre. Deptartment of Mechanical Engineering, Energetics and Materials. Campus of Tudela, Spain (b) University of La Rioja. Industrial Engineering Technical School. Department of Electrical Engineering. Logroño, Spain (a)

[email protected], (b)[email protected]

ABSTRACT

money to companies and users. The design of discrete event systems is likely to influence in a decisive way the performance in their operation. For this reason to define adequately the design process of a discrete event system to be manufactured or constructed is a very productive activity with clear consequences in the whole life of the system (Balbo and Silva 1998). On the other hand, to consider the performance of the operation of the system is an adequate approach to afford a design process that aims to achieve a desired behaviour for the system once it is in a running stage. Nevertheless, to forecast the future behaviour of a system in process of being designed and that is not a reality yet presents several handicaps.

The design of discrete event systems (DES) can be seen as a sequence of decisions, which allows obtaining a final product that comply with a set of specifications and operates with efficiency. A decision support system can alleviate the decision making as well as provide with more information and tools to make the best choice to the decision-maker. The decisions related to the design of a DES may include the choice among a set of alternative structural configurations. These alternatives may be defined by the designer by mere combinations of subsystems that solve subproblems associated to the specifications and behaviour of the DES. As a consequence, it is possible that the alternative configurations share redundant information that lead to improvements in the classical approaches to solve this type of decision problems. In this paper, the formalization of a decision problem based on a DES, underlining the characteristic feature of exclusivity between alternative configurations is presented as a tool that broadens the classical approach with new ideas and techniques to improve the efficiency in the solving of decision problems.

On the first hand, it is necessary to approximate the results, since a model of the system and not the real one should be used. On the other hand, it is needed to select the type of model to be used. In certain DES it is possible to develop physical prototypes to test certain properties of the behaviour of the systems. More commonly, formal models are developed to apply algorithmic methodologies to analyze the behaviour of the original system. Sometimes it is possible and productive to combine the construction of several models of different nature in the design process of a system: physical systems that model specific characteristics of the real system can be combined with formal models developed on a computer to forecast the behaviour of the real system.

Keywords: decision support system, discrete event systems, Petri nets, exclusive entity.

1. INTRODUCTION Many technological systems can be described as discrete event systems (DES) due, in a large number of cases, to the presence of digital computers in their control. Mobile phones, computer networks, manufacturing facilities or logistic systems constitute examples of DES whose presence is common in our technological society (Cassandras and Lafortune 2008). The efficiency and correctness in the operation of these systems can save important quantities of

The formal models developed to forecast the performance of a discrete event system in process of being designed are usually complemented with simulation in order to evaluate the behaviour of the model (Piera et al. 2004). In fact, simulation allows exploring the region of the state space of the system under specific conditions. On the other hand, it is

580

common that modelling and simulation present a rate between the information obtained from the system and cost associated to the modelling process, which is more favourable than other techniques that imply the physical implementation of the model by means of prototyping.

The model of the system is only a part of the formalization process of a decision problem stated on a discrete event system. There are other elements that can be included in the resulting formal problem as the type of solution expected for the problem, the solution space and, depending on the decision problem, the objective function that evaluates the cost or the performance of the DES after the selection of a certain solution from the solution space.

Furthermore, the design process of a DES requires stating and solving several decision problems. In particular, it is usual that it is necessary to choose among a set of alternative configurations or structures for the system (Latorre et al. 2009). This is the case, for example, when different layouts for the material or products conveying can be chosen to define a final configuration for a chain supply in process of being designed. A classical approach to this type of problems is intensive in computer resources, when the solution is searched by means of formal algorithms. This fact is due to the analysis by simulation of every one of the alternative configurations, knowing that this analysis requires launching not only one but sometimes an undetermined number of simulations.

In this paper, an overview on the statement and formalization process of a decision problem based on a discrete event system is given underlying the exclusiveness feature in the different alternative structural configurations for the DES that can be particularized by different formalisms. On the other hand, this exclusiveness can be abstracted into the concept of set of exclusive entities that leads to an interesting property that will be defined in this paper. 2.

DEFINITIONS

A discrete event system can be defined in the following way:

An analysis of the design process described in the previous paragraph, characterizes the discrete event system to be designed among a set of alternative structures with the property of mutual exclusive evolution between the structures. This idea allows developing methodologies that reduces the computational cost of performing simulations to the different alternative configurations by the removal of the redundant information present in the models of every one of them (Latorre et al. 2010b).

Definition 1. Discrete event system. Dynamic system whose behaviour can be described by discrete state variables and is governed by asynchronous and instantaneous incidences, called events, which are solely responsible for the state changes.  A discrete event system may be defined in a more or less ambiguous way by a set of specifications and some constraints and expectations in its dynamic behaviour. The ambiguity in the definition of a DES can be interpreted as freedom degrees, some of which should be particularized in the design process, while others can be specified in the operation processes. The mentioned freedom degrees may be called undefined characteristics of the discrete event system.

This concept of exclusive entity is abstracted in a more general idea defined by the exclusive entities, associated to an undefined model. On the other hand, this new idea can be particularized in a variety of formalisms to be able to represent the model of the system in a compact way able to develop fast sets of simulations to support the decision making process in the design of discrete event systems. The general approach given by a set of exclusive entities to the exclusiveness associated to the different alternative structural configurations for a discrete event system to be designed, constitutes a characteristic feature of a model defined as a disjunctive constraint in the formalization of a decision problem based on a DES. A model of this kind can be called undefined Petri net since it contains certain parameters whose values should be chosen among a domain set as result of decisions.

The previous paragraph allows to define a particular type of DES. Definition 2. Undefined DES. A discrete event system with at least one undefined characteristic is said to be an undefined DES.  The type of decisions stated in the design process of an undefined DES try to reduce the

581

ambiguity in the description of the discrete event system, by the transformation of undefined characteristics in defined ones. As a consequence, it is possible to state a decision problem in the way described in the following.

P = {p1, p2, . . . , pn} is the finite, non-empty, set of places (one type of node in the graph). T = {t1, t2, . . . , tm} is the finite, non-empty, set of transitions (the other type of node in the graph) F Ž (P × T) ‰ (T × P) is the set of directed arcs (from places to transitions and from transitions to places) in the graph, called flow relation. w : F ĺ 1* is the weight function on the arcs.

Definition 3. Decision problem based on a DES. Let D be an undefined discrete event system. A decision problem based on D is a choice, among several alternatives, in response to a question posed on any set of the undefined characteristics of D or its evolution. 

In this paper, a new approach in the definition of a Petri net will be defined.

Once a decision problem has been stated, it is necessary to solve it. For this purpose, it is convenient to represent it in a formal language.

Definition 5. Unmarked Petri net. A (generalized) unmarked Petri net (or Petri net structure) is a triple

A formal language shows important advantages from a natural language to state a decision problem. On the one hand, it provides with precision to the description of the problem, removing ambiguity and allowing the application of an algorithmic solving methodology. On the other hand, the consequence of the successful application of a solving methodology to a decision problem expressed in a formal language is one or several quantitative results, which can easily be compared with numerical references or the results of other methodologies.

N = ¢np, SJ, SvalJ² where np  1* is the number of places. SJ = { J1, J2, … , Jn } is a set of structural parameters. SvalJ = { cv1, cv2, … , cvm } is the set of feasible combinations of values for the parameters of SJ. It is verified that n = k · np, where k  1* and  cvi  SvalJ , cvi = (v1, v2, … , vn).

A first element that is convenient to include in the formal statement of the decision problem is the discrete event system itself. There are a number of formal languages that can cope with the modelling of a generic discrete event system. However, the decision of the formal language to be considered in this paper, the Petri nets (Petri 1962), is based in the versatility and double representation that may be matrix-based or graphical. On the other hand complex behaviours of collaboration and competence may be modelled in an easy and natural way (David and Alla 2005), (Jiménez et al. 2005).

This new definition of unmarked Petri net allows constructing the incidence matrices of the formalism and underlines the approach of this paper, focussed on the formalization process of a decision problem based on a discrete event system. Reducing the concept of Petri net to a collection of parameters and their feasible values, the formalization process from a discrete event system can be considered as the translation of a subset of characteristics of the DES into a set of parameters of the Petri net. The characteristics of the DES translated and included in the Petri net will depend on the degree of detail of the model.

In particular, it is possible to define an autonomous unmarked Petri net in the following way (Cassandras and Lafortune 2008) and (Silva 1993), where an introduction to the Petri net paradigm can be found as well as in (Peterson 1981) or (David and Alla 2005).

Subsequently, an undefined characteristic of the discrete event system will be modelled by means of one or several undefined parameters in the Petri net.

Definition 4. Petri net graph A (generalized) Petri net graph (or Petri net structure) is a weighted bipartite graph

The process of obtaining a formal model from an original system, the modelling process, can be interpreted by means of the translation of the undefined characteristics of the DES into a set of undefined parameters. As a consequence, an

N = ¢P, T, F, w² where

582

undefined parameter can be defined as indicated below.

system is associated to so many different Petri nets as alternative structural configurations for the DES can be found. These Petri nets can be called alternative Petri nets and belonging to the same model for the original DES should comply with a property of exclusiveness (Latorre et al. 2011). Of course it is not possible that several of them can be chosen as solution for the DES design process. The only option for the model to be coherent with the reality of the decision problem is to comply with a property of exclusiveness. This property can be imposed by means of the concept of mutually exclusive evolution defined below.

Definition 6. Undefined parameter Any numerical variable of a Petri net model or its evolution that has not a known value but it has to be assigned as a consequence of a decision from a set of at least two different feasible values. The value assigned to the undefined parameter must be unique.  A parameter of a Petri net may belong to the set of structural parameters; nevertheless, it is possible to define other types of parameters according to the role they play in the model. For example, there is a category of marking parameters that includes the initial marking of all the places of the Petri net. As a consequence, an autonomous marked Petri net can be defined in the following way.

Definition 8. Mutually exclusive evolution Given two Petri nets R and R’. They are said to have mutually exclusive evolutions if it is verified: i) If m(R) z m0(R) Ÿ m(R’) = m0(R’) ii) If m(R’) z m0(R’) Ÿ m(R) = m0(R)   As a consequence, a set of alternative Petri nets can be described as:

Definition 7. Marked Petri net. A (generalized) marked Petri net (or Petri net system) is a triple N = ¢np, SJ, SvalJ²

Definition 9. Set of alternative Petri nets. Given a set of Petri nets SR = { R1, …, Rn }, SR is said to be a set of alternative Petri nets if n>1 and  i, j such that 1 d i, j d n, Ri and Rj verify:

where np  1* is the number of places. SJ = { J1, J2, … , Jn } is a set of structural parameters. SvalJ = { cv1, cv2, … , cvm } is the set of feasible combinations of values for the parameters of SJ.

i) R and R’ have mutually exclusive evolution. ii) W(R) z W(R’). Ri is called the i-th alternative Petri net of SR.

It is verified that n = (k+1) · np, where k  1*and  cvi  SvalJ , cvi = (v1, v2, … , vn).

This classical approach of modelling an undefined DES with alternative structural configurations by means of a set of alternative Petri nets is not the only option. Even it is not necessarily the most efficient option (Latorre et al. 2009) for posing and solving a formal statement of the decision problem based on the DES.

It is possible to notice from the comparison of the definition 5 and the definition 7 that the addition of the np marking parameters have modified the definition of the unmarked Petri net by increasing the size of the set SJ and hence the number of values in the feasible combinations of values belonging to SJ.

In this search for new formalisms, it is interesting to abstract the representation of the undefined DES performed with the set of alternative Petri nets. On the first hand, it is possible to obtain a general abstraction for the mutually exclusive evolution of the alternative Petri nets by means of the concept of the exclusive entities. The alternative Petri nets can be considered exclusive entities since only one of them can be chosen at a time. In fact, a set of exclusive entities can be defined in the following way:

Furthermore, it is easy to deduce that this parametric definition of a Petri net allows easily to be extended to Petri nets with extended features as interpreted Petri nets, including timed Petri nets, and other nets that include priorities, colours, etc. As it has already been explained previously, the design process of a discrete event system is usually associated to several alternative structural configurations for the DES. A classical approach for the modelling of such a

583

Proposition 2. A set of choice variables is a set of exclusive entities.

Definition 10. Set of exclusive entities. Given a discrete event system, a set of exclusive entities associated to it is a set Sx = { X1, …, Xn }, which verifies that

Proposition 3. A natural choice colour is a set of exclusive entities.

i) The elements of Sx are exclusive, that is to say, only one of them can be chosen as a consequence of a decision.

Proposition 4. A set of alternative Petri nets is a set of exclusive entities.

ii)  i, j  1*, 1 d i, j d n it is verified that Xi z Xj. iii)  f: Sx o SR , where

All these valid representations of a set of exclusive entities lead to different formalisms able to model a discrete event system with alternative structural configurations. For more details on the elements that appear in the statements of the propositions see (Latorre et al. 2010a) and (Latorre et al. 2010c).

SR = { R1, …, Rn } is a set of alternative Petri nets, feasible models of D. f is a bijection Ÿ  Xi  Sx ! f(Xi) = Ri  SR such that Ri is a feasible model for D and  Ri  SR ! f-1(Ri) = Xi  Sx .

An additional property should be guaranteed for any representation of a set of exclusive entities.

Definition 11. Undefined Petri net. An undefined Petri net is a 4-tuple

where

Theorem. Given an undefined Petri net RU associated to a set of exclusive entities Sx, any representation of the set of exclusive entities Sz verifies that card (Sx) =card (Sz)

np  1* is the number of places. SJ = { J1, J2, … , Jn } is a set of structural parameters. SvalJ = { cv1, cv2, … , cvm } is the set of feasible combinations of values for the parameters of SJ. Sx = { X1, X2, … , Xq }, where q > 1, is a set of exclusive entities.

This last property implies that no matter which representation is chosen for the set of exclusive entities of an undefined Petri net, the cardinality of its representation is the same than its abstraction Sx. In other words, the number of alternative structural configurations of the original discrete event system is constant.

N = ¢np, SJ, SvalJ, Sx²

It is verified that n = (k+1) · np, where k  1*and  cvi  SvalJ , cvi = (v1, v2, … , vn).

The applications of the concept of set of exclusive entities can be found in the decision field that is associated to the discrete event systems.

In fact, the set of exclusive entities Sx does not provide with more structural or marking parameters to the model. It simply organizes or classifies the parameters into exclusive subsets. The specific representation of this undefined Petri net can be made according to different formalisms that should include a set of exclusive entities.

Each exclusive entity can be related to a feasible configuration of the original discrete event system. The set of all these configurations determine the complete set of possible choices to define univocally the controllable parameters of the associated Petri net model. One interesting application of the concept of set of exclusive entities consists of restricting the association to the structural configurations of the original DES to the exclusive entities. In this case it is possible to develop a methodology to choose among different structures to design, modify or control certain systems modelled by Petri nets. Every exclusive entity may be associated to several undefined or controllable parameters that lead to diverse behaviours,

3. PROPERTIES AND APPLICATIONS Several properties can be stated in relation with the idea of an undefined Petri net: Proposition 1. The feasible combinations of values for the undefined structural parameters of a compound Petri net is a set of exclusive entities.

584

however, the exclusive entities are reserved, in this approach, to the structural parameters. In this methodology, it is possible to state the following theorem:

REFERENCES Balbo, G. y Silva, M. (eds.), “Performance Models for Discrete Event Systems with Synchronizations: Formalisms and Analysis Techniques”, Editorial Kronos, Zaragoza, Spain, 1998

The search for an efficient representation of an undefined Petri net, can lead to formalisms that profit from the search methodology, from the similarities between the different structural configurations of the DES, from a single solution space, etc, enhancing the performance of the optimization algorithms aimed to take the best decisions.

Cassandras, Christos G., Lafortune, S., “Introduction to Discrete Event Systems”. Second Edition, Springer, 2008 David, R., Alla. H., Discrete, Continuous and Hybrid Petri nets, Springer, 2005

Some formalisms that can be mentioned are the sets of alternative Petri nets, the compound Petri nets, the alternatives aggregation Petri nets and the disjunctive coloured Petri nets. Their suitability for representing an undefined Petri net can be deduced from the propositions 1 to 4. 4.

CONCLUSIONS RESEARCH

AND

Jiménez, E., Pérez, M., Latorre, J.I., “On deterministic modelling and simulation of manufacturing systems with Petri nets”. Proceedings of European Modelling Simulation Symposium. Marseille, pp. 129-136. Marseille. 2005

FUTURE Latorre, J.I., Jiménez, E., Pérez, M., Martínez, E., “The design of a manufacturing facility. An efficient approach based on alternatives aggregation Petri,” Proceedings of the 21st European Modelling and Simulation Symposium (EMSS 09). Puerto de la Cruz, Spain, vol. 2, pp. 33-39, September 2009.

In this paper, a decision problem based on a discrete event system is analysed. Moreover, some important topics in the formalization process of this problem are considered. In particular, it has been underlined a relevant type of decision problem stated in the design process of a discrete event system. The new approach of considering a Petri net from a parametric point of view leads to an abstraction of a model of a discrete event system with alternative structural parameters. Some properties allow relating the set of exclusive entities with different representations that comply with the invariance of their number of elements. This property is related with the fact that the number of alternative structural configurations for the DES being designed is constant.

Latorre, J.I., Jiménez, E., Pérez, M., “The alternatives aggregation Petri nets as a formalism to design discrete event systems.” International Journal of Simulation and Process Modeling, Special Issue. 2010 Latorre, J.I., Jiménez, E., Pérez, M., “On the Solution of Optimization Problems with Disjunctive Constraints Based on Petri Nets” Proceedings of the 22nd European Modelling and Simulation Symposium (EMSS 10). Fez, Moroco, pp. 259-264, October 2010.

The topic presented and summarized in this paper is an important part in the theory that affords the solution of the decision problems based on DES with different alternative structural configurations by means of the removal of redundant information and obtaining compact formalisms that behave efficiently in the algorithms to solve the associated problems.

Latorre, J.I., Jiménez, E., Pérez, M., “Coloured Petri Nets as a Formalism to Represent Alternative Models for a Discrete Event System”. Proceedings of the 22nd European Modelling and Simulation Symposium (EMSS 10). Fez, Moroco, pp. 247-252, October 2010.

Open research lines so far are the search for new formalisms to represent undefined Petri nets, as well as to develop criteria and algorithms to choose the best formalism to solve a given decision problem.

Latorre, J.I., Jiménez, E., Pérez, M., “Efficient Representations Of Alternative Models Of Discrete Event Systems Based On Petri Nets”. Proceedings of the UKSim 13th International Conference on Computer

585

Modelling and Simulation. Cambridge, United Kingdom, March 2011. Peterson, J.L. “Petri Net Theory and the Modelling of Systems”, Prentice Hall, Englewood Cliffs, 1981. Petri, Carl A. (1962). “Kommunikation mit Automaten”. Ph. D. Thesis. University of Bonn (German). Piera, M.À., Narciso, M., Guasch, A., Riera, D., “Optimization of logistic and manufacturing system through simulation: A colored Petri net-based methodology,” Simulation, vol. 80, number 3, pp 121-129, May 2004 Silva, M. “Introducing Petri nets”, In Practice of Petri Nets in Manufacturing”, Di Cesare, F., (eds.), pp. 1-62. Ed. Chapman&Hall. 1993

586

MATRIX-BASED OPERATIONS AND EQUIVALENTE CLASSES IN ALTERNATIVE PETRI NETS. Juan Ignacio Latorre-Biel(a), Emilio Jiménez-Macías(b) (a)

Public University of Navarre. Deptartment of Mechanical Engineering, Energetics and Materials. Campus of Tudela, Spain (b) University of La Rioja. Industrial Engineering Technical School. Department of Electrical Engineering. Logroño, Spain (a)

[email protected], (b)[email protected]

ABSTRACT Petri net is a modelling paradigm used for discrete event systems (David and Alla 2005), (Cassandras et al. 2008). Their transformations are a powerful tool for the validation and verification of Petri nets as models of discrete event systems. If a simplified Petri net verifies a certain set of properties, the original Petri net will verify them if the applied net transformation preserves these properties. On the other hand, the control of Petri nets may also improve with the application of Petri net transformations. If the unobservable transitions are removed by the application of transformation rules, then the state of the simplified net evolves under the firing of the observable transitions. In this paper, the transformation of nets will be applied to modify the formalism that represents a disjunctive constraint of a decision problem stated on a Petri net model. Some formalisms may be more suitable for the modelling process than others (alternative Petri nets), while others are more compact and suitable for the development of optimization processes in an efficient way (compound Petri nets, alternatives aggregation Petri nets or disjunctive coloured Petri nets). A set of transformation rules that preserve the structure of the associated graph or reachable markings are provided in this paper, as well as an example of application in a net transformation between two formalisms. As a consequence of the application of these rules, both the original and resulting formalisms will show an equivalent behaviour. Hence, any of them can be used to state a decision problem but the efficiency of the algorithm to solve the problem may be different when considering the required computer resources and the quality of the obtained solution.

Keywords: equivalence operations, Petri net transformations, decision support system, compound Petri nets, alternative Petri nets. 1. INTRODUCTION One of the stages that can be considered in the modeling process of a discrete event system is the validation and verification (Peterson 1981), (Jimenez et al. 2005). According to (Silva, 1993), it is possible to reduce the cost and the duration of the design process of a SED by checking if certain properties are verified by the model. One of the techniques of qualitative analysis is based in the transformation of the Petri net structure. Some important issues of these techniques are described in (Berthelot 1987), (Silva 1993) and (Haddad and PradatPeyre 2006). These early developments in the theory of the Petri nets have led to other transformation techniques in the static structure and to new formalisms based on PN that are aimed to simplify the modeling of DES whose structure varies with time. For example (Van der Aalst 1997) provides with eight transformation rules, based in the previously mentioned techniques, applied to systems that experience the frequent changes in their structure. An undefined Petri net can be interpreted as a model of a discrete event system that includes freedom degrees in its structural characteristics (Latorre et al., 2009b). The undefined Petri net is an abstraction that can be particularized in a specific formalism. A classical approach to obtain a model of an undefined discrete event system is a set of alternative Petri nets (Latorre et al., 2007). This type of nets verifies the property of mutually exclusive evolutions, hence in the same Petri net alternative structural configurations of an original discrete event system can be included (Latorre et al. 2011).

587

In this paper the concept of equivalence class is applied to every alternative Petri net. This concept allows substituting any Petri net belonging to a set of alternative Petri nets by another one whose behaviour and properties are the same than the original one. Hence, the resulting set of alternative Petri nets will verify the same properties and show equivalent behaviour. Every equivalence class will be composed by Petri nets with the same behaviour and all of them will be said to be equivalent.

The operation of swapping two rows of a matrix is defined as the following function: swapr: Mmunu{1,2, …,m}u{1,2, …,m} o Mmun (A, i, j) o B  Mmun

§ a11 ¨ ¨ ¨a i1 where, A = ¨ ¨ ¨a ¨ j1 ¨ ¨ © am1

Given an alternative Petri net, the methodology to obtain equivalent Petri nets to the first one will be based in matrix-based operations, applied to the incidence matrix of the net. These matrix-based operations will lead to new Petri nets but the graphs of reachable markings will be isomorphous in the original and the resulting net. This fact ensures that the properties and the behaviour of both Petri nets is the same and, hence, that they can be considered as equivalent ones.

§ a11 ¨ ¨ ¨a j1 B= ¨ ¨ ¨a ¨ i1 ¨ ¨ © am1

The matrix based operations can be applied with the aim of transforming the set of alternative Petri nets into a compound Petri net (Latorre et al. 2010). This process requires the merging of the sets of parameters of all the nets belonging to the set of alternative Petri nets. In particular, it is necessary to merge the structural parameters that are the elements of the incidence matrices. In order to obtain a compound Petri net with the smallest set of undefined structural parameters and hence to obtain a compact model that requires a reduced computation resources to simulate the evolution of the original DES, it is convenient to apply the matrix-based operations to the set of alternative Petri nets.

 ai 2  a j2  am 2

a12  a j2  ai 2 

am 2

 a1n · ¸  ¸  ain ¸ ¸  Mmun,  ¸  a jn ¸¸  ¸ ¸  amn ¹

 a1n · ¸  ¸  a jn ¸ ¸  Mmun.  ¸  ain ¸¸  ¸ ¸  amn ¹ 

In other words, definition 1, describes the swapping of the ith and jth rows in a matrix A. This operation is denoted by swapr(A, i, j) Remark 1. When applying this operation to the incidence matrix of a Petri net it has to be taken into consideration that the ith row represents the input and output arcs of a place of the PN. Let us call this place pi. For this reason, swapping two rows, the ith and the jth, implies that the arcs associated to pi are no longer present in the ith row of the incidence matrix but in the jth. As a natural consequence, if this new incidence matrix is to be included in the characteristic equation of the Petri net it has to be considered that the ith element of m, the marking of the Petri net, does not represent m(pi) the marking of the place pi anymore. The same considerations can be made for pj, the jth row of the incidence matrix and m(pj). Therefore, the swapr(A,i,j) operation implies the swapping of the ith and jth elements in the marking vector m of the Petri net. This statement is also true for the particular case of m0. It is clear then that it is necessary to apply the same swapping operation that is applied to the incidence matrix, to the marking of the Petri net. In a subsequent section, it will be mentioned the reference name and the alias of any place of a Petri net. With these concepts it will be generalized the previous considerations on the swapping of

These operations might lead to alternative Petri nets whose incidence matrices have more similarities in the same positions (common elements). This fact imply that when the element of all the incidence matrices in a given position is the same, there is not an associated undefined parameter and when most of the elements are the same in a certain position, then the set of feasible values for the undefined structural parameter that appears is reduced. 2.

a12

TRANSFORMATION OPERATIONS

Once the objectives of the transformations are clear, the matrix-based operations will be presented and some examples will be given. Definition 1. Operation of swapping two rows of a matrix.

588



rows of an incidence matrix and the application of the same operation in the marking of the associated Petri net.

The operation described in the definition 3 is denoted by addr(A) and adds a row of zeros to the matrix A.

Definition 2 . Operation of swapping two columns of a matrix. The operation of swapping two columns of a matrix is defined as the following function:

Remark 3. The operation addr applied to the incidence matrix of a Petri net implies the addition of a new place with a particular property: every input and output arc has weight zero. In other words, this new place is an isolated node of the Petri net.

swapc: Mmunu{1,2, …, n}u{1,2, …, n} o Mmun (A, i, j) o B  Mmun where, § a11 ¨ A = ¨ a21 ¨ ¨ ¨a © m1

 a1i  a2 i    ami

 a1 j  a2 j    amj

 a1n · ¸  a2 n ¸  Mmun,  ¸ ¸  amn ¸¹

The marking of the Petri net that results from the application of this operation should include the marking of the new place, which will occupy the last position of the vector. However, being isolated, the place cannot experience any variation of its initial marking in the evolution of the Petri net. Furthermore, the marking of other places will not be influenced by the added place, hence the marking of the new Petri net, excluding the added place, will be the same to the original one. If the new place is considered in this comparison it is possible to say that the significant marking (the one that varies in at least an evolution of the PN) is the same in both Petri nets, hence the graphs of reachable markings are isomorphous.

and § a11  a1 j ¨

 a1i

 a1n · ¸

B = ¨ a21  a2 j  a2i  a2 n ¸  Mmun ¨      ¸ ¸ ¨ ¸ ¨a © m1  amj  ami  amn ¹



In other words, definition 2, describes the swapping of the columns i and j in matrix A, which is denoted by swapc(A, i, j)

Definition 4. Operation of removing a row of zeros of a matrix. The operation of removing a row of zeros of a matrix is defined as the following function:

Remark 2. The state equation of a Petri net requires representing the characteristic vector that summarizes the information contained in the sequence of transitions fired. The characteristic vector (also called firing count vector) contains elements that are different to zero in the positions that correspond to the transitions fired. If an operation swapc is applied to an incidence matrix and the state equation is represented, the characteristic vector should be modified according to this same swapc operation.

removr: S o M(m-1)un A o B, such that S = {A  Mmun | am* = (0 0 … 0) }, in other words, S is the set of matrices whose mth (last) row is a row of zeros.

Definition 3. Operation of adding a row of zeros to a matrix. The operation of adding a row of zeros to a matrix is defined as the following function:

§ a1,1 ¨  Given A = ¨ ¨a ¨ m 1,1 ¨ 0 ©

addr: Mmun o M(m+1)un A o B, such that

removr(A) = B

a1, n · ¸  ¸ M Ÿ mun  am 1, n ¸ ¸  0 ¸¹  

§ a1,1  a1, n · ¸ ¨ B= ¨    ¸  M (m-1)un ¸ ¨a © m 1,1  am 1, n ¹

§ a11  a1n · Given A = ¨    ¸  Mmun Ÿ ¨ ¸ ¨a ¸  a m 1 mn © ¹ § a11  a1n · ¨ ¸ addr(A) = B = ¨    ¸  M(m+1)un ¨a  amn ¸ ¨ m1 ¸ ¨ 0  0 ¸ © ¹

 The operation described in the definition 4 is denoted by removr(A) and removes the last row of a matrix A, which should contain only zeros.

589

to the alias p2 and so on. The same may happen with the transitions: the first column will always be associated to t1, the second with t2 and so on.

Remark 4. The operation removr applied to the incidence matrix of a Petri net implies the removal of a place with a particular property: every input and output arc has weight zero. In other words, this new place is an isolated node of the Petri net. Moreover, the place should be associated to the last row of the incidence matrix (if this last condition is not verified it is always possible to apply an operation swapr to guarantee this fact).

3.

APPLICATION OF TRANSFORMATIONS

THE

Example 1. Let A  Mmun be the incidence matrix of a Petri net R. The names of places and transitions of the Petri net can be shown in the following representation of A:

The marking of the Petri net that results from the application of this operation should not include the marking of the removed place (which occupied the last position of the vector before the operation). However, being isolated, the place could not experience any variation of its initial marking in the evolution of the Petri net. Furthermore, the marking of other places will not be influenced by the removed place, hence the marking of the new Petri net, will be the same to the original one (excluding the added place). If the removed place is included in this comparison it is possible to say, as it was mentioned in the remark 4, that the reachable significant markings are the same in both Petri nets, hence the graphs of reachable markings (including the non-significant markings) are isomorphous.

t1r



a11

 a1u

 a1v

 a1n

p1r

 A = ai1

 aiu

 aiv

 ain

 pir

  a jn

 p rj

    am1  amu  amv  amn

 pmr



tur

  a j1  a ju





t vr

  a jv





t nr

Let us now apply two operations to the incidence matrix A: B = swapr(A, i, j); C = swapc(B, u, v)

The swapping of rows and columns of the incidence matrices simply locates in a different place of the matrices the information (weights) related to the arcs that link a certain place with the transitions of the PN or a certain transition with the places of the Petri net. Furthermore, if the parameters associated to a place or a transition that changes its position in the incidence matrices do not remain attached to the position in the matrix but move with the place or transition, the behaviour of the net, and its structure, will be the same. For this reason, to apply such a transformation as the swapping of rows and columns of the incidence matrices, it is necessary to ensure that the appropriate amount of parameters are associated to the moving places and transitions. In order to facilitate this operation it is convenient to define a reference name for every place and transition, to which its parameters will be also referred. This reference name will be attached to the information (weights of arcs) of the rows and columns of the incidence matrices (that can change its position). On the other hand, it is also convenient to define an alias for every place and transition. The alias will be attached to the position in the incidence matrices, in the way that the first row of the incidence matrix will always be associated to the alias p1, the second

In other words: C = swapc(swapr(A, i, j), u, v) The resulting incidence matrix, with the new alias for the places and transitions is presented below:

C=

t1r  tvr t1  tu a11  a1v   a j1  a jv   ai1  aiv  

 tur  tv  a1u  a ju  aiu 

 tnr  tn  a1n  a jn  ain 

p1

p1r

 pi  pj 

 p rj  pir 

am1  amv  amu  amn

pm

pmr

In this representation of the incidence matrix C, it can be seen that the reference names of the swapped rows and columns have changed the position according to these swaps. This change is a consequence of the fact that the reference names are related to structural and marking parameters (among others) such as the elements of the incidence matrices and not to positions in the matrix. If the Petri net is the model of a real

590

~ W( R1m )=addc(addc(addr(W( R1 ))) ~ W( R2m )=addc(addc(addr(W( R2 )))

system, the reference names are likely to be associated to a physical meaning as well. However, in certain applications it is very useful to define an alias for every place and transition. This alias is associated to the position that the elements of the incidence matrix occupy. The aliases do not bear the superindex “r”. For example, in the matrix C = swapr(A, i, j), the alias of the place whose input and output arcs are stated in the jth row is pj, whereas its

p2

r i

reference name is p .

~ R1

t1

p3

p2

p2

t2

m 1

W( R ) =

W( R2m ) =

p3

p1 p2

t1 -1 ~ W( R2 ) = 1 1

t2 1 -1 -1

p1 p2 p3

t2 1 -2 0 0

t3 0 0 0 0

p1 p2 p3 p4

t1 -1 1 1 0

t2 1 -1 -1 0

t3 0 0 0 0

p1 p2 p3 p4

The second set of equivalence operations will be the swapping of one row and one column in W( R1m ). The purpose of this operation may be to make the largest number of elements placed in the same position of both incidence matrices to coincide in order to reduce the number of undefined structural parameters of the resulting Petri net.

Fig. 1. Simple alternative Petri nets. t2 1 -2

t1 -1 2 0 0

Fig. 4. Incidence matrices of the alternative Petri nets after increasing their size.

t2

t1 -1 ~ W( R1 ) = 2

t3

The new incidence matrices are shown in the figure 4.

2 t2

p4

p3

t3

t1

2

p2

p4

Fig. 3. Addition of isolated places and transitions to increase the size of the incidence matrices.

~ R2

p1

p1 t1

2 t2

Example 2. Let us consider the simple alternative Petri nets presented in the figure 1 and their incidence matrices shown in the figure 2. Some equivalence operations will be applied to transform the simple alternative Petri nets into matching ones able to be merged. The result of this merging is obtaining an equivalent compound Petri net with the smallest size of the set of undefined structural parameters and the smallest size of the feasible combination of values of these parameters. In this example it is not intended to obtain the optimal compound Petri net, just to illustrate the application of some equivalence operations. p1

p1 t1 2

R2m

R1m

swapc

Fig. 2. Incidence matrices of the simple alternative Petri nets.

m 1

W( R ' ) = swapr

The first equivalence operation consists of increasing the size of the incidence matrix to reach the dimensions 4u3. This process require the addition of isolated places and transitios as it can be seen in the figure 3.

t1 -1 0 0 2

t2 0 0 0 0

t3 1 0 0 -2

p1 p2 p3 p4 m

Fig. 5. Incidence matrix of R1 ' after the swapping operations. The process may be seen in the figure 5 and correspond to the operations:

The operations that have been applied are the following:

591

Latorre, J.I., Jiménez, E., Pérez, M., 2007 Macro-Reachability Tree Exploration for D.E.S. Design Optimization. Proceedings of the 6th EUROSIM Congress on Modelling and Simulation (Eurosim 2007). Ljubljana, Slovenia.

~ W( R1m ' )=swapr(swapc(addr(W( R1 ),2,3),2,4) After this last operation it is possible to merge the incidence matrices of both alternative Petri nets to obtain a single compound Petri net.

Latorre, J.I., Jiménez, E., Pérez, M., Martínez, E., 2009 The design of a manufacturing facility. An efficient approach based on alternatives aggregation Petri Proceedings of the 21st European Modelling and Simulation Symposium (EMSS 09). Puerto de la Cruz, Spain, vol. 2, pp. 33-39.

4.

CONCLUSIONS AND FURTHER RESEARCH In this paper it has been shown how it is possible to apply matrix-based operations to the incidence matrices of a Petri net that preserve the structure of the graph of reachable markings. As a consequence, the properties and behaviour of the original and resulting Petri nets are equivalent. This idea constitutes a powerful tool that allows, in the application described in this document, to transform a set of alternative Petri nets into a compound Petri net with a set of undefined parameters is expected to be reduced, comparing with a case where the matrix-based operations are not applied. As future research it is expected to extend the application of these ideas to other fields in the modeling of discrete event systems, as well as to develop new matrix-based operations that allow obtaining equivalent Petri nets.

Latorre, J.I., Jiménez, E., Pérez, M., Blanco, J., 2009 The problem of designing discrete event systems. A new methodological approach. Proceedings of the 21st European Modelling and Simulation Symposium (EMSS 09). Puerto de la Cruz, Spain, vol. 2, pp. 40-46. Latorre, J.I., Jiménez, E., Pérez, M., 2010 Control of Discrete Event Systems Modelled by Petri Nets. Proceedings of the 7th EUROSIM Congress. Prague. Latorre, J.I., Jiménez, E., Pérez, M., 2011 Efficient Representations Of Alternative Models Of Discrete Event Systems Based On Petri Nets. Proceedings of the UKSim 13th International Conference on Computer Modelling and Simulation. Cambridge, United Kingdom.

REFERENCES Berthelot, G., 1987 Transformations and decompositions of nets in “Petri Nets: Central Models and Their Properties, Advances in Petri Nets”. Lecture Notes in Computer Science, Brauer, W., Reisig, W., and Rozenberg, G. (eds.), vol. 254-I, pp. 359–376. Springer.

Peterson, J.L., 1981 Petri Net Theory and the Modelling of Systems. Prentice Hall, Englewood Cliffs.

Cassandras, Christos G., Lafortune, S., 2008 Introduction to Discrete Event Systems. Second Edition, Springer.

Silva, M., 1993 Introducing Petri nets. In Practice of Petri Nets in Manufacturing”, Di Cesare, F., (eds.), pp. 1-62. Ed. Chapman&Hall.

David, R., Alla. H., , 2005 Discrete, Continuous and Hybrid Petri nets, Springer

Van der Aalst., W.M.P., 1997 Verification of Workflow Nets. In "Application and Theory of Petri Nets 1997", P. Azéma and G. Balbo (eds.), volume 1248 of Lecture Notes in Computer Science, pages 407426. Springer-Verlag, Berlin.

Haddad, S. and Pradat-Peyre, J.F., 2006 New Efficient Petri Nets Reductions for Parallel Programs Verification. Parallel Processing Letters, pages 101-116, World Scientific Publishing Company. Jiménez, E., Pérez, M., Latorre, J.I., 2005 On deterministic modelling and simulation of manufacturing systems with Petri nets. Proceedings of European Modelling Simulation Symposium. Marseille, pp. 129-136. Marseille.

592

SYNTHESIS OF FEEDBACK CONTROLLER FOR STABILIZATION OF CHAOTIC HÉNON MAP OSCILLATIONS BY MEANS OF ANALYTIC PROGRAMMING Roman Senkerik(a), Zuzana Oplatkova(a), Ivan Zelinka(b), Donald Davendra(b), Roman Jasek(a) (a)

Tomas Bata University in Zlin , Faculty of Applied Informatics, Nam T.G. Masaryka 5555, 760 01 Zlin, Czech Republic (b) Technical University of Ostrava, Faculty of Electrical Engineering and Computer Science, 17. listopadu 15, 708 33 Ostrava-Poruba, Czech Republic (a)

[email protected], (b)[email protected]

ABSTRACT This research deals with a synthesis of control law by means of analytic programming for selected discrete chaotic system – Hénon Map. The novelty of the approach is that analytic programming as a tool for symbolic regression is used for the purpose of stabilization of higher periodic orbits, which represent the oscillations between several values of chaotic system. The paper consists of the descriptions of analytic programming as well as used chaotic system and detailed proposal of cost function used in optimization process. For experimentation, SelfOrganizing Migrating Algorithm (SOMA) with analytic programming and Differential evolution (DE) as second algorithm for meta-evolution were used.

Instead of EA utilization, analytic programming (AP) is used in this research. AP is a superstructure of EAs and is used for synthesis of analytic solution according to the required behaviour. Control law from the proposed system can be viewed as a symbolic structure, which can be synthesized according to the requirements for the stabilization of the chaotic system. The advantage is that it is not necessary to have some “preliminary” control law and to estimate its parameters only. This system will generate the whole structure of the law even with suitable parameter values. This work is focused on the expansion of AP application for synthesis of a whole control law instead of parameters tuning for existing and commonly used method control law to stabilize desired Unstable Periodic Orbits (UPO) of chaotic systems. This research is an extension of previous research (Oplatkova et al. 2010a; 2010b, Senkerik et al., 2010c) focused on stabilization of simple p-1 orbit – stable state. In general, this research is concerned to stabilize p-2 UPO – higher periodic orbits (oscillations between two values). Firstly, AP is explained, and then a problem design is proposed. The next sections are focused on the description of used cost function and evolutionary algorithms. Results and conclusion follow afterwards.

Keywords: Chaos Control, Analytic programming, optimization, evolutionary algorithms. 1. INTRODUCTION The interest about the interconnection between evolutionary techniques and control of chaotic systems is spread daily. First steps were done in (Senkerik et al. 2010a; 2010b), (Zelinka et al. 2009) where the control law was based on Pyragas method: Extended delay feedback control – ETDAS (Pyragas 1995). These papers were concerned to tune several parameters inside the control technique for chaotic system. Compared to this, current presented research shows a possibility as to how to generate the whole control law (not only to optimize several parameters) for the purpose of stabilization of a chaotic system. The synthesis of control law is inspired by the Pyragas’s delayed feedback control technique (Just 1999, Pyragas 1992). Unlike the original OGY (Ott, Grebogi and York) control method (Ott et al.1990), it can be simply considered as a targeting and stabilizing algorithm together in one package (Kwon 1999). Another big advantage of the Pyragas method for evolutionary computation is the amount of accessible control parameters, which can be easily tuned by means of evolutionary algorithms (EA).

2. ANALYTIC PROGRAMMING Basic principles of the AP were developed in 2001 (Zelinka et al. 2005). Until that time only Genetic Programming (GP) and Grammatical Evolution (GE) had existed. GP uses Genetic Algorithms (GA) while AP can be used with any EA, independently on individual representation. To avoid any confusion, based on the nomenclature according to the used algorithm, the name - Analytic Programming was chosen, since AP represents synthesis of analytical solution by means of EA. The core of AP is based on a special set of mathematical objects and operations. The set of mathematical objects is a set of functions, operators and so-called terminals (as well as in GP), which are usually constants or independent variables. This set of variables

593

is usually mixed together and consists of functions with different number of arguments. Because of a variability of the content of this set, it is termed the “general functional set” – GFS. The structure of GFS is created by subsets of functions according to the number of their arguments. For example GFSall is a set of all functions, operators and terminals, GFS3arg is a subset containing functions with only three arguments, GFS0arg represents only terminals, etc. The subset structure presence in GFS is vitally important for AP. It is used to avoid synthesis of pathological programs, i.e. programs containing functions without arguments, etc. The content of GFS is dependent only on the user. Various functions and terminals can be mixed together (Zelinka et al. 2005, Zelinka et al. 2008, Oplatkova et al. 2009). The second part of the AP core is a sequence of mathematical operations, which are used for the program synthesis. These operations are used to transform an individual of a population into a suitable program. Mathematically stated, it is a mapping from an individual domain into a program domain. This mapping consists of two main parts. The first part is called Discrete Set Handling (DSH) (See Figure 1) (Zelinka et al. 2005, Lampinen and Zelinka 1999) and the second one stands for security procedures which do not allow synthesizing pathological programs. The method of DSH, when used, allows handling arbitrary objects including nonnumeric objects like linguistic terms {hot, cold, dark…}, logic terms (True, False) or other user defined functions. In the AP, DSH is used to map an individual into GFS and together with security procedures creates the above-mentioned mapping, which transforms arbitrary individual into a program.

Figure 2: The main principles of AP 3. PROBLEM DESIGN The brief description of used chaotic system and original feedback chaos control method ETDAS (Pyragas 1995) is given here. The ETDAS control technique was used in this research as an inspiration for synthesizing a new feedback control law by means of evolutionary techniques. 3.1. Selected Chaotic System The chosen example of chaotic system was the two dimensional Hénon map in form (1):

x n 1

a  x n2  by n

y n 1

xn

(1)

This is a model invented with a mathematical motivation to investigate chaos. The Hénon map is a discrete-time dynamical system, which was introduced as a simplified model of the Poincaré map for the Lorenz system. It is one of the most studied examples of dynamical systems that exhibit chaotic behavior. The map depends on two parameters, a and b, which for the canonical Hénon map have values of a = 1.4 and b = 0.3. For these canonical values the Hénon map is chaotic (Hilborn 2000). The example of this chaotic behavior can be clearly seen from bifurcation diagram – Figure 3.

Figure 1: Discrete set handling AP needs some EA (Zelinka et al. 2005) that consists of a population of individuals for its run. Individuals in the population consist of integer parameters, i.e. an individual is an integer index pointing into GFS. The creation of the program can be schematically observed in Figure 2. The individual contains numbers which are indices into GFS. The detailed description is represented in (Zelinka et al. 2005, Zelinka et al. 2008, Oplatkova et al. 2009). AP exists in 3 versions – basic without constant estimation, APnf – estimation by means of nonlinear fitting package in Mathematica environment and APmeta – constant estimation by means of another evolutionary algorithms; meta implies metaevolution.

Figure 3: Bifurcation diagram of Hénon Map Figure 3 shows the bifurcation diagram for the Hénon map created by plotting of a variable x as a function of

594

the one control parameter for the fixed second parameter.

3.3. Cost Function Proposal for the cost function comes from the simplest Cost Function (CF). The core of CF could be used only for the stabilization of p-1 orbit. The idea was to minimize the area created by the difference between the required state and the real system output on the whole simulation interval – IJi. But another universal cost function had to be used for stabilizing of higher periodic orbit and having the possibility of adding penalization rules. It was synthesized from the simple CF and other terms were added. In this case, it is not possible to use the simple rule of minimizing the area created by the difference between the required and actual state on the whole simulation interval – IJi, due to many serious reasons, for example: degrading of the possible best solution by phase shift of periodic orbit. This CF is in general based on searching for desired stabilized periodic orbit and thereafter calculation of the difference between desired and found actual periodic orbit on the short time interval - IJs (40 iterations) from the point, where the first min. value of difference between desired and actual system output is found. Such a design of CF should secure the successful stabilization of either p-1 orbit (stable state) or higher periodic orbit anywise phase shifted. The CFBasic has the form (6).

3.2. ETDAS Control Method This work is focused on explanation of application of AP for synthesis of a whole control law instead of demanding tuning of EDTAS (Pyragas 1995) method control law to stabilize desired Unstable Periodic Orbits (UPO). In this research desired UPO is only p-2 (higher periodic orbit – oscillation between two values). ETDAS method was obviously an inspiration for preparation of sets of basic functions and operators for AP. The original control method – ETDAS has form (2): F (t ) S (t )

K > 1  R S t  W d  x(t )@ ,

x(t )  RS t  W d ,

(2)

where: K and R are adjustable constants, F is the perturbation; S is given by a delay equation utilizing previous states of the system and W d is a time delay. The original control method – ETDAS in the discrete form suitable for two-dimensional Hénon Map has the form (3): x n 1

a  x n2  by n  Fn ,

Fn

K > 1  R S n  m  x n @ ,

Sn

x n  RS n  m ,

(3)

W2

CFBasic

where: m is the period of m-periodic orbit to be stabilized. The perturbation Fn in equations (3) may have arbitrarily large value, which can cause diverging of the system outside the interval {0, 1.0}. Therefore, Fn should have a value between  Fmax , Fmax . In this preliminary study a suitable Fmax value was taken from the previous research. To find the optimal value also for this parameter is in future plans. Previous research concentrated on synthesis of control law only for p-1 orbit (a fixed point). An inspiration for preparation of sets of basic functions and operators for AP was simpler TDAS (Pyragas 1992) control method (4) and its discrete form given in (5): F (t ) Fn

K >x t  W  x(t )@ ,

K x nm  x n .

pen1  ¦ TS t  AS t ,

(6)

t W1

where: TS - target state, AS - actual state IJ1 - the first min value of difference between TS and AS IJ2 – the end of optimization interval (IJ1+ IJs) pen1= 0 if IJi - IJ2 • IJs; pen1= 10*( IJi - IJ2) if IJi - IJ2 < IJs (i.e. late stabilization). 4. USED EVOLUTIONARY ALGORITHMS This research used two evolutionary algorithms: SelfOrganizing Migrating Algorithm (Zelinka 2004) and Differential Evolution (Price and Storn 2001, Price 2005). Future simulations expect a usage of soft computing GAHC algorithm (modification of HC12) (Matousek 2007) and a CUDA implementation of HC12 algorithm (Matousek 2010).

(4) (5)

4.1. Self Organizing Migrating Algorithm – SOMA SOMA is a stochastic optimization algorithm that is modelled on the social behaviour of cooperating individuals (Zelinka 2004). It was chosen because it has been proven that the algorithm has the ability to converge towards the global optimum (Zelinka 2004). SOMA works with groups of individuals (population) whose behavior can be described as a competitive – cooperative strategy. The construction of a new population of individuals is not based on evolutionary principles (two parents produce offspring) but on the

Compared to this work, the data set for AP presented in the previous research required only constants, operators like plus, minus, power and output values x n and x n1 . Due to the recursive attributes of delay equation S utilizing previous states of the system in discrete ETDAS (3), the data set for AP had to be expanded and cover longer system output history ( x n to xn9 .), thus to imitate inspiring control method for the successful synthesis of control law securing the stabilization of higher periodic orbits

595

4.2. Differential Evolution

behavior of social group, e.g. a herd of animals looking for food. This algorithm can be classified as an algorithm of a social environment. To the same group of algorithms, Particle Swarm Optimization (PSO) algorithm can also be classified sometimes called swarm intelligence. In the case of SOMA, there is no velocity vector as in PSO, only the position of individuals in the search space is changed during one generation, referred to as ‘migration loop’. The rules are as follows: In every migration loop the best individual is chosen, i.e. individual with the minimum cost value, which is called the Leader. An active individual from the population moves in the direction towards the Leader in the search space. At the end of the crossover, the position of the individual with minimum cost value is chosen. If the cost value of the new position is better than the cost value of an individual from the old population, the new one appears in new population. Otherwise the old one remains there. The main principle is depicted in Figures 4 and 5.

DE is a population-based optimization method that works on real-number-coded individuals (Price 2005). & For each individual xi ,G in the current generation G, DE & generates a new trial individual xic,G by adding the weighted difference between two randomly selected & & individuals xr1,G and x r 2,G to a randomly selected third & & individual xr 3,G . The resulting individual xic,G is & crossed-over with the original individual xi ,G . The fitness of the resulting individual, referred to as a & perturbed vector u i ,G 1 , is then compared with the & & fitness of xi ,G . If the fitness of u i ,G 1 is greater than the & & & fitness of xi ,G , then xi ,G is replaced with u i ,G 1 ; & & otherwise, xi ,G remains in the population as xi ,G 1 . DE is quite robust, fast, and effective, with global optimization ability. It does not require the objective function to be differentiable, and it works well even with noisy and time-dependent objective functions. Description of used DERand1Bin strategy is presented in (7). Please refer to (Price and Storn 2001, Price 2005) for the description of all other strategies.

ui , G  1

xr1, G  F x xr 2, G  xr 3, G

(7)

5. SIMULATION RESULTS As described in section 2 about Analytic Programming, AP requires some EA for its run. In this paper APmeta version was used. Meta-evolutionary approach means usage of one main evolutionary algorithm for AP process and second algorithm for coefficient estimation, thus to find optimal values of constants in the evolutionary synthesized control law. SOMA algorithm was used for main AP process and DE was used in the second evolutionary process. Settings of EA parameters for both processes were based on performed numerous experiments with chaotic systems and simulations with APmeta (See Table 1 and Table 2).

Figure 4: Principle of SOMA, movement in the direction towards the Leader

Table 1. SOMA settings for AP Parameter Value PathLength 3 Step 0.11 PRT 0.1 PopSize 50 Migrations 4 Max. CF Evaluations (CFE) 5345 Table 2. DE settings for meta-evolution Parameter Value PopSize 40 F 0.8 CR 0.8 Generations 150 Max. CF Evaluations (CFE) 6000

Figure 5: Basic principle of crossover in SOMA, PathLength is replaced here by Mass

596

The Analytic Programming used following setting-up: Basic set of elementary functions for AP: GFS2arg= +, -, /, *, ^ GFS0arg= datan-9 to datan, K

5.2. Example 2 The second example of a new synthesized feedback control law Fn (perturbation) for the controlled Hénon map (8) inspired by original ETDAS control method (3) has the form (11) - direct output from AP and form (12) – with estimated coefficients.

Total number of cost function evaluations for AP was 5345, for the second EA it was 6000, together 32.07 millions per each simulation. The novelty of this approach represents the synthesis of feedback control law Fn (8) (perturbation) for the Hénon Map inspired by original ETDAS control method. a  xn2  by n  Fn

xn1

Fn

· § x xn5 xn1 K1  xn3 ¨¨  n 7  xn ¸¸ ¹ © K3 · § x n 7 K 2 ¨¨   xn6  xn4  xn1 ¸¸ xn x x x ¹ © n 5 n  2 n

Fn



(8)

Following two presented simulation results represent the best examples of synthesized control laws. Based on the mathematical analysis, the real p-2 UPO for unperturbed logistic equation has following values: x1 = - 0.5624, x2 = 1.2624. Description of the two selected simulation results covers direct output from AP – synthesized control law without coefficients estimated; further the notation with simplification after estimation by means of second algorithm DE and corresponding CF value.

xn5 xn1 xn3  25.168  0.5402 xn7  xn · § xn 7 4.4124¨¨   xn6  xn4  xn1 ¸¸ xn ¹ © x n 5 x n  2 x n

(11)

(12)

Simulation output representing successful and quick stabilization of Hénon map is depicted in Figure 7. The CF Value was 0.7540 (average error 0.01885 per iteration). 1.5

5.1. Example 1 The first example of a new synthesized feedback control law Fn (perturbation) for the controlled Hénon map (8) inspired by original ETDAS control method (3) has the form (9) – direct output from AP and form (10) – with estimated coefficients by means of the second EA. xn1 K 2  xn3  xn K1

1.0

x

0.5 0.0 0.5

Fn



Fn

0.342699 xn1 0.7  xn3  xn

(9) 1.0

0

(10)

1.0

x

0.5 0.0 0.5 1.0

100 Iteration

150

150

200

6. CONCLUSION This paper deals with a synthesis of a control law by means of AP for stabilization of selected chaotic system at higher periodic orbit. Hénon map as an example of two-dimensional discrete chaotic system was used in this research. In this presented approach, the analytic programming was used instead of tuning of parameters for existing control technique by means of EA’s as in the previous research. Obtained results reinforce the argument that AP is able to solve this kind of difficult problems and to produce a new synthesized control law in a symbolic way securing desired behaviour of chaotic system and stabilization. Presented two simulation examples show two different results. Extremely precise stabilization and simple control law in the first case and not very precise and slow stabilization and relatively complex notation of chaotic controller in the second case. This fact lends

1.5

50

100 Iteration

Figure 7: Simulation results - the second example

Simulation depicted in Figure 6 lends weight to the argument, that AP is able to synthesize a new control law securing very quick and very precise stabilization. The CF Value was 3.8495.10-12, which means that average error between actual and required system output was 9.6237.10-14 per iteration.

0

50

200

Figure 6: Simulation results - the first example

597

Pyragas K., 1995. Control of chaos via extended delay feedback, Physics Letters A, vol. 206, 1995, pp. 323-330. Senkerik, R., Zelinka, I., Davendra, D., Oplatkova, Z, 2010a, Evolutionary Design of Chaos Control in 1D, In. Zelinka I., Celikovski S., Richter H., Chen G.: Evolutionary Algorithms and Chaotic Systems, SpringerVerlag Berlin, pp.165 - 190. Senkerik R., Zelinka I., Davendra D., Oplatkova Z., 2010b, Utilization of SOMA and differential evolution for robust stabilization of chaotic Logistic equation, Computers & Mathematics with Applications, Volume 60, Issue 4, pp. 1026-1037. Senkerik R., Oplatkova Z., Zelinka I., Davendra D.,Jasek R., 2010c, Synthesis Of Feedback Controller For Chaotic Systems By Means Of Evolutionary Techniques, Proceeding of Fourth Global Conference on Power Control and Optimization, Sarawak, Borneo, 2010,. Zelinka I., 2004. “SOMA – Self Organizing Migrating Algorithm”, In: New Optimization Techniques in Engineering, (B.V. Babu, G. Onwubolu (eds)), chapter 7, 33, Springer-Verlag, 2004, ISBN 3-540-20167X. Zelinka I.,Oplatkova Z, Nolle L., 2005. Boolean Symmetry Function Synthesis by Means of Arbitrary Evolutionary Algorithms-Comparative Study, International Journal of Simulation Systems, Science and Technology, Volume 6, Number 9, August 2005, pages 44 - 56, ISSN: 1473-8031. Zelinka I., Senkerik R., Navratil E., 2009, Investigation on evolutionary optimization of chaos control, Chaos, Solitons & Fractals, Volume 40, Issue 1, pp. 111-129. Zelinka, I., Guanrong Ch., Celikovsky S., 2008. Chaos Synthesis by Means of Evolutionary algorithms, International Journal of Bifurcation and Chaos, Vol. 18, No. 4, pp. 911–942

weight to the argument, that AP is a powerful symbolic regression tool, which is able to strictly and precisely follow the rules given by cost function and synthesize any symbolic formula, in the case of this research – the feedback controller for chaotic system. The question of energy costs and more precise stabilization will be included into future research together with development of better cost functions, different AP data set, and performing of numerous simulations to obtain more results and produce better statistics, thus to confirm the robustness of this approach. ACKNOWLEDGMENTS This work was supported by the grant NO. MSM 7088352101 of the Ministry of Education of the Czech Republic and by grants of Grant Agency of Czech Republic GACR 102/09/1680 and by European Regional Development Fund under the project CEBIATech No. CZ.1.05/2.1.00/03.0089. REFERENCES Hilborn R.C., 2000. Chaos and Nonlinear Dynamics: An Introduction for Scientists and Engineers, Oxford University Press, 2000, ISBN: 0-19-850723-2. Just W., 1999, Principles of Time Delayed Feedback Control, In: Schuster H.G., Handbook of Chaos Control, WileyVch, ISBN 3-527-29436-8. Kwon O. J., 1999. Targeting and Stabilizing Chaotic Trajectories in the Standard Map, Physics Letters A. vol. 258, 1999, pp. 229-236. Lampinen J., Zelinka I., 1999, New Ideas in Optimization – Mechanical Engineering Design Optimization by Differential Evolution, Volume 1, London: McGraw-hill, 1999, 20 p., ISBN 007-709506-5. Matousek R., 2007, GAHC: Improved GA with HC station, In WCECS 2007, San Francisco, pp. 915Ǧ920. ISBN: 978Ǧ988Ǧ98671Ǧ6Ǧ4. Matousek R., 2010, HC12: The Principle of CUDA Implementation. In MENDEL 2010, Mendel Journal series, pp. 303Ǧ308. ISBN: 978Ǧ80Ǧ214Ǧ4120Ǧ 0. ISSN: 1803Ǧ 3814. Oplatková, Z., Zelinka, I.: 2009. Investigation on Evolutionary Synthesis of Movement Commands, Modelling and Simulation in Engineering, Vol. 2009. Oplatkova Z., Senkerik R., Zelinka I., Holoska J., 2010a, Synthesis of Control Law for Chaotic Henon System Preliminary study, ECMS 2010, Kuala Lumpur, Malaysia, p. 277-282, ISBN 978-0-9564944-0-5. Oplatkova Z., Senkerik R., Belaskova S., Zelinka I., 2010b, Synthesis of Control Rule for Synthesized Chaotic System by means of Evolutionary Techniques, Mendel 2010, Brno, Czech Republic, p. 91 - 98, ISBN 978-80-2144120-0. Ott E., C. Greboki, J.A. Yorke, 1990. Controlling Chaos, Phys. Rev. Lett. vol. 64, 1990, pp. 1196-1199. Price K., Storn R. M., Lampinen J. A., 2005, Differential Evolution : A Practical Approach to Global Optimization, Natural Computing Series, Springer. Price, K. and Storn, R. (2001), Differential evolution homepage, [Online]: http://www.icsi.berkeley.edu/~storn/code.html, [Accessed 30/04/2011]. Pyragas K., 1992, Continuous control of chaos by selfcontrolling feedback, Physics Letters A, 170, 421-428.

598

A SIMULATION STUDY OF THE INTERDEPENDENCE OF SCALABILITY AND CANNIBALIZATION IN THE SOFTWARE INDUSTRY Francesco Novelli SAP Research Darmstadt, Germany [email protected]

premises delivery model, to launch SaaS counterparts to their established software products, in coexistence or as replacements. We focus on two key challenges inherent in such a strategic response. On the one hand, it is indispensable for the incumbent to understand the dynamics of revenue cannibalization and its consequences on profitability and on the positioning of the firm vis-a-vis the competition. On the other hand, it is necessary to achieve the degree of technological and organizational scalability required to profitably implement a SaaS strategy.

ABSTRACT Dominant software vendors, whose applications have been predominantly delivered on-premises so far (i.e., installed, maintained and operated at customers’ premises) are challenged by the rising adoption of software as a service (SaaS) solutions, outsourced applications delivered through the web under subscription or usagebased pricing terms. As a response, these incumbent vendors extend their product portfolios with SaaS offerings, but thus risk to engender revenue cannibalization, as a newly introduced SaaS application may attract their own on-premises customers instead of expanding the market or drawing from a competitor’s customer base. At the same time they face the novel, severe scalability requirements of the technological and organizational infrastructure underlying a successful SaaS business. Using an agent-based simulation model, we study the interdependence between cannibalization and scalability in monopolistic and duopolistic software markets.

1.2. Cannibalization Cannibalization is the switching of sales from an existing product to a new one of the same firm (Newman 1967). It is an issue of paramount importance for an incumbent vendor venturing into SaaS, given the intrinsic degree of substitutability between the already established on-premises products and their SaaS siblings. This may indeed put pre-existent revenue streams and market shares at stake. As a case in point, let us consider how competition is unfolding in the office automation market. Microsoft is the dominant player with 6 billion dollars revenue from that segment in the second quarter of 2011 (as a term of comparison, 5 billion was the revenue from the Windows OS). However, the entry into the market of free online office applications (such as those by Google) has pushed Microsoft to respond with the development of two SaaS alternatives to its well-known Office suite: a free, ad-supported one with limited functionalities and a subscription-based one with enhanced collaboration features. The delicate challenge is for Microsoft to tame the cannibalization effect this move may engender, i.e., to avert a financially harmful drift of on-premises customers to its own SaaS offerings. Cannibalization may also represent a deliberate, offensive product strategy, pursued to drive growth (McGrath 2001). As a matter of fact, some on-premises vendors have successfully managed the transition to a hybrid or purely SaaS model. Concur Technologies, for instance, paired its on-premises offerings with the Application Service Provider model (predecessor of SaaS) in the late 90s already, and then transitioned to become a purely SaaS player just as this delivery model emerged (Warfield 2007). Analogously, Ariba started

Keywords: software as a service, cannibalization, scalability, agent-based modelling and simulation 1.

INTRODUCTION

1.1. Evolution of Software Delivery and Pricing Models In the last decade the software industry has witnessed the emergence of the so-called software as a service (SaaS) delivery model, whereby vendors provide webbased, outsourced software applications (SIIA 2001), dispensing customers with most installation, operation and maintenance activities otherwise needed at their premises. Moreover, SaaS solutions are usually coupled with subscription or usage-based pricing models (Lehmann and Buxmann 2009), lowering the initial investment in comparison with packaged software. As a matter of fact, several beholders of the software industry agree that the adoption of SaaS applications has gained momentum: Information Systems researchers (Benlian, Hess, and Buxmann 2009), IT market analysts (Gartner 2010), and investors, who, in the first quarter of 2011, gave SaaS public companies an average market evaluation 6.5 times their annual revenues. This trend, in turn, urges incumbent vendors, which built their dominant market positions on the on-

599

the transition in 2006 and gradually ported all its applications to a SaaS model, now the generator of most of its revenues (Wainewright 2009). Both companies initially went after a SaaS-enabled market expansion aimed towards the mid-segment but eventually – in sharp contrast with the Microsoft scenario – found it profitable to deliberately cannibalize their on-premises customers along the way.

Total Expenses as % of Total Revenues

it in the 60-70% range, and two SaaS competitors, unable, even after having successfully ridden a steep learning curve, to lower it under the 90% mark (Salesforce) or to reach operational profitability at all (NetSuite).

1.3. Scalability Understanding the financial and competitive consequences of revenue cannibalization and then attempting to avert it or to ride it are not the only concerns facing incumbents. The SaaS delivery model poses a scalability threat as well, both from a technological and from an organizational perspective. This threat stems from the peculiarities of this newly addressable market. Since SaaS lowers the technological and financial requirements for a software purchase, the market swells in number of potential buyers while the average financial and technological resources available to them decrease. As a matter of fact, since a SaaS offering is hosted and operated by the provider and accessed through the World Wide Web, very simple applications that do not demand supplemental integration and customization virtually appeal to any organization meeting the minimum technical requirement of an available Internet access. From a financial point of view, the SaaS subscription fees dilute over time the investment for the license of a given software functionality. Therefore, small and medium-size companies can, in spite of their usually more limited IT budget and technical personnel, enter application markets once populated by large enterprises only. This is exemplarily shown for the European Union in Table 1. It evidently represents a huge market opportunity to be tapped into by software vendors in terms of number of potential new accounts (enterprises) and users (employees).

Small 10-49

Medium 50-249

Large 250 or more

Number of enterprises* Persons employed* (million) * % enterprises that employ † IT/ICT specialists % enterprises with an ERP system# % enterprises with a CRM system#

1,408,000 27.9

228,000 23.4

45,000 45.2

9%

25%

53%

17%

41%

64%

23%

39%

54%

NetSuite

Oracle

Salesforce

SAP

200%

150%

100%

50%

0% 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

Figure 1: Total Operating Expenses as a Percentage of Total Revenues for selected Software Vendors (source: corporate financial reports) As a matter of fact, a SaaS software vendor must bear the additional costs of setting up and operating the technological infrastructure needed to deliver the SaaS application. Moreover, beyond those technological repercussions, scalability issues engage the SaaS provider on an organizational level as well, for this new, more fragmented segment of software buyers imposes to think a series of processes anew. For instance marketing and sales, where using dedicated sales team for each account as it is the habit with large enterprises is not possible on a large scale, and other means, such as telesales and innovative internet-based funnels, need to be employed.

Table 1: Key indicators per enterprise size class in the EU-27 area; source: Eurostat (# as of 2010, * as of 2008, † as of 2007) Size class Employees

250%

1.4. The Interplay of Cannibalization and Scalability Incumbent software vendors introducing SaaS are confronted with a typical new product introduction problem, where the new product may divert current customers from other offerings of the same firm, instead of attracting new buyers or drawing from a competitor’s customer base (Kerin, Harvey, and Rothe 1978), further complicated by the trade-off between a more saturated but highly profitable current software market of large enterprises and a fast-growing but less profitable potential SaaS market. Scalability of the SaaS business is certainly a prerequisite to target market expansion. Without it vendors face the risk of not being able to satisfy demand – i.e., failing to build the appropriate level of capacity to ride growth – or doing so inefficiently – i.e., failing to reach the scale economies that make market expansion a profitable endeavour at all. Changing perspective, limiting scalability can be a radical lever to avert cannibalization, for it puts an upper bound on the volume of intrafirm switching cus-

This opportunity for market expansion has its downsides: though larger, the new potential market is more costly to be reached and to be served. Figure 1 compares the trend in total operating expenditures over total revenues for the two leading business application vendors (SAP and Oracle), which have historically kept

600

ments. Given space constraints, in this section we describe only the most salient features of the model.

tomers. This may expose the incumbent’s flank to competition though, whereby customers might switch to a competitor instead.

3.1. Software Application A software application is characterized by the features or benefits it provides (its “quality”) to its users, the price to be paid to obtain those benefits (in terms of amount and distribution over time of the fees) and the infrastructure on which it is deployed. When the application is delivered as SaaS, it will be deployed on an infrastructure operated by the software vendor and priced under subscription terms, with an initial activation charge at the time of purchase and an anticipated, recurrent fee for each period of the simulation in which it is used. When the application is delivered on-premises, the price structure will follow the typical enterprise application pricing model and be once again made up of two components: an initial charge to purchase the licenses and an anticipated, periodical maintenance fee as a percentage of the initial charge. While the fee structure we employ for the two delivery modes is the same, the proportion between initial and periodical charge differs, with on-premises application weighting more on the front – annual maintenance rates are around 20% of the initial investment for licenses (Buxmann, Diefenbach, and Hess 2011) – and SaaS diluting the expenses more over time.

This work is organized as follows: the reasons that led to the choice of Agent-Based Modelling and Simulation as our research methodology are summarized in the following section (2). In section 3 we describe the model we based our simulation experiments on, which are then detailed and discussed in section 4. Eventually, we conclude in section 5, mentioning the limitations of the present work and possible future developments. 2.

RESEARCH METHODOLOGY

2.1. Agent-Based Modelling and Simulation to Study Microeconomics A market represents an excellent example of Complex Adaptive System (CAS), a collection of adaptive agents (suppliers and customers) concurrently engaged in local interactions (commercial transactions). Local interactions produce higher level conditions (market prices, bandwagon effects, etc.) impacting in turn the way those same interactions will evolve. Among the paradigms used to investigate such systems we chose Agent-Based Modelling and Simulation (ABMS). In ABMS each interacting component is modelled as an autonomous decision-making agent with attributes and rules defining his behavioural characteristics and how those are to evolve or adapt (North and Macal 2007). This approach lends itself neatly to the exploitation of microeconomic constructs in modelling agents’ behaviours and interactions (game theory, for instance, to dictate an agent’s strategic response) and is therefore especially suitable to study a CAS populated by economic entities. In fact, the study of economics with ABMS has reached such a respectable status to beget its own specific field of research, called AgentBased Computational Economics (Tesfatsion 2002). In this work we use ABMS to investigate at a microeconomic level a stylized business application software market where a multi-product incumbent vendor runs the risk of revenue cannibalization. ABMS suits perfectly the study of this phenomenon since, offering the possibility to observe the behaviours and decisions of individual buyers, it allows a disaggregated analysis of the components of demand. This means identifying exactly which customers switch between software applications of the same vendor (cannibalization), leave for a competitor (competitive draw), or enter the market for the first time (market expansion). 3.

3.2. Software Application Vendors The software vendors are price-making suppliers of the same class of business application (e.g., ERP, CRM, etc.) but with different price-quality schedules and delivery models. In the case of a SaaS the vendor must also operate the infrastructure on which the application is deployed. In each simulation period vendors collect the due payments from the customers that adopted one of their applications and bear the costs of the SaaS infrastructure. 3.3. SaaS Infrastructure and Scalability The infrastructure is made up of a set of technological or organizational resources (e.g., servers, sales representatives) characterized by a certain individual performance. The overall performance is, however, not just the sum of these components, and depends on the level of scalability of the infrastructure. We use Amdahl’s law (Amdahl 1967) to account for this issue and formalize the degree of non-scalability of the infrastructure through a so-called contention rate, which exerts a negative impact on the ability to efficiently scale (and, as we will see, to compete) growing exponentially with the scale requirement (Shalom and Perry 2008). The maximum total capacity K of an infrastructure with N resources of throughput τ each, designed to have a CnR rate of contention is:

MODEL

We model a closed, vertically-differentiated software application market. The market structure is a monopoly with a single vendor selling both an on-premises application and a SaaS one for the first series of experiments, and a duopoly consisting of the same vendor plus a purely SaaS challenger for the second set of experi-

ߢൌ

601

߬ ͳ െ ‫ܴ݊ܥ‬ ‫ ܴ݊ܥ‬൅ ܰ

(1)

Equation (1) gives the total capacity a certain infrastructure can attain in a given period. For instance, 20 resources with throughput of 1000 customers per period each, arranged in an architecture designed to have a 20% contention rate, would generate a total capacity of 4167 customers per period. Doubling the resources (i.e., scaling out of 20 additional resources) 4545 customers could be served (an 8% increase). However, the maximum achievable capacity would be bounded to less than 5000 customers per period, no matter how many additional resources are thrown in. Reducing contention would be a much effective lever: decreasing the contention rate of 5% would increase capacity by 25% (to 5195 customers per period).

equation (3) when comparing the surplus of her current choice with other alternatives in the market. The offerings in the market have a certain initial market share each in terms of pre-assigned customers (the incumbent’s on-premises one being the largest), but the overall addressable market includes potential customers that will take their first buying decision during the simulation. 3.5. Implementation and Runtime Environment The whole model was implemented in Java using the ABMS open-source toolkit Repast Simphony 2.0 (North et al. 2007). 4.

3.4. Software Application Customers Customers are current or potential adopters of a software application sold in the market. The decision to adopt an application is made on the basis of the obtained surplus. The surplus for the i-th customer of type ߠ when adopting an application j is: ܵ௜௝ ൌ ߠ௜ ܳ௝ െ ܶ‫ܱܥ‬௝ ሺܶሻ൅‫ܺ ן‬௝

An experiment consisted of 10 replications of 21 periods of length (the equivalent of three 7-years software application life cycles in the temporal scale we chose) for each model configuration, where a model configuration was given by the contention rate of the incumbent’s SaaS infrastructure, specified in a 0%-50% interval with 5% steps. Each experiment was conducted in different growth and competitive scenarios as detailed in the two following sub-sections.

(2)

The first term of equation (2) is the willingness to pay of a customer with marginal valuation of quality ߠ for an application of quality Q. ߠ is an input parameter set randomly for each consumer at simulation start (drawn from a uniform distribution with support between 0 and 1). The second term is the present value of the total cost of ownership of the application (detailed below). The third addend is the network externality derived from all consumers that already adopted an application with the same delivery model (i.e., the relevant network ܺ௝ is the total number of SaaS customers if j is one of the SaaS applications, or the total number of onpremises customers if j the incumbent’s on-premises application). The total cost of ownership over a horizon of T years is computed for both on-premises and SaaS applications employing the formula for the present value of an annuity: ܶ‫ܱܥ‬௝ ሺܶሻ ൌ ߮௧బ ൅ ߮௧ ൅

߮௧ ሺͳ െ ሺͳ ൅ ‫ݎ‬ሻି்ାଵ ሻ ‫ݎ‬

EXPERIMENTS

4.1. Experiments in a Monopolistic Market Our first series of experiments dealt with a monopoly in two scenarios: one of high growth of the SaaS segment, where this has a total size (in terms of number of potential customers) 10 times the on-premises segment, and one of low growth, where it merely matches the onpremises segment’s size. In a monopolistic situation the decisions of the incumbent is linked to the trade-off between revenue cannibalization and market expansion. If the potential market tapped into with SaaS is large enough in size to offset the effect of cannibalizing the high-margin onpremises sales, then the vendor should pursue a highcapacity strategy and, therefore, invest in a scalable infrastructure. Otherwise cannibalization could be averted by limiting the capacity offered in the market with a more conservative strategy. Conversely, a company that has not yet reached the needed level of scalability would unprofitably pursue growth in the SaaS segment and should refrain from it. Examining the results of these first experiments, it can be seen that, in case of high-growth in the SaaS segment, the monopolist may indeed offset (in terms of sales volume) revenue cannibalization with market expansion by pursuing a high-scalability strategy (Figure 2). On the contrary a low-scalability strategy allows minimizing revenue cannibalization in a low-growth scenario, where no significant market expansion would be possible anyway (Figure 3). Please note that throughout the remainder of this section we calculated total cannibalized revenues in terms of the projected onpremises revenues lost when the customer switch to SaaS (i.e., the discounted stream of maintenance fees, as expressed by eq. 3).

(3)

where ߮௧బ is initial charge (activation of the SaaS subscription or on-premises license charge), ߮௧ the anticipated periodical charge (the subscription fee or the maintenance fee respectively), and r the annual interest rate. When taking a purchase decision, a consumer first calculates (2) for every available application, then adopts the one with highest non-negative surplus among those with available capacity offered in the market. Although we do not explicitly model switching costs, the consumer that has already adopted an application considers the initial charge a sunk cost and drop ߮௧బ from

602

The specific contribution margins of the two software products will dictate the overall effect on the monopolist’s profit. Given the higher margins enjoyed in delivering on-premises applications (see introduction), mirrored in our model, a multi-product monopolist in a low-growth scenario would be better off slowing the rate of SaaS adoption among its own customers by limiting the offered capacity (Figure 4). On the contrary, being able to scale to expand into the SaaS segment would be, in case of high growth, the more profitable strategy.

4.2. Experiments in a Duopolistic Market In the presence of a SaaS challenger the incumbent’s on-premises customers can switch to either the incumbent’s SaaS offering or the competitor’s one, adding the risk of competitive draw to the strategic considerations of the incumbent. This risk can be more or less pronounced depending on the scalability of the challenger’s SaaS infrastructure. We therefore define four basic scenarios, showed in Table 2. Table 2: Basic Scenarios for the Duopolistic Market

600000

Challenger’s Scalability Low High (CnR = 0.3) (CnR = 0.05) Scenario Scenario LG1 LG2 Scenario Scenario HG1 HG2

Revenue Cannibalization Market Expansion

500000

Growth in the SaaS Segment

400000

300000

Low (1X) High (10X)

200000 100000

In confronting a high-scalable SaaS challenger it always pays for the incumbent to be able to match the competitor’s scale, because this allows at least retaining through cannibalization customers that would otherwise be lost (Scenario LG2, Figure 5) if not even offsetting any competitive draw or cannibalization effect by riding growth (Scenario HG2, Figure 6).

0 0

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Incumbent Contention Rate

Figure 2: Total Cannibalized On-Premises Revenues and Total SaaS Market Expansion in a Scenario of High Growth (Average for 10 replications) 2500000

Revenue Cannibalization Competitive Draw

Revenue Cannibalization Market Expansion

2000000

Market Expansion

1200000

1500000

1000000 800000

1000000

600000 400000

500000

200000

0

0

0

0

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

0.05

0.1

Incumbent Contention Rate

Total Profit

LG Scenario

0.2

0.25

0.3

0.35

0.4

0.45

0.5

Figure 5: Total Cannibalized On-Premises Revenues, Total SaaS Market Expansion, and Total Competitive Draw of On-Premises Revenues by the SaaS Challenger in Scenario LG2 (Average for 10 replications)

Figure 3: Total Cannibalized On-Premises Revenues and Total SaaS Market Expansion in a Scenario of Low Growth (Average for 10 replications) 6000000

0.15

Incumbent Contention Rate

HG Scenario

Revenue Cannibalization Competitive Draw

5000000 4000000

Market Expansion

600000 500000

3000000

400000 300000

2000000

200000

1000000

100000 0

0

0

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

Incumbent Contention Rate

Incumbent Contention Rate

Figure 6: Total Cannibalized On-Premises Revenues, Total SaaS Market Expansion, and Total Competitive Draw of On-Premises Revenues by the SaaS Challenger in Scenario HG2 (Average for 10 replications)

Figure 4: Incumbent’s Total Profit in the two Monopolistic Scenarios (HG = High Growth, LG = Low Growth; Average for 10 replications)

603

Total Profit

As shown in Figure 7, the incumbent’s total profit is generally higher in case of high-growth and negatively correlated with contention, except for the particular case of low growth and presence of a non-scalable challenger (scenario LG1), where the option to limit capacity as a lever to control cannibalization could still be viable. This is due to the lower risk of losing relevant market shares to a poorly-scalable competitor.

REFERENCES Amdahl, G., M., 1967. Validity of the single processor approach to achieving large scale computing capabilities. Proceedings of the April 18-20, 1967, spring joint computer conference (pp. 483-485). Benlian, A., Hess, T., and Buxmann, P., 2009. Drivers of SaaS-adoption–an empirical study of different application types. Business & Information Systems Engineering, 1(5), 357–369. Buxmann, P., Diefenbach, H., and Hess, T., 2011. Die Softwareindustrie: Ökonomische Prinzipien, Strategien, Perspektiven. Springer, Berlin. Gartner, 2010. Gartner Survey Indicates More Than 95 Percent of Organizations Expect to Maintain or Grow Their Use of SaaS Through 2010. Available at: http://www.gartner.com/it/page.jsp?id=1361216 [Accessed June 9, 2011]. Kerin, R., A., Harvey, M., G. and Rothe, J., T., 1978. Cannibalism and new product development. Business Horizons, 5(21), 25–31. Lehmann, S., and Buxmann, P., 2009. Pricing strategies of software vendors. Business & Information Systems Engineering, 1(6), 452–462. McGrath, M., E., 2001. Product strategy for high technology companies: accelerating your business to web speed. McGraw-Hill. Newman, J., W., 1967. Marketing management and information: a new case approach. R. D. Irwin. North, M., J., Howe, T., R., Collier, N., T., and Vos, J., R., 2007. A Declarative Model Assembly Infrastructure for Verification and Validation. Advancing Social Simulation: The First World Congress (pp. 129-140). North, M., J., and Macal, C., M., 2007. Managing business complexity: discovering strategic solutions with agent-based modeling and simulation. Oxford University Press, USA. Shalom, N., and Perry, G., 2008. Economies of NonScale. Nati Shalom’s Blog. Available at: http://natishalom.typepad.com/nati_shaloms_blog/ 2008/06/economies-of-no.html [Accessed June 26, 2011]. SIIA, 2001. Software as a service: Strategic backgrounder (Vol. 21). Retrieved from http://www.siia.net/estore/ssb-01.pdf. Tesfatsion, L., 2002. Agent-based computational economics: growing economies from the bottom up. Artificial life, 8(1), 55-82. Wainewright, P., 2009. Ariba’s Journey to Software as a Service - The Connected Web. ebiz. Available at: http://www.ebizq.net/blogs/connectedweb [Accessed June 17, 2011]. Warfield, B., 2007. Interview: Concur’s CEO Steve Singh Speaks Out On SaaS/On-Demand. SmoothSpan Blog. Available at: http://smoothspan.wordpress.com [Accessed June 17, 2011].

Scenario HG1 Scenario HG2 Scenario LG2 Scenario LG1

3500000 3000000 2500000 2000000 1500000 1000000 500000 0

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Incumbent Contention Rate

Figure 7: Incumbent’s Total Profit in the Identified Market Scenarios (see Table 2; Average for 10 replications) 5.

CONCLUSION

This work showed the multi-faceted interdependence of cannibalization and scalability in determining the success of a SaaS strategy pursued by an on-premises vendor, either in a monopolistic position or as an incumbent challenged by a SaaS competitor. Given the lower margins of a SaaS offering, the monopolist prefers to avoid cannibalization by limiting scale, unless the achievable market expansion proves substantial. In the presence of a SaaS challenger, instead, revenue cannibalization may be for the incumbent a necessary evil whereby customers are retained against the threat of competitive draw. Scalability represents then a key requirement for the incumbent to ride SaaS adoption, cannibalize and possibly expand the market. These findings were obtained by going after specific strategic interdependences in simplified market scenarios. The modelled market landscape and competitive dynamics should be extended to get a more realistic and comprehensive picture of the trends currently affecting the software industry. Moreover, a thorough validation of the experimental outcomes based on empirical market data ought to be performed. ACKNOWLEDGMENTS This work was partially supported by the German Federal Ministry of Education and Research (BMBF) under promotional references 01IA08003A (Project PREMIUM|Services)

604

SECURITY IN SENDING AND STORAGE OF PETRI NETS BY SIGNING AND ENCRIPTION Íñigo León Samaniego(a), Mercedes Pérez de la Parte(b), Eduardo Martínez Cámara(b), Juan Carlos Sáenz-Díez Muro(a) (a)

University of La Rioja. Industrial Engineering Technical School. Department of Electrical Engineering. Logroño, Spain (b) University of La Rioja. Industrial Engineering Technical School. Department of Mechanical Engineering. Logroño, Spain [email protected], [email protected], [email protected], [email protected]

ABSTRACT The aim of this paper is double. On the one hand, to provide a standard way to hide all or part of a Petri net that could contain sensitive information, such as a company that represents a secret production process through Petri nets (privacy). On the other hand also as standard ensure that Petri net has not been altered (integrity) and that who sends or firm that Petri net is who he say he is (non-repudiation). To ensure the privacy of an entire Petri net (or a part of it) the best solution is not to prevent access to such information, such as hiding in a safe or behind a firewall, but encrypt that information, even being to view. Today it is easier to open a safe or circumvent a firewall than to break an encryption standard algorithm (which, incidentally, is impossible nowadays). As for the integrity and non-repudiation, the solution again is not to deliver the Petri net 'in hand' to avoid disruptions and to know who delivers it (since we are in the Internet age). The solution is to digitally sign all or part of a Petri net so that reliably to know who has performed the firm, and be able to detect any unauthorized modification of any of the signed data. The aim of this paper is to show how to encrypt the selected part of the graph and to sign the Petri net, so that the obtained file compliances with the desired signature and encryption. So, in this final file, all the information (and only that) referred to the shaded part is encrypted and will not be interpretable. In particular, anything will be know about the nodes p1 and p2 or transitions t1 and t3: their constitute a secret process. In addition, this file will contain additional information that will verify the integrity of the file to prevent anyone to modify and information about who has signed this Petri net. The solution we propose is to use PNML representation of Petri nets and XMLEncryption standards for encryption and for signing XMLSignature.

1.

INTRODUCTION This paper consits on the application on Petri Nets of some of the latest standard technologies used in computer security. The idea is to provide security and protection of information in data storage and sharing. In particular, we will achieve privacy, authentication, integrity and non-repudiatability data. To achieve this, we introduce some concepts such as XML, digital signature, encryption and PNML (Petri Net Marked Language). Throughout the whole paper standard Technologies are used, but, in order to implement them, in some cases it is necessary to introduce a transformation to the data (without loss of information). 1.1. Privacy This term is related with the prevention of unauthorized access to information. The solution is not to prevent physically access to such information, eg in a safe hiding or behind a firewall, but to encrypt the information. Nowadays it is easier to open a safe or to circumvent a firewall than to break an encryption standard algorithm (which today is impossible). 1.2. Integrity The integrity of the data will be obtained if we can avoid or at least detect unauthorized modification of information. 1.3. Authentication Authentication ensures that people assuring that they say or sign the data, are actually who they say they are. This avoids receiving data from a person posing as another. 1.4. Non-repudiatability It will be obtaided if we can prevent anyone saying that has not sent or modify something done. It should be possible to assure that a preson have done something.

Keywords: Petri nets, Encryption, Digital signature, Privacy, Integrity, Authentication, non-repudiatability

The solution for authentication, integrity and non repudiatability fails to deliver the information 'in hand'

605

to avoid disruptions and to know who gives it, as we are in the era of Internet and technology. The solution is then to digitally sign all or part of the data so that we know who has made the signing of a reliable and be able to detect unauthorized modification of any of the signed data.

and furdermore it may not have been another (nonrepudiation). Let us suppose we have the following Petri net.

1.5. XML XML is a metalanguage for defining other languages. XML is not really a particular language, but a way of defining languages for different needs. XML is also a standard way to exchange structured information. It is based on distributed hierarchical labels containing data. The XML files are text files. The work is based on this format for implementing information security. 1.6. PNML Marked Petri Net Language (PNML) is an XML language designed to represent Petri nets. With this language a Petri net can be stored in a text file (XML), without loss of information.

Figure 1: Petri net with a part that want to be hidden It will be proposed how to encrypt the selected part of the graph, and to sign the Petri net, so that the obtained file meets the desired signature and encryption. So in this final file, all the shaded information (and only that) is encrypted and will not be interpretable. In particular, we will not know anything about the node p2 or transitions t1 and t3: it is a secret process. In addition, this file will contain additional data that will verify the integrity of the file to prevent that anyone modify it, as well as data about who has signed this Petri net.

1.7. Digital certificate A digital certificate is a digital file non-transferable and non-modifiable generated by a trusted third party called Certificate Authority (CA), that associates a public key to a person or entity. For a certificate to perform its tasks need to use a private key that only the owner possesses.

3.

1.8. Digital signature It is equivalent to the conventional signature. It is an addition to the document you signed and indicates that it agrees with what is said in it. The digital signature provides authentication features, integrity, and non repudiation. Computationally speaking, it is a process thst transforms the original message using the private key, and anyone with the signer's public key can verify this.

3.1. XMLEncryption Encryption is a standard of XML files. It can be used symmetric or asymmetric encryption, but in this case, it is preferable to use symmetric encryption, because it is computationally less demanding. The idea behind this encryption is to replace the XML elements that want to be encrypted by another piece of XML that contains encrypted information and data about the algorithms used for encryption. When a file non-XML is encrypted, the only option is to encrypt it completely. When we apply this technology to XML, it permits to define specific fragments of the document that want to be encrypted or even to transform the document before applying encryption. Whatever the origin of data, the result is always an XML element. Typically, this document has all the information needed to be deciphered. The information that can be found is: • Encryption algorithm: is the name of a method for encrypting information. It may not be included, being necessary to be know by both the part that encriptes the file and the part that decryptes it. • Encrypted information: this part must always be present. • Name of the password used: it is optional. It is used when there is a set of keys, and have to be also

1.9. Encrypt and Decrypt Encryption is the process to convertí in unreadable some information considered as important. Decoding is the reverse: from the encrypted content becomes legible original content. Keys are used to encrypt and decrypt. In the case that the key to encrypt and decrypt is the same, it is called symmetric encryption. If encryption is made with a key but decryption is made with a different key, it is called asymmetric encryption. 2.

TECHNOLOGIES

APPROACH

The goal of this work is to hide all or part of a Petri net that may contain sensitive information, such as a company that represents a secret production process through Petri nets (privacy). On the other hand, another goal is to ensure with standard resources that a Petri net has not been altered (integrity), and that the sender or firmer of the Petri net is who sais to be (authentication)

606

known by both the part that encriptes the file and the part that decryptes it.. • Encrypted password: it is optional. The part that encrypts the document must have a public or a private key. With this key it can encrypt the password used to encrypt the content. The part that decrypts the document must have the other key. Below is an example of XMLEncryption. This is the original document: John Smith 4019 2445 0277 5567 Example Bank 04/02

Figure 2: Original document with XMLEncryption This is the document after encrypt the credit card (Figure 3). In this example we have only the encrypted information and we have no information about the key or the encryption algorithm. John Smith A23B45C56 Example Bank 04/02

Figure 4: XML obtained after applying XMLSignature A signature as must contain, accordingly with XML Signature: • Canonicalization method: Two XML documents are equivalent if they represent the same information. A method of canonicalization transforms an XML document into another equivalent one. All XML documents equivalent, since they are canonicalized using the same method, result in the same XML. It is applyed before signing. If a method is not specified, one of them is assigned by default. • Reference: There may be several references within a single firm. In each reference the part of the document that is signed and the hash algorithm used are indicated. A summary algorithm generates a sequence of bytes of fixed length from contents of arbitrary length. This sequence of bytes is different for each content. • Information on the key signature can optionally include the data necessary for validation. This part can indicate the public key directly, through a sequence of characters that identifies it or through a URL. Additionally it can also have more information about who has signed it: name, organization, country...

Figure 3: Encrypted document with XMLEncryption 3.2. XMLSignature It is a standard of digital signature of files, not necessarily XML files. However, the final file is always an XML document. It requires digital certificates and public and private keys for its operation. There are three alternatives: • Envelope: The result is the original XML file to which a signature element is added in the XML file itself. • Enveloping: The result is an XML file with the signature, and within it, there are the original elements of the original XML file. • Detached: The result is the original file and, separately, an XML file with the signature of that file. It really does not matter which one to use. They are simply different ways of organizing the generated signature.

607

• Transformations: It is possible that what want to be signed is not the complete document, but some information of it. With the changes you can do almost anything, from selecting only certain parts, to change the structure of XML, or to include other XML fragments. If it is not necessary to apply any transformation before signing you can skip this part. The end result of applying XML Signature is an XML element of the form shown in Figure 4.

Code: if (p1>0) then p1 0) that part from a hidden transition but become to a visible place or from a visible place to a hidden transition (0}*720000-(P{#P34>0}+P{#P26>0}+ P{#P12>0})*2880*0.3)40*(NOper1+NOper2+NOper5+NOper8+NOper9+NOper_ 10)-20000*Mach_Delay;

In order to present all this information, the following tables are presented. There are 5 experiments or optimization methods we have applied. Experiments

The first expression of this formula (P{#P37>0}) represents the throughput of the whole system given that place P37 is the place positioned just before being performed the last task (done by operator 11). This expression will represent the probability that there is more than zero tokens in place P37, and this is exactly the meaning of the throughput (considering that the maximal number of tokens in place P37 is 1 because there is a P-invariant that contains this place and also places P37 and Oper11). The amount of 720.000 corresponds to the gain that the company is having considering a mean selling price for all the windows produced of 25 Euros per unit produced and considering that there is a shift of 8 hours (28800 seconds). The next term (E{#P34}+E{#P26}+ E{#P12})*30*8*0.1) corresponds to the energy consumption term that is associated with the use of the different machines that are involved in the process. In this particular case there are three operation machines that are represented in places P34, P26 and P12. Computing the utilization of these machines during a shift of eight hours and considering the mean cost of the energy equal to30 kwh and considering a cost of energy equal to 0.1€/kwh

EXP1:TwoPhaseApproach:Temp_anneal_scaleparameter100 EXP2:TwoPhaseApproach:Temp_anneal_scaleparameter50 EXP3:TwoPhaseApproach:Temp_anneal_scaleparameter20 EXP4:TwophaseApproachwithReductionofSearchSpaceinthesecondPhase EXP5:TwophaseApproachwithTemperatureParameterinvariablesinsecondPhase

Experiment Time(Minutes) EXP1 EXP2 EXP3 EXP4 EXP5

Experiment EXP1 EXP2 EXP3 EXP4 EXP5

Simulations

2765.33 1090.33 589.7 7286 289.12

QUALITY

Profit 3505 1802 707 2351 184

271179.9 266887.5 269647.6 245302.2 252209.5

100.00% 39.43% 21.32% 50.12% 10.46%

Simulations 100.00% 51.41% 20.17% 67.08% 5.25%

Timing 100.00% 98.42% 99.43% 90.46% 93.00%

6. CONCLUSIONS AND FUTURE RESEARCH Here we have presented a set of approaches that have been applied to a real Flexible Manufacturing System where a first approach to the introduction of energy consumption information is introduced into the optimization problem adding an extra value to the optimization process and giving another solution to the companies in order to reduce the expenses associated with this energy consumption.

The next part corresponds to the cost associated with the utilization of the different operators that has been estimated in 40 Euros for each worker and for and 8 hour shift. Finally, the last part corresponds to the cost associated to the inclusion into the system of a quicker Numerical Machine Center that will increase the price according to the mean operational speed (20000 € /second)

Some possible future research topics will be related with the introduction of more energy related information into the models so that the optimization process will be more concentrated into this topic.

5. RESULTS AND COMPARISON The results we are interested to compare between the two approaches previously shown are related with productivity measures. It will be considered the number of pieces produced per time unit for each type of product (32 different types can be produced in the FMS). Another performance measure we will consider will be the utilization of the different operators that are present into the system

REFERENCES Aarts, E., Korst, J. “Simulated Annealing and Bolzmann Machines”, Wiley (1989) Ajmone Marsan, M., Balbo, G., Conte, G., Do-natelli, S., Francheschinis, G. “Modelling with Generalized Stochastic Petri Nets”, Wiley (1995) Balbo, G., Silva, M.(ed.), “Performance Models for Discrete Events Systems with Synchronisations:

624

Formalism and Analysis Techniques” (Vols. I and II), MATCH Summer School, Jaca (1998) DiCesare, F., Harhalakis, G., Proth, J.M., Silva, M., Vernadat, F.B. “Practice of Petri Nets in Manufacturing”, Chapman-Hall (1993) Ingber, L. “Adaptive simulated annealing (ASA): Lessons learned”, Journal of Control and Cybernetics, 25 (1), pp. 33–54 (1996) Rodriguez, D. “An Optimization Method for Continuous Petri net models: Application to Manufacturing Systems”. European Modeling and Simulation Symposium 2006 (EMSS 2006). Barcelona, October 2006 M. Silva. “Introducing Petri nets, In Practice of Petri Nets in Manufacturing” 1-62. Ed. Chapman&Hall. 1993 Zimmermann A., Rodríguez D., and Silva M. Ein effizientes optimierungsver-fahren fr petri netz modelle von fertigungssystemen. In Engineering kom-plexer Automatisierungssysteme EKA01, Braunschweig, Germany, April 2001. Zimmermann A., Rodríguez D., and Silva M. A two phase optimization method for petri net models of manufacturing systems. Journal of Intelligent Manufacturing, 12(5):421–432, October 2001. Zimmermann, A., Freiheit, J., German, R., Hommel, G. “Petri Net Modelling and Performability Evaluation with TimeNET 3.0”, 11th Int. Conf. on Modelling Techniques and Tools for Computer Performance Evaluation, LNCS 1786, pp. 188-202 (2000).

625

A SIMULATION-BASED CAPACITY PLANNING MODEL: A CASE STUDY IN A CONTRACT FURNISHING SME Nadia Rego Monteil(a), David del Rio Vilas(b), Diego Crespo Pereira(c), Rosa Rios Prado(d), Arturo Nieto de Almeida(e)

(a), (b), (c), (d), (e)

(a)

Integrated Group for Engineering Research (GII). University of A Coruna, Spain

[email protected], (b)[email protected], (c)[email protected], (c)[email protected], (d)[email protected]

Uncertainty is a well-known characteristic of make-to-order (MTO) philosophy based companies. It is difficult (if not impossible) to predict the time at which a customer will place an order, the order due date, its quantity and nature and accordingly, the process and material requirements to fulfil it. Generally, orders’ uncertainty affects more to SMEs because of their weaker negotiation position (Achanga, 2006). Usually, diversifying is a way to reduce the impact of such condition. For instance, civil engineering companies try to combine projects for public clients with other works for private ones. Also shipyards do the same by offering both ship repairing and maintenance activities and new shipbuilding. However, this product variability and simultaneous production with shared resources make the planning and scheduling problem even more complicated. The available effective capacity to be used during the execution depends on the operational dynamics, which in turn depends on planning decisions. This circularity in planning is more complex in unsynchronised shops where the variety of products follows diverse routings (Albey and Bilge 2011). All these factors - uncertainty in demand, product variability, simultaneous conditions and diverse routings- are present in the normal activity of this case study. As said before, the short period of time between the effective confirmation of an order and the moment the manufacturing process actually starts does not allow a proactive planning. The impact on the production levels may lead to bottleneck formations, subcontracting, difficulties to meet the due dates and expensive raw material acquisition. In other words, competitiveness relies on efficient production and capacity planning. Clearly, deterministic processing times and deterministic demand rates based models are inappropriate to consider uncertainty in models. Mula et al. (2006) have carried out a literature review of the uncertainty treatment on production planning models. The use of simulation in production planning has been considered for several authors in the literature. For instance, Albey and Bilge (2011) state that simulation

ABSTRACT The contract furnishing sector –stores, hotels and public facilities– is characterized for working under a MTO philosophy, having worldwide clients, a flexible and highly manual process and a high mix of products. In the presence of these sources of variability, the lack of a proactive planning drives to outsourcing and cost overruns. This paper presents a case study of capacity assessment in a Spanish SME of manufacturing, distribution and assembly of contract furnishing. To do so, a Monte Carlo simulation approach was adopted, with stochastic values for demand, process and product parameters. Based on historical data and expert interviews a spreadsheet-based model was proposed in order to represent the variability. As a result, capacity anticipation under different scenarios was provided: (i) at present conditions, (ii) with an increased demand and, (iii) in case of a change in the type of production orders. Keywords: Capacity Planning, Monte Carlo Simulation, Demand Forecasting, Contract Sector, SME 1. INTRODUCTION Contract furnishing sector is made up by companies that offer a complete furnishing service to hotels, stores, offices and public buildings, including design, manufacturing, distribution and final on-site assembly. These companies typically work under a Make to Order (MTO) philosophy, meaning that the manufacturing starts only after a customer's order is received. The case study is a family-owned Spanish medium enterprise. Its main client is one of the world’s largest fashion distributors representing circa 60% of the company’s production and sales. Being a SME and having powerful worldwide clients is a complex balance; failing or even delaying an order is not an option. However, projects’ planning is usually carried out in a reactive manner making subcontracting necessary in order to meet the tight deadlines.

626

is a “natural solution” to represent the complex dynamic operational behaviour of unsynchronised shops. Several works propose simulation production planning. Hung and Leachman (1996) collect flow time statistics instead of machine capacities while simulating a production plan. Byrne and Bakir (1999) solve a multi-period multi-product production planning problem by using a hybrid simulation-analytical approach. Kim and Kim (2001) update both machine capacities and flow times by collecting relevant statistics during simulation. However, Buxey (2005) points out three reasons why business has ignored researcher’s efforts so far. Those are basically related to (i) their huge need of data –impossible, difficult or expensive–, (ii) to the model underlying assumptions which make them inappropriate for the process or (iii) because they are too complex and they are seen as worthless by the managers. Moreover, empirical evidence shows that practitioners using advanced planning methods are on average less satisfied with their plans than those who use simpler and less accurate methods (Jonhson and Mattson 2003). In that sense, Tenhiälä (2011) links appropriate capacity planning techniques with process types, set-up and nature. He also highlights the wide use of non-systematic planning methods and claims that optimization is not always desirable in complex real-world planning situations so more pragmatic research in operations management is needed. Considering these circumstances, the main goal of the study was obtaining an estimate for the fitted allocation of production resources (both in time and quantity) for the following year under two plausible demand scenarios: 1. 2.

orders are given with a certain level of anticipation. On the other hand, 40 % of production serves a variety of hotels, stores and business with more flexibility in due dates, but with much more uncertainty both in its amount and nature and in its occurrence. After a failed statistical attempt of characterizing this sort of demand, it was decided not to include it within the scope of the study. Besides, due to the possibility of delay in these type of order deliveries, it has been considered that the company’s capacity estimation is mainly influenced by the rest of the “predictable” orders workload. From now on, demand and orders will refer to the main client’s one. The normal company’s operating scheme is now briefly described. The whole process starts when the fashion distributor provides the furnishing company with its own interior designs catalogue for all the stores of the year. This catalogue allows a basic preliminary estimate of materials as well as an early and conceptual technical design of the furnishing elements. From that moment on, specific orders may be placed in the form of a store plan, between 10 weeks -normal order- and 5 weeks -urgent order- before the opening date, where all the furnishings should be assembled and ready to be used. Once the order arrives, the number of pieces of furniture and the amount and type of materials is calculated by the Technical Department. Then, the material supplying and the generation of production orders (in term of parts’ groups that are manufactured together) may start. From that time, the manufacturing process takes place sequentially through four manufacturing sections. Each of them, with different machines, operators and responsible, is now described. Also a fifth section is included, although it does not take place within the production plant.

The next year presents the same tendency than the previous ones. Changes in the type of orders may occur: a. There is the possibility of establishing an important contract with a chain of resort hotels, increasing the workload around a 20%. b. Their main client may change the type of order, from the complete interior store manufacturing one to the partial refurbishing of actual stores, which would mean smaller orders (-25%)

1. Machining, comprising cutting and machining operations (drills, shapers, CNC machining centre, etc.). 2. Finishing, consisting of sanding, varnishing and automatic and manual painting operations. 3. Cabinet making, where carving operations, assembly and subassembly take place. They are the most qualified and experienced workers. 4. Packing, that can be either automatic or manual packing. All the finished packages are placed on the storage area until the whole order is ready so it can be dispatched. 5. Assembly. Finally, the pieces of furniture and the assembly workers are sent to the final location. Currently, the cabinetmakers are also in charge of these operations.

To do so, we decided to build a spreadsheet-based simulation model aiming at representing how demand implied workload on the different work centres. Variability modelling is obtained from historical databased probability distributions for demand, process and product parameters. Monte Carlo simulation is then used for the risk assessment of the solutions provided.

Some of the reasons for the complexity when applying a planning method are described now:

2. CASE STUDY Demand is composed of several components. On the one hand, 60 % of production is dedicated to their main client. Although time deadlines are strict, at least

-

627

There is a high uncertainty on demand, considering that orders are different in quantities and due dates.

-

-

1.

The manufacturing process is defined in the literature as an HMLVS, i.e. High-Mix LowVolume Simultaneous production (Prasad, 2011). Its main characteristic is that product mix and bottlenecks keep changing frequently over time. Distribution and final assembly take place out of the manufacturing plant, increasing the uncertainty due to unpredictable external events. Besides, assembly and cabinet-making compete for the same resources.

2.

3.

Number of operators, related to the direct manufacturing costs. Failure probability, related to the risk of a plan. It is calculated as the number of weeks where the required workload exceeds the available workload (given a certain capacity level). Occupation level. The average occupation level is related with the efficiency of the plan. MODEL

INPUT DATA

3. APPROACHING THE PROBLEM The followed methodology is shown in Figure 1. From historical demand data (2003-2010) we tried to determine the number and dates of the orders that could be usually expected any year, according to the past events. The same information was used for determining how project usually developed throughout the process operations. From its analysis, probability distributions are obtained in order to represent the following events: -

OUTPUT DATA

Time

Demand

Workload Aggregate Unit: Workhours

1 year- horizon 1 week – period Backward scheduling

Each month probability of orders incoming

Product

Capacity assessment

Process

Aggregate Unit: Workhours

Risk quantification: Parameters

4 Manufacturing workcenters + Assembly stage

Figure 1: Model Development 4. MODEL DEVELOPMENT Attending to the eight years record, when characterizing the occurrence of demand in terms of how likely is to have a fixed due date, a set of five month’s behaviours within a year was identified. Autocorrelation tests were performed without leading to the identification of any significant pattern. By means of Maximum Likelihood Estimation five Poisson distributions were found to model the orders fulfilment process. When one or more orders are initially expected to be fulfilled in the same month, we have considered that they all will have to be ready for the first week (so assuming a worst case). As a result of the demand estimation, at each simulation N-orders arrive at a certain delivery week (DW) The amount of manufacturing and assembly hours (H) is composed of two terms. First, the average value was obtained from historical data and verified with the responsible of the Production Department. Then, a variable noise was modelled. The distribution of total hours in each process sector was quite regular (distribution of orders in work centres is shown in Figure 2). In order to simplify the model it was considered fixed (p1, p2, p3, p4 y p5).

Monthly probability of a particular number of order deliveries. For each order: - Sizing. Distribution of the global amount of hours and its allocation among the different main operations. - Timing. Distribution of the operations durations and their respective delays.

Probability distributions are the input data of the model. Regarding the system modelling, time and process considerations have to be regarded. On the one hand, a marketing medium term approach (one year) has been proposed as a useful way of linking the managerial issues with the production requirements. Intermediate-range planning is facilitated by aggregating the many products of a company into a single unit of output (Aggregate Production Unit). In our case, each order will be translated into working hours. Also, we have considered as the planning period time the working week, because it is the usual unit in which production activities are referred within the company. On the other hand, the process has been divided in their main operations, that is to say, Machining, Finishing, Cabinetmakers, Packing and Assembly. A backward scheduling has been adopted, starting from the assembly tasks and finishing in the machining. The resulting workload -the output data- is then analysed both in its quantity (maximum and average) and attending to its distribution along the year. The workload sizing and timing will be the basis of the capacity planning. Decisions regarding capacity planning have different risk implications. For the assessment of capacity planning, three different parameters will be used:

MACH

FINI

2

4

CABI

PACK

ASSE

100% 80% 60% 40% 20% 0% 1

3

5

6

7

8

9

10 11 12

Figure 2: Monthly Distribution for the 2006 Year Orders Attending to the Defined Process Sectors Given a certain delivery week, (DW), the process scheduling is established backwards. Accordingly, the

628

variety on operations lengths was modelled. For example, most times (63% of frequency) an order machining takes place in four weeks. During these four weeks, 32% of the machining hours happen the first week, 41% the second, 15% the third and 13% the fourth. However, 37% of times it takes place in three weeks. This has been done for all the work centres of the process. In addition, production orders have to go through the whole process in a certain order attending to technological constraints. The sequential progress of the order is characterized by the following parameters: -

As a result, the number of hours per operation, week and order (K)i can be obtained: (K)i=din . M

(3)

The global amount of work, Kt is the composition of each Ki in a common time axis. The workload (Q), a 52x5 matrix, is obtained by adding the same operations each week.

x: End of packing – start of assembly delay. r1: Start of cabinetmaker work – start of packing delay. r2: Start of finishing – start of cabinet maker work delay. r3: Start of machining – start of finishing delay Figure 4: Workload Matrix Composition

The variety on delays between operations has been modelled in the same way as the variety on lengths. Graphically, the evolution of an production order is depicted in Figure 3.

In the same execution, there is a variation in terms of amount of hours, operations lengths and delays. Between two successive runs there is variation in terms of different number of incoming orders and their corresponding delivery date. So, we can talk of an inter and intra- year variation. 5.

RESULTS

5.1. Workload sizing The average workload, in hours per week, is the average of the average workload of each operation. The maximum workload is the average of the maximum values. The obtained distributions for average and maximum cabinet makers workload after 1000 simulations are shown in Figure 5. The results for the rest of departments are described on Table 1.

Figure 3. Process Evolution (for the Centres, including the Assembly). Each order has to progress along the different operations according to the evolution parameters (x, d1, r1, d2, r2, d3, r3 and d4). Those different operations take a fixed rate of the global H, (p1, p2, p3, p4 and p5), so the amount of hours of each process (din) can be obtained as a vector: din= HT (p1, p2, p3, p4, p5)

Figure 5. Average (left) and Maximum (right) Distribution of Workload for the Cabinet Department.

(1)

All the production sequence is included in a 5xD matrix (M), where the dimension depends on the final duration of the manufacturing and assembly process, as it follows: D= d0 + x+d1+ r1 + r2 + r3

Table 1. Average and maximum workload per year

(2)

Workload

Mac

Fin

Cab

Pac

Asse

Average (h/week) Max (h/week)

120.2 548.7

145.4 616.6

216.8 1597.0

110.6 449.3

294,52 3268.9

However, as a result of (i) the variability in number of hours and time evolution for every order, and (ii) the seasonable behaviour of the demand, one single value is not accurate enough to describe the expected workload. Therefore, results will be presented

Each element on M is the fraction of the corresponding production department dedicated hours (in columns) for a particular week (in rows). This way, M shows the evolution of the different operations for a single order.

629

dividing in quarters of year (Q1, Q2, Q3 and Q4) and a range of more likely values (range of 50% confidence level) are presented together with a boxplot. In Table 2, for instance, machining workload in the first thirteen weeks has been between 643 hours and 1636 hours in 500 of the 1000 simulations. As it can be noticed, second and third quarters of the year show the greater values of both workload and variability in all the departments except in assembly. More details in the year distribution of the workload will be found in section 5.2.

D.

5.2. Workload timing Yearly workload distribution (along 52 weeks) is now showed for the different departments on Figure 6. On Table 3 all operations workload distribution is described in terms of shape and Coefficient of Variation (CV), and maximum and minimum level of occupations.

Table 2. Workload’s Departments by Quarter Range Q1 643-1636

M.

Q2 977- 2148 Q3 1618-3027 Q4 493-1340

Figure 6: Year Workload Distribution for the different Departments

Q1 782-1942

F

Table 3: Operation Workload Distribution: CV, Shape, Maximum and Minimum Occupation

Q2 1148-2530 Q3 2005-3687 Q4 645-1611 Q1 1054-2891

C

Q2 1644-3868 Q3 2912-5682

Q1 1054-2891 Q2 1644-3868

Shape

Mac

1,20

Smooth

Fin

1,14

Smooth

Maximum Occupation. AugustSeptember May-June

Cab

1,64

Concentrated in short periods

Pac

1,07

Smooth

Asse

2,29

Concentrated in very short periods

One or two weeks in June, July and August End of July, beginning of August One week in August and one in September

Minimum Occupation December DecemberJanuary and April March and December End of October, beginning November March-April

The different shapes of the workload are related to the tasks nature. For instance, cabinet makers and assembly respective workload appear in a more concentrated way. Besides, these end stages of the process strongly influence the product quality. According to both factors, these work centres are considered the most critical in the process. Currently, assembly and cabinetmakers share the same labours. This condition aims at reducing the impact of the concentrated assembly works. However, it also implies that sometimes the end of the process is almost unfilled. As a result the work flow is sometimes interrupted. It has been advised to the managers to separate these work centres and to try to negotiate a resource pool (variable number of working hours that

Q3 2912-5682 Q4 776-2374 Q1 2192-4680

A

CV

May –July

Q4 776-2374

P

Op.

Q2 1466-3859 Q3 4611-8471 Q4 1027-2949

630

can be compensated with longer holiday days) with the workers. The plant total workload can be obtained by adding the different operations (machining, finishing, cabinet makers and packing) each quarter. Results are shown on Figure 7.

These graphs could be used for aiding in the decision process of establishing the capacity level. For example, the managers might decide that a 20% failure probability for the Finishing section is acceptable (one of five weeks the expected amount of work exceeds the Finishing capacity), which would imply that seven labours would be enough. As it has been stated, each department workload reaches different levels and presents different behaviours. So, different capacity levels would be needed for meeting the same requirements (in terms of failure probability). However, experience shows that it is not necessary the same confidence level in all departments. For instance, a delay in machining would be much less severe than a delay in assembly. Being machining the first department, the excess of workload would probably be transferred to the following week without more consequences, while an excess of workload in assembly would probably lead to an eventual delay in the order. Accordingly, an operation cost indicator for each department can be built as follows:

Figure 7: Plant Workload First and fourth quarter of the year usually present the lowest level of work while the third quarter reaches its top. This information is useful for setting prices and hiring policies.

‫ ݊ = ܥ‬+ 52 ȉ

஼ಷ ஼ಹ

ȉ ܲ(݂݈ܽ݅‫)݁ݎݑ‬

(1)

Being C an indicator of the cost of operating with ஼ n workers in a certain department, ಷ the relation

5.3. Capacity Estimation According to the workload on each operation, different levels of capacity can be established. In a first approach, a constant annual level of capacity will be studied. Those levels have implications in terms of probability of risk, efficiency and of course, cost (parameters introduced on section 3). Capacity range for each department will be studied between a minimum and maximum level. The adopted criteria are that the minimum/maximum level for each department corresponds with covering the average/maximum workload values (Table 1). The results are shown in Figure 8 (only for one operation). As it could be expected, the higher the number of operators is, the less the failure probability (lower blue line) and the occupation level are (red line).

஼ಹ

between the cost of a failure and the cost of hiring an extra worker. This cost increases with the number of workers and decreases with the failure probability. We could say, for instance, that a single failure in Assembly would cost 3 times more than hiring an extra worker, while a failure in Machining would only cost 0.5 times more (these values have been chosen for illustration purposes, and do not have to be close to reality). When representing these expressions, in Figure 9, the optimum number of workers (7 for machining and 16 for assembly) is obtained.

Finishing 70% 60% 50% 40% 30% 20% 10% 0% 4

5

6

7

8

P(failure)

9

10

11

12

13

14

15

Occupation

Figure 8: Operators in Machining as a Function of Failure Probability (lower bars) and Occupation Level (upper bars).

Figure 9. Machining and Assembly Cost Indicator depending on the Number of Workers.

631

5.4. Increased demand According to the commercial department’s expectations, a new contract with an important international hotel client would imply a 20% increase in the number of orders. As a result, the average workload is a 20 % higher compared to the previous situation. The maximum workload is 12-16 % higher (Figure 10), depending on the work centre. Assembly is the department that shows the higher increase of maximum workload. Referring to the CV, a smoothing effect in every task is observed. However, the lowest decrease in CV corresponds to the Assembly. It appears to be the strictest in its concentrated nature.

6. VERIFICATION AND VALIDATION For verification and validation purposes, the observed sample distribution (historical data) was compared with the modelled distributions in terms of four variables. The first is the number of orders, which accounts for demand level. The second is the total amount of workload hours per year. Yet the number of hour per quarter of year, week or department would have been a more accurate indicator for comparing results, this information was not available within the company’s historical data. The third parameter accounts for the time gap between the start of the manufacturing and the final delivery. This value is obtained from the simulation as the addition of each operation estimate represented by different probability distributions. Then, it is compared to the historical time interval between the order incoming and the final delivery. Finally, the number of hours of each order is compared with the actual values from the database. The p-values for the null hypothesis of the averages being the same and the standard deviations being different are shown in Table 4 and Table 5. Model values for the average and standard deviations come from 1000 simulations so they were considered with a negligible error. They were then compared to the available number of observations (n) in the real data.

ASSE

PACK CV

CABI

WL_Max WL_Ave

FINI

MACH -10%

-5%

0%

5%

10%

15%

20%

25%

Figure 10: Workload Distribution when Demand increases 20%

Table 4. Test T for Differences in Variable Means

5.5. Changes in the “order type” Before a plausible change in the main client’s type of works, from “complete new stores” to refurbishing existing ones, the Sales Department forecasts smaller projects (25% less, in average) but an increase in the number of orders (20%). Were this demand scenario, the workload would change as showed in Figure 11. The average workload decreases around a 10% per work centre. A similar smoothing effect in every task takes place. However, it is remarkable that the maximum workload decreases a 17% for the cabinet centers whilst the assembly only decreases a 14%. In fact, assembly has the lower decrease in maximum workload. It can be concluded that assembly department is more sensitive to changes in the number of orders than it is in the orders’ size.

Factors

WL_Ave

MACH -15%

-10%

-5%

p-value

15,1

15,28

7

0,956

Hours / year

46764,6

47867,1

7

0,903

Weeks / order

11,4

9,6

106

0,001

Hours / order

3094,6

3109,8

106

0,903

Std. Dev. Model

Std. Dev. Data

n

p-value

Orders / year

3,8

6,9

7

0,003

Hours / year

12719,9

22976,2

7

0,003

Weeks / order

5,3

5,8

106

0,103

Hours / order

835,6

1275,0

106

0,001

It can be concluded that the estimations of demand per year, workload per year and workload per order averages do not significantly differ from those observed in the historical data (Table 4). On the other hand, significant differences were found in weeks per order average. This might be explained by the aforementioned differences in time durations, but further research should be conducted in order to assess the practical relevance of such a difference. Significant differences in estimated standard deviations were found for all the tested variables (Table 5). This suggests that the model systematically

CV WL_Max

-20%

N

Orders / year

Factors

PACK

FINI

Average Data

Table 5. Test F for Differences in Variables Standard Deviations

ASSE

CABI

Average Model

0%

Figure 11: Workload Distribution when the Type of Order Changes.

632

underestimates variability levels in workload rates and, consequently, the forecasted failure probabilities. A plausible explanation is given by the way that the annual demand is generated. Although standard autocorrelation tests did not show any autocorrelation patterns in monthly orders, this is not a sufficient proof for independence. Only under actual independence among monthly demands, annual demand variability would be accurately estimated from monthly variability. Available data were not enough to conduct a more profound autocorrelation analysis. This shortcoming of the model can be corrected by adopting a more risk-averse position in the decision making process.

agregada de la producción de Impsa”, Iberoamerican Journal of Industrial Engineering, Vol. 2, No.2, pp. 90-112 Buxey G., 2005, “Aggregate planning for seasonal demand: reconciling theory with practice”, International Journal of Operations & Product, Vol. 25 No.11, pp. 1083-1100 Geneste L., Grabot B., Letouzey A., 2003, “Scheduling uncertain orders in the customer-subcontractor context”, European Journal of Operational Research, Vol. 147, pp. 297-311 Hung Y.F., Leachman, R.C., 1996, “A production planning methodology for semiconductor manufacturing based on iterative simulation and linear programming calculations”, IEEE Transactions on Semiconductor Manufacturing, Vol. 9, pp. 257–269. Jonnson P., Mattsson, S-A., 2003, “The implications of fit between planning environments and manufacturing planning and control methods” International Journal of Operation & Production Management, Vol. 73, pp. 165–173 Kim B., Kim S., 2001, “Extended model for a hybrid production planning approach” International Journal of Production Economics, Vol. 73, pp. 165–173 Mula J., Poler R., García-Sabater J.P., Lario F.C., 2006, “Models for production planning under uncertainty: A review”, International Journal of Production Economics, Vol. 103, pp. 271-285 Prasad V., 2011, “Production Scheduling for Job Shops”, White Paper of OPTISOL (Optimal Solutions for Production Scheduling), available http://www.optisol.biz/job_shop_scheduling.ht ml, 30/03/2011 Tenhiälä A., 2010, “Contingency theory of capacity planning: The link between process types and planning methods”, Journal of Operations Management, Vol. 29, pp. 65-77 Thompson S.D., Davis W.J., 1990, “An integrated Approach for Modeling Uncertainty in Aggregate Production Planning”, IEEE Transactions on Systems, Man and Cybernetics, Vol. 20, No. 5, pp. 1000-1012

7. CONCLUSIONS A workload and capacity planning based on historical data in a Spanish SME of manufacturing, distribution and assembly of contract furnishing has been presented. A simulation approach has been adopted in order to represent the high variability associated to their MTO philosophy and job-shop production schema. The spreadsheet-based model within a Monte Carlo simulation approach allows introducing stochastic values for demand, process and product parameters. As a result, workload estimation under different scenarios was provided. Also, by means of a set of three general performance parameters – labour costs, failure probability and occupation level- the assessment of the production resources necessary to cope with the corresponding workload is achieved. When complemented with overall cost information, this planning methodology can be the basis for optimised capacity estimation according to the nature of each department. This work aims at connecting the operational level with strategic considerations by means of a simple but comprehensive and precise tool for decision making. REFERENCES Albey E., Bilge U., 2011, “A hierarchical approach to FMS planning and control with simulationbased capacity anticipation.” International Journal of Production Research, Vol. 49, pp. 3319-3342. Achanga P., Shehab E., Roy R., Nelder G., 2007, “Critical success factors for lean implementation within SMEs”, Journal of Manufacturing Technology Management, Vol. 17, No. 4, pp. 460-471 Byrne M.D., Bakir M.A., 1999, “Production planning using a hybrid simulation-analytical approach”, International Journal of Production Economics, Vol. 59, pp. 305–311 Bitran G., Yanesse, H., 1984, “Deterministic approximation to stochastic production problems”, Operations Research, Vol. 32, pp. 999-1018 Boiteux O., Forradella R., Palma R., Guiñazo H., 2010, “Modelo matemático par la planificación

AUTHORS BIOGRAPHY Nadia Rego Monteil obtained her MSc in Industrial Engineering in 2010. She works as a research engineer at the Integrated Group for Engineering Research (GII) of the University of A Coruna (UDC), where she is also studying for a PhD. Her areas of major interest are in the fields of Ergonomics, Process Optimization and Production Planning. David del Rio Vilas holds an MSc in Industrial Engineering and has been studying for a PhD since 2007. He is Adjunct Professor of the Department of Economic Analysis and Company Management of the UDC. He has been working in the GII of the UDC as a research engineer since 2007. Since 2010 he works as a

633

R&D Coordinator for two different privately held companies in the Civil Engineering sector. He is mainly involved in R&D projects development related to industrial and logistical processes optimization. Diego Crespo Pereira holds an MSc in Industrial Engineering and he is currently studying for a PhD. He is Assistant Professor of the Department of Economic Analysis and Company Management of the UDC. He also works in the GII of the UDC as a research engineer since 2008. He is mainly involved in the development of R&D projects related to industrial and logistical processes optimization. He also has developed projects in the field of human factors affecting manufacturing processes. Rosa Rios Prado works as a research engineer in the GII of the UDC since 2009. She holds an MSc in Industrial Engineering and now she is studying for a PhD. She has previous professional experience as an Industrial Engineer in an installations engineering company. She is mainly devoted to the development of transportation and logistical models for the assessment of multimodal networks and infrastructures. Arturo Nieto de Almeida received his PhD in Economics from the UDC in 2010. He has been Associate Professor of the Department of Economic Analysis and Company Management of the UDC during the last 18 years. He also owns a management and technical consulting firm in A Coruna (Spain).

634

PN AS A TOOL FOR INNOVATION IN INDUSTRY: A REVIEW Jesús Fernández de Miguel(a), Julio Blanco Fernández(b), Mercedes Pérez de la Parte(b) (a)

(b)

Grupo ECO3G, Logroño, La Rioja, Spain University of La Rioja. Industrial Engineering Technical School. Department of Mechanical Engineering. Logroño, Spain (a)

[email protected], (b) [email protected] [email protected]

ABSTRACT PNs are a basic tool widely used in scientific and technical fields, in areas as diverse as automation, computer science, management, modelling, simulation, optimization, etc. Many important research groups around the world work both in the progress of the PN in themselves (behaviour, properties, analysis, etc.) and its application fields of science, industry, or services. However, such an important tool for industrial development experiences a lack of exploitation in the productive sector, compared with the potential applicability of Petri nets in industry and services management. In this paper, we review the actual use of the PN in industry, especially in the high Ebro Valley, in Spain, where ECO3G company is dedicated to the management of innovation in the industry for many years and therefore it has information based on experience. The review is also extrapolated to the world, based on information in the scientific and technical literature, and in the mentioned ECO3G experience in innovation, and national and European research projects. It also presents special attention to the use of PN in patents and industrial property as well as the know-how in companies of different industrial sectors. Finally, an analysis of the cause of this lack of exploitation in the productive sector, and a study of ways to improve this current state, are developed.

and easily model by graphic systems, and secondly to analyze the properties of the model through mathematical and computational techniques - The richness of existing modelling formalisms in the paradigm of the PN - The richness of knowledge generated in previous research works on behaviour, properties, and techniques of analysis or simulation of PN. From the foregoing it would appear that the PN constitute a tool fully integrated in the industry due to the benefits that could be obtained in productive systems, and some service companies, with their use. However its use in industry and services is much lower than expected, given its potential as a source of innovation. In this paper, we review the actual use of the PN in industry, especially in the high Ebro Valley, in Spain, where ECO3G company is dedicated to the management of innovation in the industry for many years and therefore it has information based on experience. The review is also extrapolated to the world, based on information in the scientific and technical literature, and in the mentioned ECO3G experience in innovation, and national and European research projects. It also presents special attention to the use of PN in patents and industrial property as well as the know-how in companies of different industrial sectors. Finally, an analysis of the cause of this lack of exploitation in the productive sector, and a study of ways to improve this current state, are developed. All the analysis is developed in a general way, and also applied to the most important industrial sectors of La Rioja (the Country of the authors in Spain).

Keywords: workstation design, work measurement, ergonomics, decision support system 1. INTRODUCTION There are certain characteristics of the PN that make them especially suitable for modelling discrete systems, among which are: - Easy representation of concurrent systems (with parallel evolutions and synchronizations) - The ability to condensation in a simple model of an underlying state space that suffers the state explosion problem that makes it impossible in practice the exhaustive analysis of actual cases by capacity problems of computational effort. - The duality of graphic/mathematical representation, which allows on one hand to intuitively

2.

POTENTIAL APPLICABILITY OF PN IN THE AREA (LA RIOJA) In this first phase of study we conducted an analysis of applications made in 5 sectors that clearly characterize the industry of La Rioja (Spain) and make it recognizable both nationally and in some cases internationally. So, these sectors have been chosen for its economic and strategic importance within the region, seeking to evaluate in this article their degree of

635

maturity for a potential introduction of the PN in their productive activity. Thus, as will be founded later, we have chosen the following sectors: • Wine Production: the most recognizable industrial sectpor in the region through its designation of Origin Denomination and greater international presence through exports. It is also one of the most proactive R&D areas in recent years. • Auxiliar Automotive industry: which encompasses all techniques characteristic of the automotive production; in La Rioja they are represented mainly in the rubber industry, metal, and chemical auxiliary. Their presence in the value chain is significant nationally through its connection to some of the major automotive plants in Spain and Europe. • Footwear Sector: La Rioja, and more specifically its West region (Arnedo), are the basis of the footwear producer industrial district of Spain, and one of the top 5 in Europe, even being a leader in subsectors such as safety shoes. The growth of investment in R&D in this sector is significant. • Sector snuff: Imperial Tobacco has a plant in La Rioja that has nearly 2000 employees (representing a significant 1.5% of the population of the Community according to the latest data) and is now the largest company in Spain of the group and one of the most prominent in Europe. In recent years, Imperial Tobacco, earlier known as Altadis, has been gathering at its plant in La Rioja many production plants in Spain, gradually increasing investment and providing greater strategic importance to this factory. The special feature of this product makes that it is included in the food industry.

resource for the evaluation of alternatives after the research phase of R&D projects, avoiding costly processes of reanalysis and correction of the prototypes, with the consequent consumption of resources. In other sectors, such as footwear, Carpanzano et al. (2004) posed the NP as a complement to the development of a modular production that is controlled by flexible production systems (RMS) integrated directly into the production line. Finally, sectors such as furniture had similar proposals such as Gradisar and Music (2007) where, through tools such as MATLAB, proposed the definition of an algorithm to determine the most appropriate control strategy for a given environment and productive process. Thus, through these representative examples we can see how the transfer of PN counts with relevant background and case studies that allow us to appreciate the competitive advantage that they provide to the industry. However, in the last years, it can be determined that there is a clear evolution towards a more experimental research and somewhere, more focused on sectors with higher technological capability and maturity in the production side, where the most representative example is the automotive. Influenced by Japanese production techniques arising under the Automobile Industry (JIT-Kanban, 5S, LEAN ...) the PN found in this sector increased responsiveness and a route application much more straightforward than in traditional sectors, as is highlighted in Miao and Xu (2009). References in this field are numerous in recent years (2009-2011), from the most theoretical ones, such as Zhang et al. (2010), which seeks to define the technique of detection and evaluation of the critical points of contradiction as the key to improving the definition of PN based on classic models in engineering solutions, to the most applied, such as Wang (2009), which works in the application of fuzzy PN in the production chain, especially in technical delivery, targeting a logistics guidance that subsequently has been consolidated. In parallel, there are numerous research groups working in the development of PN powerful techniques of a more global approach, and from which can be implemented oriented tools for various industrial sectors, almost always with a productive approach. Representative examples of this more general research are Han et al. (2009) or Li et al. (2010) and Wang (2009), and even more and Chen Xiao (2010), applying DSM mathematical techniques, or Xu et al. (2010), who posed a transfer of know-how to the field of logistics, which currently represents one of the busiest lines in applied research in PN, especially in seaports, airports, docks or logistics centers of activity. Finally, in this overview is intended to highlight the link between PN with applied R&D projects in industry, a trend that has manifested itself more strongly nowadays. Bartz (2010) proposed an improvement in the management of the information accumulated in the development of R&D project exemplified in the case of the automobile. Especially interesting is the

3.

OVERVIEW OF THE APPLICATION OF PN Research in PN is now at the point of greatest activity in the world, although it could stand a greater intensity of research groups in Asia, perhaps encouraged by a strong tradition in the introduction of new production techniques that improve productivity. However, we determined that since the 90's there were some references focusing on the applicability, especially as a complement to the development of R&D activities in the productive environment. For example, in (Japan-USA, 1992) already pointed to the PN as a particularly useful tool in addition to production techniques such as mechatronics, microrobotics or CAD-CAM techniques. Surprisingly, the research on PN was originally closely tied to certain sectors of technics to which they intended to supplement, especially through a very focused approach to automate certain tasks or at least integrating them. So in 2001, works such as Zhan and Luo (2001) already pointed to the usefulness of PN in highly automated industrial environments, such as the snuff, proposing a PN modeling that would allow assessing the productivity of selected alternative processes, mainly in packaging lines. Jeng et al. (2004) seeking the application of PN in the development of new types of semiconductors in the electronics industry. It is particularly interesting the approach that the work gives to the use of PN as a

636

contribution of the work in the search for synergies through the application of techniques WfMC (Workflow Manager Coalition). This also allows a more precise analysis of the different possibilities around a project with a more flexible decision-making and effective management emerged from the data acquired during the analysis of the production process. Currently the firm cooperation in R&D in both national and European projects is essential for peak performance in their development. Applied and reference examples that illustrate the effectiveness of PN in research and development is the case of China's research project about optical LAMOST (Modeling of control system based on LAMOST for Petri net workflow, X. Lu) or Costa et al. (2010) which highlight some of the PN modeling tools more useful for decision making in a project, supported by graphical and multilanguage code editors.

biochemical pathway models using clinical data (WO02099569 (A2)) which shows the usefulness of PN in the evaluation methods or fields of biochemistry and method for System abnormal event detection of continuous operation in the Industrial Processes, WO2006031635 (A2) of 2006 already laid the foundations for industrial operation from the use of PN. In addition, other references as traces from Constructing Petri Nets for diagnostics, US2008320437 (A1), or Method to Improve Unfolding in Petri Nets, US2009172013 (A1) represent the results of experimental research lines and emerging result of the search for a deeper knowledge in the field of methodology and definition of the PN. 5.

METODOLOGY The introduction of PN in R&D projects of La Rioja should be closely linked to a flexible methodology that allows business to appreciate their added value without consuming a large amount of resources. Keep in mind that with few exceptions, such as Imperial Tobacco, we speak of SMEs with a very short experience in terms of R&D and that assess the return on investment in these technologies through the results of billing partners. Similarly, being traditional sectors, a methodology of integration of PN into the Business activities should be seeked more than a methodology of implementation, looking for an easy transition. This will need to define an integration strategy. This strategy, through the University of La Rioja, the Administration or the social agents involved in research should seek to organize the integration initiatives to support more effectively the strategies of both lines of business and technology in the companies. That is, to seek the development of the technology without consuming resources that could make companies perceive a threat to their business activity. Thus, in a first phase of diagnosis, we define the response of each company or sector to some fundamental questions: • How can PN provide greater business value through the integration of applications or their improvements? • Where do we start? What is the critical technical area that needs to be revised or which is the one that the company intended for R&D? • In what order should be integrated processes and applications that the company can carry out easily? • How to take advantage of the investments? • Can the company receive aids from the Administration to facilitate the transition? • How to ensure that investment in integration is maintained over time? • How to organize the integration effort? Does the company have staff trained to understand or at least apply the methodology defined by the PN? • How to ensure that they are aligned with the defined strategy? Does the company have ability or experience in cooperation other companies, Universities, or agents of innovation?

4.

PETRI NETS AND INDUSTRIAL PROPERTY. PATENTS A search of the most important patents around the PN has been developed as a basis for this work. It is difficult to find patents around the PN as advances are usually framed mainly in the field of Intellectual Property through publications such as those discussed above. Still, the application of PN in certain production processes means that there are interesting references to cases of successful applied research in PN. The results are essentially codes G06 (computers, calculators and counting) and especially G05 (Control systems or automatic control in general; functional elements, monitoring or testing devices or elements). The geographic distribution of these patents, as in the case of articles presents a greater intensity in the Asian region, with significant references to leading companies in sectors as disparate MICROSOFT CORP (Constructing Petri Nets from traces for diagnostics, US2008320437 (A1 )), SCHNEIDER ELECTRIC Automotion (Method for orchestrating services of a service-oriented orchestration and automation system machine, US2010292810 (A1)), SAMSUNG (Configuring learning petri nets, EP1335320 (A2)) and even EXXONMOBIL (System and method for abnormal event detection of continuous operation in the Industrial Processes, WO2006031635 (A2)) that have patents applied directly to their main lines of work which are closely linked to the most cutting edge in the industry today. Such records allow us to appreciate that macroindustrial level, the PN has not only proven its effectiveness, but the companies that have chosen to integrate them have ended up reaching patents reflect the success and utility of PN as a tool . Within the references can be found in the attached document include, for direct application to fields of industry Virtual Production Control System and Method and Computer Program Product thereof (US2011040596 (A1) is a reference to recent virtual control system a manufacturing industry, Information processing method for Evaluating and using

637

• Is it necesary prior training? • Is the company capable of perceiving the received added value through the use of PN to deal with future phases of the research and the innovation? Once defined the particular scope of the proposal in each sector through the diagnosis, the development of a PN model is the next step, in companies with capacity to develope it, prioritized by the needs and the benefits of a detailed description of the behaviour and the knoledge of how key initiatives could be undertaken for research proposals to facilitate the evolution of the maturity level of integration required to support the proposed initiatives. At this point we should also assess the investment capacity of the company before moving to the next phases. Not surprisingly, obtaining a model would constitute for 90% of the companies a significant progress in its strategic capacity of production management. Later, in a more advanced stage of the research, the following stages could be to perform simulations, analysis, implementation, and a final phase of the optimization of the process studied under the PN model in order to obtain the final diagnosis. All these phases depend on the outcome of further analysis of individual sectors, and particularly their degree of maturity and the awareness that the industrial structure of La Rioja and the Government have towards this type of research initiatives. In summary, the methodology of strategic development for the diagnosis-modelling phase consists of three levels of information acquisition, which are consolidated to yield an overview of the baseline, the state of the art and skills of each company or sector, the gap between the current models and the models obtained by PN, and a plan of initiatives to reduce this gap . This process can be seen in Figure 1. Data collection and the diagnosis is made through three fundamental points: • Current business and productive strategy: must be done with site visits to each company or industry in order to define the vision and business systems. Current ways of business, production model, systems and technology, experience in R&D issues and business needs; it is done with sessions of interviews to the various areas involved and their managers • Architecture or state of applications and technology: based on questionnaires given to the responsible for application and production areas. Formats should be simple, very protected, allowing us to capture data from the production or R&D project for further analysis as accurately as possible. • Development of diagnosis: With the information gathered, the analysis of the current situation in each company or sector will be developed, where the technical maturity will be evaluated, as well as the sensitivity to the integration of new technology, its perception, or the elements of methodology of integration previously used by the organization that may be useful, as well as the objectives of subsequent phases.

Figure 1: Methodology of analysis of PN implementation in companies of La Rioja Finally, it would be very interesting to develop a deliverable for Business named "Strategy and foundations of a PN model for profuction management", which quantifies technical gap, the technical means available and necessary, and a series of proposals reflected in a starting "road map" that the companies should take to achieve the successful implementation of the methodology of pre-PN phase of modeling, simulation and analysis. Additionally a "cash flow" or plan of ROI showing the added value provided globally by the PN process, including potential or repayable financial support of the Administration, Could complement the document.

6.

CONCLUSIONS AND MAIN RESULTS. Accordingly to this preliminar analisys, industry from La Rioja is based primarily on the so-called "traditional" sectors, especially if we value that the production of wine, shoes and furniture, account for over 50% of economic activity in La Rioja. Additionally, these companies have a strategic organizational model and very traditional: the transition from the first generation of many SMEs to the second generation, better trained and adapted to the new methodologies, is being much slower than in other regions, mainly because the good economic figures before the crisis occasioned that the crisis affected La Rioja later than the other regions. Therefore, in order to deal with projects like the proposed around the PN, the region has a first line of strategic-cultural handicap because R&D has been seen by many of the companies more as a complement (or even something unnecessary) than to normal and neccesary activity. This has led many companies to face later incorporation into the new strategic models, despite the existence of financial and human resources provided by the authorities that have facilitated and facilitate nowadays the transition. Therefore, in general, we can say that the industry of La Rioja lacks mature enough to tackle a research project of optimization of the production based on PN (ranging from modeling to optimization). It is therefore

638

necessary to address the first phase of awareness the support of the Administration and later to deal modelling projects on companies that presents a sufficient level of maturity through specific programs related to projects or implementations of projects of R&D, as a supplement to for example the implementation of nacional regulations UNE 166000 (Management of R&D and innovation). So a series of conclusions and general recommendations (as well as some more specific ones) are made in this work, for the analyced sectors, in order to promote the integration of PN techniques in the industry of La Rioja.

global impact instead of only local (without particular solutions). • As seen below, support for research in some of the most characteristic sectors should increase, like wine, which has found an interesting niche development that should be addressed as a way of obtaining value for the sector and by extension for La Rioja. We now proceed to discuss in Table 1 specific proposals for each of the sectors analyzed in. It is noted that there are some priority sectors of application, such as wine or snuff, if what is sought is that research has a high degree of novelty. However, if the focus is on modernizing the production model in La Rioja, the focus should be more towards furniture and footwear as most representative sectors.

6.1. General recommendations The industry of La Rioja must become sensitized in the application of innovative techniques and the tools provided by the PN for the implementation of R&D projects. In order to achieve greater maturity, it would be interesting to carry out some of these initiatives: •Campaign awareness pre-production enhancement techniques. Disclosure of them, especially the PN through specific workshops or presentations, in situ, in an informative and forthcoming way. • To adopt systems and formats that fit the business profile of La Rioja industry. If the initial maturity is low, we should perform a simple and close approximation to the strategic management of the companies that show, especially through case studies similar to those described above, the interest of the PN for modelling, simulation and optimization. Approaches would also be interesting in order to perform return on investment that the company should carry out in PN, demonstrating their added value. • To seek support from national and regional administrations and to work in cooperation with companies and industry organizations for the development of PN. Bringing the triple helix model (Industry, University, Government) to companies in a tangible and easily way. Given that La Rioja has a University and an active Development Agency, it would be interesting to carry out a series of audits on industrial organization and R&D in which the balance of the interest of the inclusion of PN within the Rioja business network could be determined. • That such cooperation includes, where possible, tools of financial support that at least allow companies to assess their baseline pre-PN modeling with a low cost. This would be particularly interesting to search for knowledge transfer from the incorporation of technologists from the University of La Rioja, because of its proximity and experience to businesses. We also consider that such support should take precedence collaboration between companies, the aforementioned incorporation of technologists, and that the studies should be addressed as a whole (or cluster) in order to optimize resources and achieve that the PN have a

REFERENCES 1992 Japan - USA Symposium on Flexible Automation Advances in Materials Manufacturing Science and Technology XIII: Modern Design Theory and Methodology, MEMS and Nanotechnology, Material Science and Technology in Manufacturing, Materials Science Forum, Volume 628 629, 2009, 734p Bartz, R. 2010 Contribution to a workflow-based information management in automotive testing and data análisis, Proceedings of the IEEE International Conference on Industrial Technology, 2010, Article number 5472558, Pages 1026-1031 Carpanzano, E.a , Cataldo, A.a , Tilbury, D.b 2004 Structured design of reconfigurable logic control functions through sequential functional charts, Proceedings of the American Control Conference, Volume 5, 2004, Pages 4467-4471 Chen, J.a , Zhang, L.-W.b , Luo, J.-Q.b 2009 Study on reconfiguration cost modelling of Reconfigurable Manufacturing System IET Conference Publications Volume 2009, Issue 556 CP, 2009 Cicirelli, F., Furfaro, A., Nigro, L. 2010 A servicebased architecture for dynamically reconfigurable workflows, Journal of Systems and Software, Volume 83, Issue 7, July 2010, Pages 1148-1164 Cicirelli, F., Furfaro, A., Nigro, L. 2010 Using time stream Petri Nets over a service architecture for workflow modelling and enactment Spring Simulation Multiconference 2010, SpringSim'10, 2010, Article number 131 Costa, A.a c , Gomes, L.a c , Barros, J.P.b c c , Oliveira, J.a , Reis, T.a 2010 Petri nets tools framework supporting FPGA-based controller implementations , Proceedings - 34th Annual Conference of the IEEE Industrial Electronics Society, IECON 2008, 2008, Article number 4758345, Pages 2477-2482

639

SECTOR

STRATEGIC MATURITY

Wine Sector

Medium-Low. Big companies are addapted to the new environment but there are plenty of small traditional wineries

Auxiliar Automoti ve Sector

Medium-High. It is a sector heavily influenced by the groups to which they provide, which are generally part of the same group. Thus many companies already assume management techniques from its parent companies. In any case it is a sector that traditionally has bet, also in La Rioja, by advanced management techniques. Medium-Low. The sector in La Rioja presents a significant gap between large firms with capacity and experience in strategic management and other family companies in a intergenerational transition period.

Footwear Sector

MATURITY IN R&D

APPLICABILITY OF PETRI NETS

ADDED VALUE OF PETRI NETS

Medium. There are R&D projects mainly related to big companies. Currently there is a tendency to increase investment in this area, although there are still many companies that do not have wine experts or technicians and R&D departments. Medium-High. There are R&D projects for years and many companies have their R&D to continuously generate ideas. However a need exists for greater cooperation between companies although there are already ongoing corrective measures.

Very high. There are hardly any references. The traditional origin of production processes has hindered the technology transfer.

It is recommended the bid for product differentiation, using the PN to study the impact in production and distribution of new wines based on R&D from new varieties.

Medium. In this case the degree of novelty of the application of PN is quite low. Locally, it may be considered a productive innovation, although it PN integration would consist, in the first phase, mainly in integrating solutions or models already implemented successfully.

PN are recommended to be used as support tool for optimization and improvement of productive activety. Also as supplement in the R&D carried out in this sector.

Medium. In the last 5 years, companies have tried, more or less, to develop, internasly or with their suppliers, improvements to compete with the Asian market. In addition, the presence of the Technology Centre in the environment is a basic support.

Medium-High. PN can be considered a new and important support in decision-making in R&D, especially in the case of safety footwear.

In this case it is proposed that PN are used to support the transition from classical to a production model based on the R&D through a gradual adjustment: NP are proposed as a tool for analysis and development of projects. In this case the contribution that can bring Petri nets is much lower than in other sectors less mature. That is why we must seek that PN add value in any area of special interest, highlighting the possible improvement in product distribution and logistics. The Rioja furniture companies need to vary its traditional production model to proprietary products, characterized by modularity, flexibility of production and innovation. All these values can be provided through PN models.

Snuff Sector

Very high. Imperial Tobacco is a multinational group that applies the most modern production techniques and R&D results both to productive and strategic management.

Very high. It has different areas of R&D and is continually developing new projects to adapt to legislation and consumer, or as a strategy of differentiation.

High. While there are solutions in the industry previously, Imperial Tobacco has both the capacity and interest enough to develop projects based on PN from modeling to optimization. Perhaps this is the most suitable plant for this in La Rioja.

Furniture Sector

Medium-Low. In La Rioja some traditional companies coexist with companies applying more productive techniques. However, the most tipical profile is a cooperative or a family company with classic strategic-functional structure.

Medium. There exists very interesting R&D projects in course, but generally developed through collaboration with suppliers or on topics of little technological interest. It is facing an evolution in R&D to provide competitive advantage

Medium. There exist previous success cases and companies that have sensitivity towards innovation. However, it is necessary a strategic and productive evolution in the short to medium term to advance from only modeling to simulation and optimization.

Table 1: Summary of analysis of PN applicability in the main industrial sectors of La Rioja Systems, Man, and Cybernetics Part A:Systems and Humans. Volume 34, Issue 1, January 2004, Pages 102-112 Kleyner, A.a , Volovoi, V.b 2010 Application of Petri nets to reliability prediction of occupant safety systems with partial detection and repair Reliability Engineering and System Safety Volume 95, Issue 6, June 2010, Pages 606-613 Li, X.-P.a c , Zhao, W.b c c , Liu, D.-X.a c c , Yuan, C.Y.a , Zhang, S.-K.b c c , Wang, L.-F.b c c , 2010, A supply chain modeling technology based on RFID discovery service , Tien Tzu Hsueh Pao/Acta Electronica Sinica Volume 38, Issue 2A, February 2010, Pages 107-116 Mendes, J.M.a , Restivo, F.a , Leitão, P.b , Colombo, A.W.c 2010 Petri net based engineering and software methodology for service-oriented

Gradišar, D.a b , Mušiþ, G.a 2007 Productionprocess modelling based on productionmanagement data: A Petri-net approac, International Journal of Computer Integrated Manufacturing, Volume 20, Issue 8, December 2007, Pages 794-810 Han, K.-H.a , Yoo, S.-K.b , Kim, B.c 2009 Integration of UML and Petri Net for the process modeling and analysis in workflow applications Proceedings of the 13th WSEAS International Conference on Computers - Held as part of the 13th WSEAS CSCC Multiconference, 2009, Pages 255-262 Jeng, M.a , Xie, X.b , Chung, S.-L.c 2004 ERCN* Merged Nets for Modeling Degraded Behavior and Parallel Processes in Semiconductor Manufacturing Systems, IEEE Transactions on

640

industrial automation IFIP Advances in Information and Communication Technology Volume 314, 2010, Pages 233-240 Miao, Z., Xu, K.-L. 2009 Research on control policy for lean production systems based on petri net, 2009 International Conference on Information Management, Innovation Management and Industrial Engineering, ICIII 2009 Volume 2, 2009, Article number 5370440, Pages 557-560 Telmoudi, A.J.a , Nabli, L.b , M'hiri, R.c 2009 Modeling method of robust control laws for manufacturing system to temporal and non temporal constraints through Petri nets, International Review on Computers and Software, Volume 4, Issue 2, March 2009, Pages 266-277 Wang, H., Dong, T., Zhang, J., Wang, H. 2010 Simulation and optimization of the camshaft production line based on Petri net Advanced Materials Research Volume 139-141, 2010, Pages 1506-1509 Wang, J. 2009 Automotive supply chain performance influencing path analysis based on fuzzy petri net , 2009 International Conference on Information Management, Innovation Management and Industrial Engineering, ICIII 2009 Volume 1, 2009, Article number 5368162, Pages 359-362 X.Lu 2010 Modeling of control system for LAMOST based on Petri net workflow, Proceedings of SPIE - The International Society for Optical Engineering, Volume 7738, 2010, Article number 77381I Xiao, R., Chen, T. 2010 Research on design structure matrix and its applications in product development and innovation: An overview, International Journal of Computer Applications in Technology Volume 37, Issue 3-4, March 2010, Pages 218-229 Xu, Y., Zhang, M., Tang, S., 2010 Research on workflow model of cooperation between 4PLs and 3PLs based on Petri-net 3rd International Conference on Information Management, Innovation Management and Industrial Engineering, ICIII 2010; Kunming; 26 November 2010 through 28 November 2010; Category number P4279; Code 83865 Yu, X.-L., Jiang, J., Xia, B.-Q., Pan, Z.-K. 2010 Petri-net-based analysis method for grid services composition model CICC-ITOE 2010 2010 International Conference on Innovative Computing and Communication, 2010 AsiaPacific Conference on Information Technology and Ocean Engineering 2010, Article number 5439240, Pages 180-184 Zhan, Y.-D., Luo, Y. 2001 Modeling and simulation research of material handling automatic system based on Petri Net , Xitong Fangzhen Xuebao/Acta Simulata Systematica Sinica, Volume 13, Issue 4, 2001, Pages 501-504 Zhang, D., Zhang, P., Jiang, P., Tan, R. 2010 Contradictions determination method in product

design using Petri net 5th IEEE International Conference on Management of Innovation and Technology, ICMIT2010; Singapore; 2 June 2010 through 5 June 2010; Category number CFP10795-ART; Code 81208 Zhao, N., Dong, S., Ding, W., Chen, L. 2010 The modularization of the tobacco distribution center simulation ICLEM 2010: Logistics for Sustained Economic Development Infrastructure, Information, Integration - Proceedings of the 2010 International Conference of Logistics Engineering and Management Volume 387, 2010, Pages 22712277

641

STOCHASTIC OPTIMIZATION OF INDUSTRIAL MAINTENANCE STRATEGIES Castellanos F.(a), Wellens A.(b) Facultad de Ingeniería – UNAM, México (a) [email protected], (b) [email protected]

Abstract 1. Introduction Since even industrial equipment of the same type tends to behave in a different way, maintenance techniques consider equipment as independent items, which mean that maintenance decisions are made for individual equipment. The common practice is that decisions are made regarding the most problematic equipment and afterwards these decisions are generalized to other equipment of the same type. This causes maintenance to be effective, although usually with very high costs and unnecessary in some equipment. Industrial equipment computers have a feature that generally is not taken into account in the determination of its reliability: the equipment may have failures that do not limit their production capacity, but increase the proportion of defective products, operation or preparation time, etc. This paper discusses a computer application designed as a decision support tool for selection of maintenance activities taking into account the risks and costs associated with choosing different maintenance strategies. Rather than searching for a solution to a problem: “what maintenance strategy would lead to the best reliability and dependability parameters of system operation”, in this approach different maintenance scenarios can be examined in “what-if” studies and their reliability and economic effects can be estimated. The proposed model represents the reality of the machines and can be successfully implemented as a maintenance management system as for example Reliability Centered Maintenance (RCM). The model makes use of stochastic process theory, specifically semiMarkov chains in continuous time (SMC). The

The dynamics of business requires industries to produce more and more at the lowest cost, highest quality and a high level of reliability (availability and reliability of their equipment) to meet the stringent technical, economical and legal requirements, and to remain efficiently and competitively in the market. The aim of this research is the development of a mathematical simulation model for the optimization of maintenance policy in a complex industrial production system, in which maintenance practices and production are integrated into a single procedure, in which control variables are the maintenance policies. This was done using a combination of maintenance management techniques and stochastic processes. This work combines an efficient methodology for maintenance management, such as RCM, with mathematical modeling, to be able to optimize the availability of a complex system. To be able to compare different maintenance strategies and the corresponding reliability, availability and economics, the model’s output variables are the system and parts availability in a given time. Two different simulation strategies were used, on one hand being the modeling of the complex system using an Excel database and the interaction of variables through constraints and likelyhood behavior, and on the other hand Matlab’s Simulink / SimEvents, feeded by equipment characteristics, behavior, the companies’ environment, needs and maintenance strategy actions. Keywords: reliability, stochastic processes, simulation

maintenance,

642

maintenance procedures may cost more money than they should or essential equipment may be unnecessarily taken out of service for prolonged periods of time. Preventive maintenance policies are either aimed at detecting deterioration of the equipment before it fails or simply are adopted on the a priori assumption that the equipment has deteriorated and requires replacement. In either case, there is a need to select the maintenance frequency and extent so that the desired objectives are achieved. As well, the operators want to maximize benefits (which would results in reduced equipment deterioration and reduced cost of equipment replacement and repairs) and to minimize the costs of the maintenance activities.

first difficulty to model the problem is its mixed discrete/continuous condition. On one hand, the equipments degree of damage is considered to be continuous in time; on the other hand the maintenance inspection system which detects the state of the team and take actions for it, is discrete in time. This makes the mathematical system to predict the total system state at time t a hybrid stochastic system (HSP) and thus quite complex. To overcome this problem, simulation will be used to optimize the system. The second difficulty with respect to the modeling of the system is that the behavior of a real maintenance system is usually unknown: due to its nature, accurate data is difficult to obtain as in case of actual failure, inspection and reparations are carried out, so future state of the equipment without maintenance cannot be known. To overcome this limitation, subjective information from maintainers is taken into account in the model data.

2.1 Life cicles The simulation model will determine the maintenance strategy that maximizes the availability of equipment; the availability is linked to the equipment’s aging and corresponding life cycle, which in turn is also linked with the maintenance activities. These life curves represent the relationship between the equipment’s technical or financial condition, and time. Since there are many uncertainties in predicting equipment life, the probabilistic analysis of failure rates should be done carefully to construct and evaluate the life curves properly. Figure 1 shows an example of a life curve and its modeled equipment status over time with different maintenance policies. The simulation results are similar to the presented ones.

1.1 Other research The simulation of maintenance models generally do not represent a complete the reality of the machines by focusing on very specific aspects of individual machines. In the aging model proposed by Jaroslaw Sugie refers to a maintenance model which does not depend on changes in the maintenance, rather focuses on the deterioration of the machine, so it does not propose changes to improve efficiency the machine, its life or reduce costs. This system proposes a new approach, the machine as part of a system which responds to the decisions arising from the maintenance management and selection policies. This model shows costs of each policy, use of the machine and economic performance in the system.

2. Modeling System The deterioration process of the equipment will, to a large extent, depend on the adopted inspection and maintenance policy. However, it is often difficult to determine with reasonable confidence just what the best frequency of equipment inspection is, or what should be inspected. As a result, some

Figure 1: Typical life cycle of industrial equipment.

643

The three examples of maintenance policies are: 1) Stop all maintenance actions. 2) Continue current maintenance policies. 3) Reduce partially existing maintenance policies after some time.

2.3 Model Explanation The simulation model consists of a machine that has a number of stages during his life. Each stage is represented as D1 to D4, each has different levels of productivity, efficiency and quality of product produced. For example, D1 represents brand new machine with 0 defects and 100% efficiency, D4 represents the machine in the last stage of their productive life with low efficiency and many defective products.

2.2 Application Scenarios Optimization of maintenance policy using the proposed tool is applicable to any company that meets the following initial conditions: • • •

The transitions between states are represented by Dn a Markov chain with a continuously variable λ which determines the values of the exponential distribution for the state changes represented as PDij. When the team arrives at D4 and changes state again becomes F means failure of the equipment.

Being a medium sized manufacturing industrial Have a maintenance department or maintenance operations The machinery must be analyzed and removable reparable

All variables in the model must be analyzed for each case study, the values that will in many cases are subjective information by the maintenance technicians.

During his lifetime the machine have a system to prevent this reaches F, this is maintenance. Each determined time inspections are performed (In) to the machine to determine the degree of wear and take decisions about whether to be left like this, send it to light maintain (Mn1)

Figure 2. Graphical representation of the proposed Markov model.

644

3 states of deterioration, depending on the skill of workers; in Sust have a high cost for the acquisition of new machinery. If the system reaches the final state of Sust, the simulation stops because it would be the life of a second machine. The graphic model system shows in Figure 2.

or intensive maintenance (Mn2). These decisions will be taken by the analysis and experience of the inspector. This value should be studied for each simulated case and will likely Pn0, Pn1 and Pn2. The inspection system is represented by a discrete markov chain imbued. The maintenance system is represented by a semi-markov chain.

The model input variables are divided into 2 types: variables and variables of computer maintenance. Parameters of the equipment are all variables that are intrinsic to the equipment, facilities, the external environment as demand and costs. Maintenance variables are variables that can simulate different maintenance strategies, these variables are: time between inspections, experience levels of inspectors and inspection fidelity.

If the machine was sent to maintenance, it going to be out from production system, this will cause delays also maintenance generates costs and a timeout to return to activities. Waiting times and maintenance costs are normally distributed with μ and σ, which are greater in Mn2 case. At the end of the holding time before maintenance, the machine will return to production at some stage Dn which will be simulated by a binomial distribution. Depending on the situation the machine and maintenance process, as result may have a minor, higher or same wear state after maintenance.

These variables fed the model and responds with results, also varying levels to design the best strategy. The model results are divided into 3 categories, each of which will have a different priority in each case: financial, time available and maintainability.

When the machine reaches state F has a special status of the inspection (IF) which determines the conditions of the machine and if it can be repaired through an exhaustive maintenance (MF) or must be replaced completely (Sust). In MF has high waiting times and high costs, the machine could return to any of the first

The machinery is in a industrial system of production inputs and outputs. This system is designed so that each time t (representing 1 day) has outputs useful for decision-making as downtime, number of parts produced, defects,

Table 1: Main model results

645

as follows, Figure 3 shows the curves of life of different models to the point of wear.

delays, etc. Those are transformed into costs for the analysis of the financial part of the model. It takes into account each day that the machine is under maintenance or inspection to assess the percentage of the working life of machinery. The model collected how often the machine has been sent to light or intensive maintenance to determine the maintainability of the equipment.

As noted on table 1, you have different factors to select the most appropriate policy. If you search the availability of the machine would select the alternative 5, if you want to reduce costs is due to select Alternative 1 and to extend the life of machine to the maximum recommended Scenario 3.

2.4 Setting the model

4. Conclusions

The proposed model has great flexibility in representing various types of maintenance systems and situations, which may range from corrective maintenance, preventive and situational analysis of the team.

The optimization of maintenance activities through the implementation of techniques such as simulation, situational analysis and maintenance management techniques play a key role in industries which increase their competitiveness by achieving better results with high quality at lower cost possible. The paper presents a markov model adaptation method that allows adjustment of the basic model to user-expected changes in maintenance policy. The model has the ability to adjust to working environments commonly found in industries and can be extended to represent more complex systems, confirming the effectiveness of a system or proposing a new one with a tuning of policies.

Figure 3: Results of life curve of equipment for default and generated maintenance policy from markov model. As input variables are the specific qualities of the maintenance system, the main ones: the set time between inspections, the rigor of the inspection, maintenance time and costs and the length of time between the various states of impairment, as well the company variables such as demand, production costs, sales costs, compliance costs and rework costs.

3. Application example Model was used to simulate production machinery within a company dedicated to selling radiator cooling tubes. The system was tested in 5 different conditions: default policy, inspections of 10, 20 and 40 days, without inspections. There were 3 replicates of each type and the results were

646

References BIOGRAPHICAL NOTES Abdel-Hamed, Mohamed S. (2004). Optimal predictive maintenance policies for a deteriorating system: The total discounted cost and the long-run average cost cases. Communications in statistics. Ascher H.; Feingold H. (1984). Repairable systems Reliability. Modelling Inference, misconceptions and their causes. Marcel- Dekker Christer, A.H. (1999). Developments in delay time análisis for modelling plant maintenance. Journal of Operations Research Society. Jarosław Sugie (2007). Modeling Changes in Maintenance Activities through Fine-Tuning Markov Models of Ageing Equipment. 2nd International Conference on Dependability of Computer Systems.

Francisco Castellanos obtained his bachelor degree in Industrial Engineering (Tecnológico de Monterrey, México). Currently is studying a master's degree on System engineering at UNAM. Ann Wellens is a chemical engineer with postgraduate studies in Industrial Administration (KUL, Belgium) and a master degree in Environmental Engineering (UNAM, Mexico). At the moment she is a full-time lecturer in the Systems Department of the Industrial and Mechanical Engineering Division of the National University of Mexico (UNAM). She has been working in air pollution issues for the last 15 years, dictating courses, collaborating in research projects and participating in conferences related with mathematical modeling of air pollution dispersion and statistics.

647

DESIGN AND DEVELOPMENT OF DATA ANALYSIS MODULES FOR THE AERMOD AND CALPUFF SIMULATION MODELS Wellens A.(a), García G.(b) (a)(b)

Facultad de Ingeniería, UNAM-MÉXICO

(a)

[email protected], (b)[email protected]

ABSTRACT Pollutants enter the environment in diverse ways. The dispersion of industrial chimney pollutants depend on many correlated factors, as for example:

Mathematical models for atmospheric dispersion are being used in a wide variety of industrial applications. Even simplified models have improved their formulation incorporating up-to-date knowledge regarding micrometeorology and dispersion, and can be used to estimate air pollution concentrations around, for example, industrial facilities. Two dispersion models based on Gaussian modeling of the dispersion plume, are of special interest in the simulation of small and medium scale dispersion: AERMOD and CALPUFF are recommended by the US EPA to determine air pollution dispersion.

x x x x

The physical and chemical nature of the effluent. The meteorological characteristics in the environment. The location of the chimney with respect to possible obstructions for the free movement of air. The nature of the area located downwind the chimney.

With very few exceptions, the basic approach of the current regulative platform of the EPA for air pollutant modeling in the surroundings of an industrial source has been maintained fundamentally without changes from the beginning of the air programs, approximately 30 years ago. Most used models have been Gaussian ones; they give quick results, but their development is based on quite severe assumptions. Nevertheless, in the last years significant scientific advances have been reached: these have been incorporated in the ISCST3 model (Industrialist Source Complex – Short Term Model) to design advanced Gaussian models able to evaluate pollutant transport at long distances and in complex topographical and meteorological conditions. Two of these advanced models are CALPUFF and AERMOD.

Both models are freely distributed by EPA, although independent developers offer graphical interfaces (for example ISC-AERMOD, View, Breeze, CALPUFF View) to be able to integrate in a friendly way the topographic, land use and meteorological data, and to represent the results graphically. Although these graphical interfaces are quite complete, specific research projects may need some data manipulation not provided by these interfaces. This paper proposes some external modules to AERMOD and CALPUFF that extract and prepare in a specific way the output data for certain research needs, such as comparison of the model data with DOAS measurements, etc.

1.1 The Gaussian model

Keywords: consequence analysis, dispersion modeling, CALPUFF, AERMOD.

The Gaussian model is a particular solution of the general equation for pollutant concentration transport.

1.

Figure 1 illustrates the problem to be studied, including the used coordinate system where the origin is located at the base of the chimney.

INTRODUCCIÓN

Mathematical models are used extensively in a variety of applications related to the study of air pollution. Examples are, among others, emission modeling, pollutant dispersion modeling, determination of the minimum chimney height, safety audit studies or the modeling of accident consequences. There exists a big variety of models that differ in application type, generated model output, spatial scale, temporary resolution, complexity, method of solution, reference system and required resources.

In a simple Gaussian model the concentration c of a compound located at the coordinate point (x, y, z) can be described by:

c( x , y , z )

648

Q u 2S V y V z

ª 1 y 2 z  H e 2 º exp «  » 2 V z 2 ¼» ¬« 2 V y

modeling system consists of three main components and a set of preprocessing and postprocessing programs. The main components of the modeling system are CALMET (a diagnostic 3-dimensional meteorological model), CALPUFF (an air quality dispersion model), and CALPOST (a postprocessing package). Each of these programs has a graphical user interface (GUI). In addition to these components, there are numerous other processors that may be used to prepare geophysical (land use and terrain) data in many standard formats, meteorological data (surface, upper air, precipitation, and buoy data), and interfaces to other models such as the Penn State/NCAR Mesoscale Model (MM5), the National Centers for Environmental Prediction (NCEP) Eta/NAM and RUC models, the Weather Research and Forecasting (WRF) model and the RAMS model. This model has been evaluated and improved by institutions and/or groups like the Interagency Workgroup on Air Quality Modeling (IWAQM-US), the EPA and other north-American and foreign organizations, and at present it is one of the most used models due to the EPA support.

Figure 1. The Gaussian dispersion model. 1.2 The AERMOD model Of all available models, few are widely accepted. The US Environmental Protection Agency (EPA) proposes and recommends AERMOD to model the dispersion of pollutants of fixed sources.

Both AERMOD and CALPUFF are models recommended by the United States Environmental Protection Agency and have been used in a variety of applications and countries. In México, they have been used for example by the Universidad Nacional Autónoma de México in air pollution studies around petroleum refineries or electricity generation facilities (Ruiz Suárez et al., 2010; Jazcilevich et al. 2009; Grutter et al. 2008). However, the nature of these research projects requires transformation of model results provided by AERMOD or CALPUFF to be able to compare them with real-time measurements.

AERMOD, developed by the US-EPA and the American Meteorological Society, was designed to support the EPA’s regulatory modeling programs (EPA 2010). AERMOD is a regulatory steady-state plume modeling system and includes a wide range of options for modeling air quality impacts of pollution sources, making it a popular choice among the modeling community for a variety of applications (Lakes AERMOD VIEW USER GUIDE). Together with the AERMOD code, the US-EPA provides three complementary components: AERMAP (AERMOD Terrain Preprocessor) AERMET (AERMOD Meteorological Preprocessor) and AERSURFACE, a tool that produces surface characteristics data. As AERMOD includes recent scientific knowledge with respect to the understanding of the planetary boundary layer, it has a more realistic approach in treating plume interaction with the earth’s surface than older Gaussian air dispersion models like ISCST3. Since December 2006, AERMOD replaced ISCST3 as the standard regulatory model. AERMOD, as any other mathematical dispersion model, provides only an estimate of the atmospheric concentration of environmental pollutants, and its results depend on the quality of the corresponding input data, and the methodology used for its determination.

2.

PROPOSED METHODOLOGY

Both AERMOD and CALLPUFF are written in Fortran, which is a not a very flexible programming language. For the proposed external modules, new programming technologies will be used, independently of the original Fortran code of the specific model. In this stage, the development platform .NET is proposed, as it is a powerful system, which provides quick and reliable results. The methodology used in the.NET platform will help to improve significantly the tasks and the interaction of the modules to be developed with the original AERMOD and CALPUFF code. In .NET a three layer architecture is used, in which every part of the programming is organized in the most efficient way in order to access the information rapidly and efficiently. The programming time is short due to the large number of tools provided in .NET and the versatility of its use.

1.3 The CALPUFF model CALPUFF is an advanced non-steady-state meteorological and air quality modeling system, adopted by the U.S. Environmental Protection Agency in its Guideline on Air Quality Models as the preferred model for assessing long range transport of pollutants and their impacts on Federal Class I areas and on a case-by-case basis for certain near-field applications involving complex meteorological conditions. The

649

Figure 4. Architecture of the three .NET layers.

Figura 3. Petacalco thermoelectricity facility.

Initially, two modules will be developed: the first integrates concentration values at different vertical levels to a sole integrated concentration (in ppm m) to be compared with DOAS integrated concentration data. The other one integrates monthly or seasonal data in an annual concentration, as – due to the size of the meteorological input data – the yearly average cannot be defined in a sole simulation run. In the future, a third module is planned, determining the region were predefined ambient standards are violated when AERMOD or CALPUFF surface concentrations are known. Once the requirements of each module are defined, the graphical interfaces are developed; the trial version will be evaluated to correct programming errors or to obtain information for future implementations. 3.

3.2 Objectives The general target is to compare the Gaussian model of dispersion of atmospheric pollutants AERMOD and CALPUFF with the spectroscopic skill DOAS in the industrial Petacalco complex, the Warrior's State by means of the creation of external modules to these models that realize the above mentioned comparison and help to work in a practical way the information. The specific targets of the same sound the following ones: 1. To use the skill DOAS to estimate by implication and wind below the entire emission of SO2 as well as its spatial distribution about the industrial Petacalco complex. 2. To shape the dispersion of this pollutant gas with AERMOD and CALPUFF using the topography and available meteorological information. 3. To evaluate the exits of the models AERMOD and CALPUFF using the sets of meteorological data, with the results of the measurements of the DOAS. 4. To use the modules designed to work with the information of exit of both measurements.

CASE STUDY

3.1 Study area The infrastructure of the Mexican Federal Commission of Electricity (CFE) includes 154 energy generation facilities. The number of thermoelectric generation plants distributed in Mexico is 79, of which the Petacalco thermoelectric facility is one of the most important. The Plutarco Elías Calles facility in Petacalco has a capacity of 2100 MW, in six production units. The electric power produced is transported through fifteen transmission lines between 115 and 400 kV. The plant uses coal as a primary fuel to produce high pressure steam (between 120 and 170 kg/cm2) and high temperature (of the order of 520°C), to move the electrical generator connected to the rotor of the steam turbine. The electricity generation plant is an important pollution source, emitting among other SO2, NOx, particulate matter and CO.

3.3 DOAS The Optical Spectroscopy of Distinguishing Absorption (DOAS, for its initials in English) is a method to determine the gas concentrations in the ambience by means of the analysis of the light, principally in the spectral status of the ultraviolet and visible one. The light that travels across the ambience is partially absorbed by the gases along the covered trajectory. Analysis in the vertical one For this project, the radiation dispersed by the blue sky is collected by means of a telescope and it analyzed espectroscópicamente to obtain column concentrations

650

in units of ppm*m. The concentration in column of a gas represents the concentration integrated along an indefinite trajectory.

Figure 8. DOAS transversal to be compared with CALPUFF simulation: May 12, 2009.

Figure 7. DOAS function The remarks with the DOAS, on having been measured from a vehicle realizing passages below the pen, give the possibility of estimating the gas flows making use of the speeds of spread of the pen. The column concentration measures itself to this skill while it passes below a pen or cloud of pollutant gases. The target of this strategy is to obtain the measurement of a "slice" of the pen that disperses perpendicularly over the measurement trajectory.

(a) 2 m

The facilities chimneys emit, among other pollutants, SO2 from 100m high chimneys. Besides this facility, four minor SO2 pollution sources, with lower emission heights (20 to 53 m), are found in the southwest area of the modeling region (Figures 8 and 9). To be able to evaluate CALPUFF model results, Differential Optical Absorption Spectroscopy (DOAS) was used to obtain experimental information on the column concentrations of SO2 in concentrations of ppm*m. The DOAS technique collects scattered radiation by the blue sky with a telescope and by analyzing the absorbed radiation spectroscopically, integrated SO2 concentration in the vertical column can be obtained.

(b) 12 m

Observations were performed with the passive DOAS technique assembled on a van traveling around the industrial complex, below the emitted gas plumes. The traversals downwind were used to evaluate CALPUFF performance. Figure 8 shows a specific DOAS transversal for May 12, 2009 around the electricity generation facility, obtained between 17:03 and 17:19). However, as CALPUFF provides point concentrations, and not vertically integrated concentrations, an extern Fortran module was written to integrate CALPUFF point concentrations in different layers in a column concentration comparable with DOAS results. To assure an appropriate horizontal resolution, only 5 vertical layers of different height could be considered in a first approach, as CALPUFF’s number of receptors is limited. At the moment, the programming is being extended to include more vertical layers and refine the results.

(c) 60 m

651

integrated concentration, and thus the quality of the comparison between CALPUFF and the DOAS measurements, depended strongly of the number and value of the chosen heights to obtain the integrated concentration. The extern module for vertical integration of CALPUFF model results is being adapted at the moment to be able to take into account easily receptors at more than 5 heights and to change in an easy way the chosen heights to integrate, as both variables depend on the specific information of the sources in the modeling domain.

(d) 200 m

4.

CONCLUSIONS

Extern modules for CALPUFF and AERMOD were developed to adapt standard output concentrations to specific research needs. Fortran models were written for vertical integration of point concentration in different layers, for integration of simulation data for different trimesters into an annual result, among others. These Fortran models were integrated to the CALPUFF and AERMOD simulation models to offer more flexibility in the results. These extern modules are being adapted at the moment to be able to make them more flexible to research needs and specific case studies.

(e) 500 m Figure 9. Point concentrations at different heights. CALPUFF, May 12, 2009 (17:00 a 18:00). As can be seen in Figure 9, the resulting concentration is quite different in different vertical layers: at lower height (below 100 m), the smaller sources in the southwest region of the domain generate higher SO2 concentrations, while at higher heights (see for example at 500 m), the electricity generation facility is becoming more and more important.

REFERENCES x Ruiz Suárez, L.G., Grutter de la Mora, M., Rosas Pérez I., Torres Jardón R., García Reynoso, A., Granada Macías, L.M., Torres Jaramillo, J.A., Wellens Purnal, A., Padilla Gordón, H., Belmont Dávila, R., García García, A., Rebulloza, R., Basaldud, R. (2010), Diagnóstico ambiental de la zona de influencia de la CTPPEC durante la construcción – puesta en servicio de la Unidad 7. Subproyecto aire. Informe final, presentado por el CCA-UNAM a CFE. x Jazcilevich, A., Siebe, C., Wellens A., Rosas, I. (2009), Impacto ambiental y en la salud de los habitantes de las actividades mineras en el Distrito Molango, Hidalgo. Subproyecto: estudios ambientales. Informe final de la segunda etapa del proyecto, proyecto 100662 del CCA-UNAM, presentado al International Development and Research Center of Canada (IDRC). x Grutter de la Mora M., García Reynoso A., Torres Jardón R., Limón Sánchez T., Wellens Purnal A., Basaldud R., García Escalante J. (2008), Emisiones a la Atmósfera y Calidad del Aire en la Central Termoeléctrica Plutarco Elias Calles. Primer informe parcial. Proyecto CFE/PUMA-UNAM. Mayo de 2008. x EPA (2010) Support Center for Regulatory Air Models, Technology Transfer Network. http://www.epa.gov/scram001/. Fecha de consulta: enero 2010.

The resulting vertically integrated concentration of SO2 is quite different from the default surface concentration (see Figure 9(a)) given by CALPUFF or in its case AERMOD, as concentrations in different vertical layers differ a lot. Comparison of CALPUFF/AERMOD results with DOAS was considerably better when comparing the vertically integrated concentration instead of the point concentration at the surface (Figure 10).

Figure 10. Integrated SO2 concentration in the vertical column (concentrations in ppm m). CALPUFF, May 12, 2009 (17:00 a 18:00) As could be observed after analysis of modeling results for different simulation dates and hours, the final

BIOGRAPHICAL NOTES

652

Gamar Castillo G. obtained his bachelor degree in Industrial Engineering (UNAM, México). Ann Wellens is a chemical engineer with postgraduate studies in Industrial Administration (KUL, Belgium) and a master degree in Environmental Engineering (UNAM, Mexico). At the moment she is a full-time lecturer in the Systems Department of the Industrial and Mechanical Engineering Division of the National University of Mexico (UNAM). She has been working in air pollution issues for the last 15 years, dictating courses, collaborating in research projects and participating in conferences related with mathematical modeling of air pollution dispersion and statistics.

653

'(9(/230(172)$6,08/$7,21722/)25&216(&8(1&($1$/sDŽĚƵůĞ



   

 &ŝŐƵƌĂϳ͘>sͲĞƚĞƌŵŝŶĂƚŝŽŶŽĨ&ĂŶĚd͘



 &ŝŐƵƌĞϱ͘WĂƌĂŵĞƚĞƌƐŽĨƚŚĞĞdžƉůŽƐŝŽŶďĂƐĞĚŽŶƚŚĞĚŝƐƚĂŶĐĞ ĐůŝŵďŝŶŐ͘     





 &ŝŐƵƌĂϴ͘>sͲĞƚĞƌŵŝŶĂƚŝŽŶŽĨĨƌĂŐŵĞŶƚƐ

657

x x x x x

&ŝŐƵƌĂϵ͘>ŝƋƵŝĚĚŝƐĐŚĂƌŐĞƐ

x



0,1,67(5,2'(75$%$-2 0 is a finite set of transitions (with P ∪ T = ∅ and P ∩ T = ∅). – Cl is a set of colours. – Co : P ∪ T → Cl is a colour function defining place marking colours and transition occurrence colours. ∀p ∈ P, Co(p) = {ap,1 , ap,2 , . . . , ap,up } ⊆ Cl is the set of up possible colours of tokens in p, and ∀t ∈ T, Co(t) = {bt,1 , bt,2 , . . . , bt,vt } ⊆ Cl is the set of vt possible occurrence colours of t. – P re(p, t) : Co(t) → Co(p)M S is an element of the pre-incidence function and is a mapping from the set of occurrence colours of t to a multiset over the set of colours of p, ∀p ∈ P, ∀t ∈ T . It can be represented by a matrix whose generic element P re(p, t)(i, j) is equal to the weight of the arc from p w.r.t colour ap,i to t w.r.t colour bt,j . When there is no arc with respect to the given pair of nodes and colours, the element is 0. – P ost(p, t) : Co(t) → Co(p)M S is an element of the post-incidence function, which defines weights of arcs from transitions to places with respect to colours. M (p) : Co(p) → N is the marking of place p ∈ P and defines the number of tokens of a specified colour in the place for each possible token colour in p. Place marking can be represented as a multiset M (p) ∈ Co(p)M S and the net marking M can be represented as a k × 1 vector of multisets M (p). M0 is the initial marking of a Colured Petri net. 3.1. Timed models As described in (Bowden 2000), there are three basic ways of representing time in Petri nets: firing durations (FD), holding durations (HD) and enabling durations (ED). The FD principle says that when a transition becomes enabled it removes the tokens from input places immediately but does not create output tokens until the firing duration has elapsed. When using HD principle, a firing has no duration but a created token is considered unavailable for the time assigned to transition that created the token. The unavailable token can not enable a transition and therefore causes a delay in the subsequent transition firings. With ED principle, the firing of the transitions has no duration while the time delays are represented by forcing transitions that are enabled to stay so for a specified period of time before they can fire.

682

are 0, the arc is omitted. Let • tb ⊆ P × Cl denote the set of places and colours which are inputs to occurrence colour b ∈ Co(t) of transition t ∈ T , i.e., there exists an arc from every (p, a) ∈ • t to t with respect to colours a ∈ Co(p) and b ∈ Co(t). To determine the availability and unavailability of tokens, two functions on the set of markings are defined. The set of markings is denoted by M. Given a marking and model time, m : P × M × T S → Co(p)M S defines the number of available coloured tokens, and n : P × M × T S → Co(p)M S the number of unavailable coloured tokens for each place of a TPN at a given model time τk ∈ T S. Two timed markings can be added (denoted +τ ) in a similar way as multisets, i.e. by making a union of the corresponding multisets. The definition of subtraction is somewhat more problematic. To start with, a comparison operator is defined. Let M1 and M2 be markings of a place p ∈ P . By definition, M1 ≥τ M2 iff m(p, M1 , τk ) ≥ m(p, M2 , τk ), ∀τk ∈ T S, ∀a ∈ Co(p). Similarly, the subtraction is defined by the number of available tokens, and the subtrahend should not contain any unavailable tokens. Let M1 , M2 and M3 be markings of a place p ∈ P , M1 ≥τ M2 , and m(p, M1 , τk ), m(p, M2 , τk ), and m(p, M3 , τk ), be the corresponding numbers of available tokens at time τk , and n(p, M2 , τk ) = 0. The difference M3 = M1 −τ M2 is then defined as any M3 ∈ M having m(p, M3 , τk ) = m(p, M1 , τk ) − m(p, M2 , τk ). Using the above definitions, the firing rule of a CTPN can be defined. Given a marked CT P N = (N , M ), a transition t is time enabled at time τk w.r.t occurrence colour b ∈ Co(t), denoted M [tb τk iff m(p, M, τk ) ≥ P re(p, t)(b), ∀p ∈ •t. An enabled occurrence transition can fire, and as a result removes tokens from input places and creates tokens in output places. If transition t fires w.r.t occurrence colour b, then the new marking is given by M  (p) = M (p) −τ P re(p, t)(b)@τk +τ P ost(p, t)(b)@(τk + f (t, b)), ∀p ∈ P . Here the subtraction operation is implemented in such a way that in case of several choices, the token with the oldest timestamp is always removed first. If marking M2 is reached from M1 by firing tb at time τk , this is denoted by M1 [tb τk M2 . The set of markings of TPN N reachable from M is denoted by R(N , M ).

The ED concept is more general than HD. Furthermore, in (Lakos and Petrucci 2007) an even more general concept is used, which assigns delays to individual arcs, either inputs or outputs of a transition. This way both ED and HD concepts are covered, and the enabling delay may even depend on the source of transition triggering while holding delay may differ among different activities started by the same transition. When modelling several performance optimization problems, e.g. scheduling problems, such a general framework is not needed. It is natural to use HD when modelling most scheduling processes as transitions represent starting of operations, and generally once an operation starts it does not stop to allow another operation to start in between. HD principle is also used in timed version of CPNs defined by Jensen, although the unavailability of the tokens is only defined implicitly through the corresponding time stamps. While CPNs allow the assignment of delays both to transition and to output arcs, we further simplify this by allowing time delay inscriptions to transitions only. This is sufficient for the type of examples investigated here, and can be generalized if necessary. To include a time attribute of the marking tokens, which implicitly defines their availability and unavailability, the notation of (Jensen 1997) will be adopted. Colours are adjoined to token number by ‘c notation and coloured tokens are accompanied with a timestamp, which is written next to the token number and colour and separated from the colour by @. E.g., two c-coloured tokens with time stamp 10 are denoted 2‘c@10. A collection of tokens with different colours and/or time stamps is defined as a multiset, and written as a sum (union) of sets of timestamped coloured tokens. E.g., two c-coloured tokens with time stamp 10 and three d-coloured tokens with timestamp 12 are written as 2‘c@10+3‘d@12. The timestamp of a token defines the time from which the token is available. Time stamps are elements of a time set T S, which is defined as a set of numeric values. In many software implementations the time values are integer, i.e. T S = N, but will be here admitted to take any positive real value including 0, i.e. T S = R+ 0 . Timed markings are represented as collections of time stamps and are multisets over T S: T SM S . By using HD principle the formal representation of a Coloured Timed Petri net is defined as follows. CT P N = (N , M0 ) is a Coloured Timed Petri net system, where: – N = (P, T, P re, P ost, Cl, Co, f ) is a Coloured Time Petri net structure with (P, T, P re, P ost, Cl, Co) as defined above.

4. COLOURED PETRI NET MODELLING OF SCHEDULING PROBLEMS An important concept in PNs is that of conflict. Two transition firings are in conflict if either one of them can occur, but not both of them. Conflict occurs between transitions that are enabled by the same marking, where the firing of one transition disables the other transition. The conflicts and the related conflict resolution strategy play a central role when modelling scheduling problems. This may be illustrated by a simple example, shown in Figure 1. The example involves two machines M = {M1 , M2 }, which should process two jobs J = {J1 , J2 }, and where J1 = {o1 (M1 ) ≺ o2 (M2 )} and J2 = {o3 (M1 )}. Job J1 therefore consist of two opera-

– f : Co(t) → T S is the time function that assigns a non-negative deterministic time delay to every occurrence colour of transition t ∈ T . – M (p) : Co(p) → T SM S is the timed marking, M0 is the initial marking of a timed Petri net. 3.2. Firing rule Functions P re and P ost define the weights of directed arcs, which are represented by arc inscriptions in the matrix form. In the case when the all the weights in the matrix

683

J1

t

M1

O3

p

J2

t3b

p11 t2b

t

po1

1b

1f

p po3

t

po2

M2

m1

t

2f

p

12

J1

p

M1

m2

p

3f

p

M2

m1

J1,J2

p

m2

21

Figure 2: A CTPN model of a simple scheduling problem t1

t2

p1

t

t2

p11

1

p12

t3

p2

t4

p3

p4

J1 M1

O3 J2

p2

O2

O1

p

t2

p1

1

J

pJ

J1

t

p

J2

p

O2

O1,O3

O2

O1

p

t3

M1

p

m1

M2

p

M2

M3

M4

Figure 3: A simple job shop problem

m2

p21

J2

Table 1: Operation durations for a simple job shop problem Operation\Job o1 o2 o3 o4

Figure 1: A PN model of a simple scheduling problem tions, the first one using machine M1 and the second one machine M2 , while job J2 involves a single operation using machine M1 . Obviously, the two jobs compete for machine M1 . This is modelled as a conflict between transitions starting corresponding operations. Place pm1 is a resource place. It models the machine M1 and is linked to t1b and t3b , which start two distinct operations. Clearly, the conflict between t1b and t3b models a decision, whether machine M1 should be allocated to job J1 or J2 first. Similarly, other decisions are modelled as conflicts linked to resource places. The solution of the scheduling problem therefore maps to a conflict resolution in the given Petri net model. The transitions that model finishing of operation (tif in Figure 1) are not relevant for scheduling and can be removed. Same holds for intermediate buffer places since the holding duration interpretation of transition delays guarantees that a subsequent transition can not fire before the precedent transition delay expires. The model can be therefore simplified as shown in the lower part of Figure 1. The occupation of a shared resource Mi during the evolution of the system is marked by a presence of unavailable token in the corresponding place pmi . With the introduction of token and occurrence colours, the resource sharing as described above can be represented in even much more compact model. Several jobs that go through a similar operation sequence can be folded together and represented by a single place/transition sequence with different token colours. The transition occurrence colours enable to distinguish different jobs both in terms of operation durations as well as in terms of their dependence on shared resources. The model from Figure 1 therefore maps to the model in Figure 2. The two jobs are represented by two token colours while a third colour is added to model resource availability. Both remaining transitions appear with two occurrence colours to model different durations where the absence of the second operation in

J1 54 34 61 2

J2 9 15 89 70

J3 38 19 28 87

J4 95 34 7 29

Table 2: machine requirements for a simple job shop problem Operation\Job o1 o2 o3 o4

J1 3 1 4 2

J2 4 1 2 3

J3 1 2 3 4

J4 1 3 2 4

job J2 is simply modelled by setting the duration to zero. A more elaborated example is shown in Figure 3. The model is based on a test example from Taillard (1993). It consists of four jobs and four machines. Every job includes four operations. Operation durations are shown in Table 1 and resource requirements in Table 2. Note that arc weights are not shown in the figure, they will be shown in the sequel. Nevertheless, only the arcs with at least one nonzero weight for any occurrence colour are shown. Further compaction can be achieved by folding the operation places. Job sequences, operation durations and resource requirements are coded by different sets of colours and corresponding transition guards and expressions (Mujica, Piera and Narciso 2010). Since the transition guards and expressions are not supported by the type of CTPNs used in this paper, this representation can not be used here. The proposed representation is therefore not the most compact one but has the advantage of a very efficient coding in a general mathematical analysis software, e.g. Matlab. In Matlab, the flow matrices of a CPN can be represented as cell matrices of size |P | × |T |, where each element is a cell containing weight matrix of size |Co(p)| × |Co(t)|. E.g. for example in Figure 3 the corresponding

684

pre-incidence matrix is ⎡ I4 ⎢ 0 ⎢ ⎢ 0 ⎢ ⎢ 0 ⎢ P re = ⎢ 0 ⎢R ⎢ 11 ⎢R ⎢ 21 ⎣R31 R41

>> PTPre=cell2mat(Pre)

Time (LPT), can be introduced when solving the conflicting situations. By introducing different heuristic dispatching rules (priority rules) decisions can be made easily. In this way, only one path from the reachability graph is calculated, which means that the algorithm does not require a lot of computational effort. The schedule of process operations can be determined by observing the marking evolution of the net. Depending on the given scheduling problem a convenient rule should be chosen. Usually, different rules are needed to improve different predefined production objectives (makespan, throughput, production rates, and other temporal quantities). A more extensive exploration of the reachability tree is possible by PN-based heuristic search method proposed by Lee and DiCesare (1994). It is based on generating parts of the Petri net reachability tree, where the branches are weighted by the time of the corresponding operations. Sum of the weights on the path from the initial to a terminal node gives a required processing time by the chosen transition firing sequence. Such a sequence corresponds to a schedule, and by evaluating a number of sequences a (sub)optimal schedule can be determined. Recent reports in scheduling literature show an increased interest in the use of meta-heuristics, such as genetic algorithms (GA), simulated annealing (SA), and tabu search (TS). Meta-heuristics have also been combined with Petri net modelling framework to solve complex scheduling problems (Tuncel and Bayhan 2007). With such an approach, the modelling power of Petri nets can be employed, and relatively good solutions of scheduling problems can be found with a reasonable computational effort. Compared to reachability tree based search methods, metaheuristics require less memory.

unfolds the pre-incidence matrix from the previous example into a pre-incidence matrix of an equivalent P/T Petri net and the command

5. COLOURED PETRI NET SIMULATION BASED EXPLORATION OF THE SOLUTION SPACE

0 I4 0 ··· R12 R22 R32 R42

··· 0 I4 0 ··· R13 R23 R33 R43



0 0 ⎥ ⎥ 0 ⎥ ⎥ I4 ⎥ ⎥ 0 ⎥ R14 ⎥ ⎥ R24 ⎥ ⎥ R34 ⎦ R44

(1)

where I4 stands for 4 × 4 identity matrix and zeros should be interpreted as 4 × 4 zero matrices. Rij define i-th resource requirements of j-th operation within jobs:



R11 = 0 0 1 1 R12 = 1 1 0 0 R21 = 0 0 0 0 R22 = 0 0 1 0 R31 = 1 0 0 0 R32 = 0 0 0 1 R41 = 0 1 0 0 R42 = 0 0 0 0



(2) R13 = 0 0 0 0 R14 = 0 0 0 0 R23 = 0 1 0 1 R24 = 1 0 0 0 R33 = 0 0 1 0 R34 = 0 1 0 0 R43 = 1 0 0 0 R44 = 0 0 1 1 Furthermore, the cell matrix can be any time converted to an incidence matrix of the corresponding unfolded P/T net and reverse, the P/T net can be folded back. The only information necessary consists of the place and transition colour sets of individual nodes in the CPN. For example, the Matlab command

>> Pre=mat2cell(PTPre, ncolP, ncolT)

In our previous work (L¨oscher, Muˇsiˇc and Breitenecker 2007, Muˇsiˇc, L¨oscher and Breitenecker 2008) different ways of solution space exploration were studied. Extensive testing of the reachability tree search based approaches has been performed. The approach is very general, as it can be applied to any kind of scheduling problem that can be represented as a Petri net. Unfortunately, the approach does not perform very well on the standard job shop benchmarks (Muˇsiˇc 2008). This motivated the exploration of alternative approaches, including local search based techniques. In (L¨oscher, Muˇsiˇc and Breitenecker 2007) the approach is presented, which extends the Petri net representation by sequences and priorities. Priorities are used as a way of parametrizing the conflict resolution strategy. For this purpose a priority ranking is assigned to transitions. If there is a conflict between a pair of transitions the transition with higher priority will fire. Another way of parametrization is to select disjoint groups of transitions and map them to sequences. A firing list is defined by ordering transitions within the group. During the model evolution a set of sequence counters is maintained and all transitions belonging to sequences are disabled except of transitions corresponding to the current

reproduces back the original cell matrix, provided that vectors ncolP and ncolT contain information about numbers of token and occurrence colours for all p ∈ P and t ∈ T . E.g., for the above exam

T and ple ncolP = 4 4 4 4 4 1 1 1 1

T ncolP = 4 4 4 4 . This way the CPN framework can be used to efficiently encode various scheduling problems into a compact representation. Later the CPN representation can be analyzed directly, or can be translated into an equivalent P/T Petri net, which enables the application of standard PN analysis methods as well as PN based scheduling techniques. 4.1. Derivation of optimal or sub-optimal schedules A derived Coloured Petri net model can be simulated by an appropriate simulation algorithm. During the simulation, the occurring conflicts are resolved ’on the fly’, e.g. by randomly choosing a transition in conflict that should fire. Instead, heuristic dispatching rules (Haupt 1989), such as Shortest Processing Time (SPT) or Longest Processing

685

leads to a neighbourhood solution of a given solution and the related modification is usually defined through a neighbourhood function. In the work of (L¨oscher, Muˇsiˇc and Breitenecker 2007) several neighbourhood functions as well as different local search strategies were implemented in the PetriSimM toolbox for Matlab and some results are shown in (L¨oscher, Muˇsiˇc and Breitenecker 2007, Muˇsiˇc, L¨oscher and Breitenecker 2008).

M1

M2

M3

M4

color 1 color 2 color 3 color 4

0

50

100

150

200

250

Figure 4: A possible solution of the given job-shop problem

5.1. Generation of feasible neighbourhood solutions from a CTPN model The problem in the previously described approach is that by perturbing sequence index vectors the resulting transition firing sequence may easily become infeasible, which results in a deadlock during simulation. The search procedures implemented in PetriSimM were designed so that such an infeasible solution is ignored and a new perturbation is tried instead. While this works for many problems, in some cases the number of feasible sequences is rather low and such an algorithm can easily be trapped in an almost isolated point in the solution space. The job shop scheduling approaches reported in the OR literature started to address the issue of efficient neighbourhood generation quite a while ago. With the wide acceptance of the Tabu search algorithm as the most promising methods for schedule optimisation the design of efficient neighborhood generation operator become the central issue and several such operators have been proposed (Blazewicz, Domschke and Pesch 1996, Jain, Rangaswamy and Meeran 2000, Watson, Whitley and Howe 2005). The question is how to link these operators and related effective schedule optimization algorithms with Coloured Petri net representation of scheduling problems. As mentioned above the Petri net scheduling methods have advantages in unified representation of different aspect of underlying manufacturing process in a well defined framework. Unfortunately, the related optimization methods are not as effective as some methods developed in the OR field. The link of two research areas could be helpful in bridging the gap between highly effective algorithms developed for solving academic scheduling benchmarks and complex real-life examples where even the development of a formal model can be difficult (Gradiˇsar and Muˇsiˇc 2007). A possible way of such a link is the establishment of a correspondence of a critical path and the sequence index vectors described previously. In a given schedule the critical path CP is the path between the starting and finishing time composed of consequent operations with no time gaps:

state of the sequence counters. After firing such a transition the corresponding sequence counter is incremented. This way the transition firing sequence can be parameterized. If the model represents a scheduling problem, the sequence obtained by a simulation run of the Petri net model from the prescribed initial to the prescribed final state is a possible solution to the problem, i.e. it represents a feasible schedule. E.g., the model from Figure 3 can be simulated by applying SPT rule (Haupt 1989) as a default conflict resolution mechanism. The resulting sequence represents a possible schedule, shown in Figure 4. The same schedule can be obtained by fixing the sequential order of transitions in conflicts related to shared resources in the system. E.g. in the above example the shared resources are machines M1 to M4. Related sets of transitions are: SM 1 SM 2 SM 3 SM 4

= {t1,c3 , t1,c4 , t2,c1 , t2,c2 } = {t2,c3 , t3,c2 , t3,c4 , t4,c1 } = {t1,c1 , t2,c4 , t3,c3 , t4,c2 } = {t1,c2 , t3,c1 , t4,c3 , t4,c4 }

(3)

where ti,cj denotes cj occurrence colour of ti and colour cj corresponds to job Jj . If these sets are mapped to four independent sequences, and a set of index vectors V = {V1 , V2 , V3 , V4 } is adjoined, where Vi is a corresponding permutation of integer values i, 1 ≤ i ≤ 4: V1 V2 V3 V4

= {1, 4, 2, 3} = {1, 2, 4, 3} = {1, 3, 4, 2} = {1, 3, 2, 4}

(4)

a supervised simulation run, which forces the prescribed sequential order of conflicting transitions, results in the same schedule as above. The sequence supervised simulation is implemented by a simple modification of the regular CTPN simulation algorithm. After the enabled transitions are determined in each simulation step, the compliance of the set of enabled transitions to the state of the sequence counters is checked. Transitions that take part in defined sequences but are not pointed to by a counter are disabled. The exploration of the solution space and the related search for the optimal schedule can then be driven by modifications of sequence index vectors. Such a modification

CP = {Oi : ρi = ρi−1 + τi−1 , i = 2 . . . n}

(5)

where Oi are operations composing the path, ρi is the release (starting) time of operation Oi , and τi is the duration of Oi . The operations Oi on the path are critical operations. Critical operations do not have to belong to the same machine (resource) but they are linked by starting/ending times.

686

M3

M4

color 1 color 2 color 3 color 4

M1

M1

M2

M2

M3

M4

color 1 color 2 color 3 color 4

t1,c3 t2,c3

t3,c2

t4,c2

0

t2,c4 t3,c4 t4,c4

50

100

150

200

250

Figure 6: An optimized solution of the given job-shop problem

Figure 5: A critical path within a schedule and critical transitions

are often encountered when unrestricted permutations on the index vectors are used. Based on this observation a set of neighbourhood functions can be defined which limit the permutations of the index vectors in a way that will produce feasible firing sequences only. Several widely used move operators can be implemented. Such a neighbourhood function permits the optimisation of schedules represented as PN or CTPN models by a wide variety of local search optimization techniques. It is also important to note that such a neighbourhood function is comparable to exploring the reachability tree in an event driven manner. It is possible that certain feasible firing sequence imposes one or more intervals of idle time between transitions, i.e. some transitions are enabled but can not fire due to sequence restrictions. This is different from the exploration in a time driven manner when a transition has to be fired whenever at least one transition is enabled. The difference is important in cases when the optimal solution can be missed unless some idle time is included in the schedule as shown in (Piera and Muˇsiˇc 2011). The described neighbourhood generation procedure was coded in Matlab and used in combination with a simple Simulated annealing (SA) search algorithm. Comparison of the minimum makespan for the above job shop problem calculated by the proposed algorithm and some other standard algorithms is shown in Table 3. SA-SPT denotes the combined algorithm with the Simulated Annealing and the SPT rule (Muˇsiˇc 2009), RT-search stands for a reachability tree based heuristic search (Lee and DiCesare 1994, Yu, Reyes, Cang and Lloyd 2003), and SA-CPN-N1 denotes Simulated Annealing and a CPN-based neigbourhood function as proposed in this paper. Clearly, the reachability tree based search and SACPN-N1 outperform other algorithms with regard to the result. It must be noted, however, that the computational effort in the case of reachability tree based search is much higher.

Critical path can be decomposed in a number of blocks. A block is the longest sequence of adjacent critical operations that occupy the same resource. The length of the path equals the sum of durations of critical operations and defines the makespan Cmax :  τi (6) Cmax = Oi ∈CP

Figure 5 shows a redrawn gantt chart from Figure 4 with indication of the critical path and the sequence of critical operations. The shown critical path consists of 5 blocks. Critical operations in Figure 5 are denoted by transition labels that trigger the start of a critical operation when fired. A transition that triggers a critical operation will be called a critical transition. The scheduling literature describes several neighborhoods based on manipulations (moves) of critical operations (Blazewicz, Domschke and Pesch 1996). One of the classical neighborhoods is obtained by moves that reverse the processing order of an adjacent pair of critical operations belonging to the same block (van Laarhoven, Aarts and Lenstra 1992). Other neighbourhoods further restrict the number of possible moves on the critical path, e.g. (Nowicki and Smutnicki 1996). Clearly every critical transition participates in one of the conflicts related to shared resources, e.g. sets (3) for the given case. If these transitions are linked to predefined firing sequences parameterized by index vectors Vi (4), a move operator corresponds to a permutation of an index vector. For example, in the schedule shown in Figure 5 a move can be chosen, which swaps the two operations in the third block on the critical path. This corresponds to the swap of transitions t4,c2 and t2,c4 in the sequence SM 3 , which is implemented by the exchange of third and fourth element within V3 index vector: move(V3 ) : {1, 3, 4, 2} → {1, 3, 2, 4}

Table 3: Calculated makespan for a simple job shop problem

A new schedule obtained by simulation with modified V3 is shown in Figure 6. When the move is limited to swap of a pair of the adjacent operations in a block on the critical path this narrows down the set of allowed permutations. The most important feature of such a narrowed set of permutation on the index vector is that every permutation from this set will result in a feasible firing sequence, i.e. a feasible schedule. Therefore no deadlock solutions can be generated, which

Algorithm SPT LPT RT-search SA-SPT SA-CPN-N1

687

Makespan 286 341 272 286 272

Table 4: Calculated makespan for a set of 15 jobs/15 machines problems Makespan ta01 ta02 SPT* 1462 1429 LPT* 1701 1674 RT-search 1592 1465 SA-SPT 1359 1358 SA-CPN-N1 1299 1326 optimum 1231 1244 * min out of 100 runs

academic scheduling benchmarks and complex real-life examples where even the development of a formal model can be difficult.

Algorithm

ta03 1452 1655 1637 1352 1357 1218

ta04 1668 1751 1590 1362 1353 1175

ta05 1618 1828 1568 1352 1344 1224

ACKNOWLEDGMENTS The presented work has been partially performed within Competence Centre for Advanced Control Technologies, an operation co-financed by the European Union, European Regional Development Fund (ERDF) and Republic of Slovenia, Ministry of Higher Education, Science and Technology.

Results in Table 3 were obtained by implementation of simple neighbourhood based on N1 move operator of (van Laarhoven et al. 1992) which was also used in (Taillard 1993) - the notation N1 is taken from (Blazewicz, Domschke and Pesch 1996). Only a single critical path was considered. Other neighbourhoods can be easily implemented and these as well as some other extensions of neighbourhood generation algorithm are currently being tested. The computational complexity drawback of reachability tree based search is much more obvious with complex problems. Table 4 shows the preliminary results of a set of standard benchmark problems with 15 jobs and 15 machines (Taillard 1993). For reference also the optimal values are listed (source: http://mistic.heigvd.ch/taillard/). The reachability tree based search has to be limited to predefined maximum tree size in order to complete in a reasonable time. In contrast to that the SA-SPT and proposed SACPN-N1 algorithms are able to improve the initial SPT solutions with a moderate effort. A prototype implementation of tabu search algorithm (TS-CPN-N1) has also been tested and the obtained results are comparable to the SA based search. It is expected that the tests with other neighbourhood operators would further improve the obtained results, which is one of the tasks for the future work.

REFERENCES Basile, F., Carbone, C. and Chiacchio, P., 2007. Simulation and analysis of discrete-event control systems based on Petri nets using PNetLab, Control Engineering Practice, 15, 241–259. Blazewicz, J., Domschke, W. and Pesch, E., 1996. The job shop scheduling problem: Conventional and new solution techniques, European Journal of Operational Research, 93, 1–33. Bowden, F. D. J., 2000. A brief survey and synthesis of the roles of time in petri nets, Mathematical & Computer Modelling, 31, 55–68. Brucker, P., 2001. Scheduling Algorithms, Springer-Verlag Berlin Heidelberg. Dell’Amico, M. and Trubian, M., 1993. Applying tabu search to the job-shop scheduling problem, Ann. Oper. Res., 41, 231–252. Gradiˇsar, D. and Muˇsiˇc, G., 2007. Production-process modelling based on production-management data: a Petri-net approach, International Journal of Computer Integrated Manufacturing, 20 (8), 794–810. Haupt, R., 1989. A survey of priority rule-based scheduling, OR Spectrum, 11 (1), 3–16. Jain, A., Rangaswamy, B. and Meeran, S., 2000. New and ”stronger” job-shop neighborhoods: A focus on the method of nowicki and smutnicki(1996), Journal of Heuristics, 6 (4), 457–480. Jensen, K., 1997. Coloured Petri Nets: Basic Concepts,Analysis Methods and Practical Use, Vol. 1, 2 edn, Springer-Verlag, Berlin. Lakos, C. and Petrucci, L., 2007. Modular state space exploration for timed Petri nets, International Journal on Software Tools for Technology Transfer, 9, 393– 411. Lee, D. Y. and DiCesare, F., 1994. Scheduling flexible manufacturing systems using Petri nets and heuristic search, IEEE Transactions on robotics and automation, 10 (2), 123–132. L¨oscher, T., Muˇsiˇc, G. and Breitenecker, F., 2007. Optimisation of scheduling problems based on timed petri nets, Proc. EUROSIM 2007, Vol. II, Ljubljana, Slovenia. Mujica, M., Piera, M. A. and Narciso, M., 2010. Revisiting state space exploration of timed coloured petri net models to optimize manufacturing system’s performance, Simulation Modelling Practice and Theory, 18, 1225–1241.

6. CONCLUSIONS The presented results indicate that the proposed combination of CPN models, sequence based conflict resolution and local search performs relatively well with a moderate computational effort. The approach may be interesting for practice, in particular because of the ability to use various existing PN or CPN models of different problems. In general, any scheduling problem can be optimized that can be represented by a timed Petri net in such a way that relations among jobs and shared resources are fixed and a shared resource always participates in the given job’s operation sequence, regardless or the determined schedule. The Petri net scheduling methods have advantages in unified representation of different aspect of underlying manufacturing process in a well defined framework. The investigations show, however, that the related optimization methods are not as effective as some methods developed in the Operations Research field. The link of two research areas could be helpful in bridging the gap between highly effective algorithms developed for solving

688

Muˇsiˇc, G., 2008. Timed Petri net simulation and related scheduling methods: a brief comparison, The 20th European Modeling & Simulation Symposium, Campora S. Giovanni (Amantea, CS), Italy, pp. 380–385. Muˇsiˇc, G., 2009. Petri net base scheduling approach combining dispatching rules and local search, 21th European Modeling & Simulation Symposium, Vol. 2, Puerto de La Cruz, Tenerife, Spain, pp. 27–32. Muˇsiˇc, G., L¨oscher, T. and Breitenecker, F., 2008. Simulation based scheduling applying Petri nets with sequences and priorities, UKSIM 10th International Conference on Computer Modelling and Simulation, Cambridge, UK, pp. 455–460. Nowicki, E. and Smutnicki, C., 1996. A fast taboo search algorithm for the job shop problem, Management Science, 42 (6), 797–813. Piera, M. A. and Muˇsiˇc, G., 2011. Coloured Petri net scheduling models: Timed state space exploration shortages, Math.Comput.Simul., p. in press. Pinedo, M. L., 2008. Scheduling: Theory, Algorithms, and Systems, 3rd edn, Springer Publishing Company, Incorporated. Taillard, E., 1993. Benchmarks for basic scheduling problems, European Journal of Operational Research, 64, 278–285. Taillard, E. D., 1994. Parallel taboo search techniques for the job shop scheduling problem, Informs Journal on Computing, 6, 108–117. Tuncel, G. and Bayhan, G. M., 2007. Applications of Petri nets in production scheduling: a review, Interna-

tional Journal of Advanced Manufacturing Technology, 34, 762–773. Vaessens, R. J. M., Aarts, E. and Lenstra, J., 1996. Job shop scheduling by local search, INFORMS Journal on Computing, 8, 302–317. van Laarhoven, P., Aarts, E. and Lenstra, J., 1992. Job shop scheduling by simulated annealing, Operations Research, 40, 113–125. Watson, J. P., Whitley, L. D. and Howe, A. E., 2005. Linking search space structure, run-time dynamics, and problem difficulty: A step toward demystifying tabu search, Journal of Artificial Intelligence Research, 24, 221–261. Yu, H., Reyes, A., Cang, S. and Lloyd, S., 2003. Combined Petri net modelling and AI based heuristic hybrid search for flexible manufacturing systems-part II: Heuristic hybrid search, Computers and Industrial Engineering, 44 (4), 545–566.

AUTHOR BIOGRAPHY ˇ C ˇ received B.Sc., M.Sc. and Ph.D. deˇ GASPER MUSI grees in electrical engineering from the University of Ljubljana, Slovenia in 1992, 1995, and 1998, respectively. He is Associate Professor at the Faculty of Electrical Engineering, University of Ljubljana. His research interests are in discrete event and hybrid dynamical systems, supervisory control, planning, scheduling, and industrial informatics. His Web page can be found at http://msc.fe.uni-lj.si/Staff.asp.

689

PLANT CAPACITY ANALYSIS IN A DAIRY COMPANY, APPLYING MONTECARLO SIMULATION Joselito Medina-Marin(a), Gilberto Perez-Lechuga(b), Juan Carlos Seck-Tuoh-Mora(c), Norberto HernandezRomero(d), Isaias Simon-Marmolejo(e) (a,b,c,d)

Advanced Research Centre in Industrial Engineering, Autonomous University of Hidalgo State, Pachuca, Hidalgo, México (e) Superior School of Cd. Sahagún, Autonomous University of Hidalgo State, Cd. Sahagún, Hidalgo, México (a)

[email protected], (b)[email protected], (c)[email protected], (d) [email protected], (e)[email protected]

1.1. Company description The plant capacity analysis reported in this paper was developed in a dairy company, which for confidentiality reasons the name of the company is omitted. This company is one of the highest milk distributers in all the Mexican territory, it processes almost 3 million of milk liters every day, and more than 900 million per year. Furthermore, the company produces more than 100 milk products, which generates setup changes. The company has a raw milk harvesting system, which is collected from ranches of associated ranchers, according to a morning schedule. The raw milk is transported in tankers from the ranches to the production plant, where the raw milk is pasteurized and ultra pasteurized. The production plant has silos for the harvested milk, silos for the processed milk (pasteurized and ultra pasteurized), and equipment for bottling and packing. Furthermore, according to client demand, ultra pasteurized milk is added with flavors, which needs silos for the bottling process, where the milk is mixed with the flavor required. In addition to fluid milk products, the production plant processes products made from milk, such as yogurt, cream, butter, and cheeses.

ABSTRACT In this paper, results of plant capacity analysis made to a dairy company are reported. This enterprise is one of the best-positioned companies in production and distribution of milk products in México. The enterprise has only one production plant and seven distribution centers. Because of the enterprise does not cover the client demand, the Planning Department planned to buy more equipment, in order to increase the production rate. This analysis was performed to determine the plant capacity, in order to know the quantity of additional equipment that production plant will need. An annual increasing rate was considered in the study, which was calculated with data of two years ago. In this work, the Monte Carlo method was applied to carry out the simulation of the production plant processes, and ProModel software was used to implement the simulation model. Keywords: production plant, plant capacity analysis, Monte Carlo method, computer simulation 1. INTRODUCTION Plant capacity means the maximum quantity that can be produced by time unit in the plant with the existing equipment. (Fare, Grosskopf, and Kokkelenberg 1989) The knowledge about the production plant capacity is very important because it defines the competitive limits of the enterprise, i.e. the plant capacity sets the response rate, the costs structure, the composition of the personnel, and the general strategy for inventory. If the plant capacity is inadequate to satisfy the market demand, a company could lose its clients. On the other hand, if the plant capacity is excessive, probably the company will reduce the prices of its products to stimulate the demand, underutilize personnel, keep overstocked, and produce other products, less profitable, in order to be in operation.

1.2. Plant capacity concepts The aim of this work is focused in the plant capacity analysis; hence, following paragraphs presents some plant capacity concepts. (Blackstone 1989) 1. Design Capacity (DC): Is the maximum possible production rate in a process, given the current designs of the product, mixing, operation policies, human resources, plant installations, and equipment. 2. Effective Capacity (EC): Is the maximum production rate that can be obtained in a reasonable way, taking into account preventive maintenance times, setup changes, and production system limitations.

690

3. Real Capacity (RC): Is the effective production rate achieved in the process. Normally, it is a time function and it changes constantly. RC is affected by the equipment wear, wastes and reworks, limited machinery assembly, employee’s absenteeism, inadequate production master planning, and other similar factors that contribute to decrease the real capacity rates.

2.2. Fluid milk processing area The processing area has two silos with a capacity of 100,000 liters each one, ten silos with a capacity of 150,000 liters each one and one silo with a capacity of 30,000 liters. The last one is only used to process cream. Before the milk is pumped to process silos, it passes for a clarification process. There are four clarifiers, two of them have a capacity of 25,000 liters per hour, and the other two have a capacity of 30,000 liters per hour. Fluid milk processing area is divided in two production lines: pasteurized milk line, and ultra pasteurized milk line.

As a relationship among these concepts, it can be seen that DC > EC > RC. Moreover, according to given concepts, some indicators can be obtained, such indicators are the utilization factor and the efficiency:

2.2.1. Pasteurized milk lines In order to produce pasteurized milk, these production lines take fluid milk from process silos; the milk is pumped to a homogenizer, and then, the milk is sent to two pasteurizers, each one with a tank with a capacity of 18,000 liters. After that, pasteurized milk is bottled by three bottle filling machines with a capacity of 18,000 liters per hour. Sometimes, another machine with a capacity of 9,000 liters is used in this process; this machine is shared with other production lines.

(1) (2) 1.3. Monte Carlo method basis Monte Carlo method is a generic form to call to mathematical procedures whose common feature is the use of random numbers and probability distributions, such as normal, exponential, uniform, beta, among others. It uses random variables defined in a finite dimensional space and the expectation value is calculated to find the approximated solution of a problem. (Kalos M.H., et. al., 2008) Monte Carlo method is widely used, because it can be applied to solve stochastic problems, or those that can be set out in a stochastic way.

2.2.2. Ultra pasteurized milk lines The production of ultra pasteurized milk uses seven lines. Five lines are used to process milk in presentation of one liter, and the other two are used to process in presentation of 250 ml. The use of the production line depends on the type of product demanded. These lines produce whole milk, lacto free milk, light milk, cholesterol free milk, and flavored milk, among others. Each line has between two or four bottle filling machines, with their respective capacities. Table 1 shows the capacities of ultra pasteurized milk lines.

2.

DESCRIPTION OF PRODUCTION PLANT PROCESSES The production plant is divided in three main areas: raw milk reception area, fluid milk processing area, and milk derivative processing area.

Table 1: Ultra Pasteurized Lines Capacities. Line Cap Product Bottle Cap (*) presentation filling (*) machine I 24 1 liter b1 12 b2 12 II 24 1 liter b3 6 b4 6 b5 6 b6 6 III 24 1 liter b7 6 b8 6 b9 6 b10 6 IV 16 1 liter b11 6 b12 6 V 30 1 liter b13 12 b14 12 b15 6

2.1. Raw milk reception area This area receives raw milk from dairy farms that belongs to shareholders of the company. Raw milk is transported in tankers with capacity of 25,000 liters. Before the raw milk is received, it is analyzed by the Quality Control Department in order to determine the quality of the product. There are six reception lines, where each line pumps milk to 40,000 liters per hour, toward one of the four reception silos, with capacity of 150,000 liters each one. One of them stores milk used to produce milk derivatives. In the pumping process, the milk passes through deareators, filters, and coolers. The capacity of these equipments is 40,000 liters per hour. Moreover, there is one reception line for cream, which pumps cream to 15,000 liters per hour. The cream is pumped into two tanks, with a capacity of 40,000 liters each one.

691

Plant III-A: Processes a type of yogurt mixed with fruits. Moreover, this plant is shared with pasteurized milk line.

Table 1: Ultra Pasteurized Lines Capacities. (Cont.) Line Cap Product Bottle Cap (*) presentation filling (*) machine VI 7 250 ml b16 1.5 b17 1.5 b18 1.875 b19 1.5 VII 10 250 ml b20 5 b21 5

3. PROBABILITY DISTRIBUTIONS In order to take into account the seasonality of the system, the data provide by the company were analyzed for every month. The company had a set of data from three years. The modeling of real system started from the arrivals of tankers to reception area. Inter arrival times were analyzed, and the probability distributions are shown in table 3.

*: in thousands of liters per hour.

2.3. Milk derivative processing area This area is divided in two subareas: cream area and yogurt area.

Table 3: Probability Distributions for Inter Arrival times of tankers. Month Probability distribution (in minutes) January E(18.4805) February E(18.46) March E(18.4804) April E(19.39) May E(18.7) June E(18.15) July E(18.49) August E(18.49) September E(18.4) October E(18.2) November E(18.52) December E(19.476)

2.3.1. Cream area Cream stored in the two cream reception tanks is pumped toward to three tanks used for a standardization process; these tanks have a capacity of 9,000, 9,000, and 14,000 liters per hour, respectively. Then, the cream is pumped to pasteurization and homogenization process. There are three pasteurizers, with a capacity of 7,000, 4,000, and 7,000 liters per hour, respectively. Every pasteurizer is connected to its homogenizer, where homogenizer speed is synchronized with the pasteurizer speed. After that, the cream is pumped into four tanks, with a capacity of 5,000 liters each one. The first tank is used to bottle cream manually in a presentation of four liters. The second, third, and fourth tank, are alternatively used to bottle cream in three bottled lines. Table 2 shows the characteristics of these lines. Line c1 c2 c3

The next variable to analyze was the milk contents of tankers, although the capacity of tankers is 25,000 liters, its real content is different. Table 4 shows the obtained results. Table 4: Probability Distributions for Milk Contents of Tankers. Month Probability distribution (in thousands of liters) January Normal(22.7500, 3.0000) February Normal(22.2500, 2.6800) March Normal(22.2500, 2.5000) April Normal(22.5710, 2.5000) May Normal(22.1140,2.5000) June Normal(22.6630, 2.7000) July Normal(22.3120, 2.3000) August Normal(22.9000, 2.2500) September Normal(22.7500, 2.8000) October Normal(22.3500, 2.7990) November Normal(22.4190, 2.9250) December Normal(22.7500, 2.4000)

Table 2: Cream Bottled Lines Source Capacity Presentation tank (*) 2, 3, 4 10 450 ml 900 ml 2, 3, 4 10 450 ml 2, 3, 4 10 200 ml

*: in thousands of liters per hour.

2.3.2. Yogurt area Yogurt process is divided according to product presentation. There are four lines in this area: Plant I, Plant II, Plant III, and Plant III-A. Raw milk stored in one of the four reception silos is used to process milk derivatives. Milk is pumped from this silo to three standardization tanks, then, milk is pumped to one of the four plants depending on the product presentation: Plant I: Processes drinking yogurt and fruit yogurt. Plant II: Processes yogurt with cereal, creamy yogurt, and whipped yogurt. Plant III: Processes whipped yogurt, drinking yogurt and fruit yogurt.

Production Master Planner provided us data about weekly production for every product. We obtained the corresponding probability distribution for every product in every month of the year. Fitted distributions for some products in January month are shown in table 5.

692

With the statistical analysis, we found an annual growth rate of 6.16%.

Table 5: Probability distributions for product depending on client demand. Product Probability distribution (in thousands of liters) Whole milk 1 liter LogNormal(229537.93, 69776.62) Light milk 1 liter Weibull(10.119, 36666.36) Strawberry flavored Normal(63342.21, 16008) milk 1 liter Drinking yogurt Triangular(25967.76, 250 ml 40661.22, 221153) strawberry-coconut Drinking yogurt LogNormal(130621.49,475 250 ml 59.78) pineapple-coconut Ultra pasteurized Triangular(0.65235, milk with fruits 0.65235, 35401.05) 250 ml strawberry Ultra pasteurized Triangular(0.00276, milk with fruits 0.00276, 24286.53) 250 ml mango Yogurt with cereal Uniform(1775.45, 150 ml 3791.46) strawberry-nut Yogurt with cereal Triangular(0.02711, 150 ml peach-nut 0.02717, 11662.38278) Cream 200 ml Weibull(32.022, 1467150.23) Cream 450 ml Weibull(7.0071, 858032.40) Yogurt 150 g Weibull(9.970, 201340) strawberry Yogurt 150 g peach Normal(831686, 182164.92)

4. COMPUTER SIMULATION The simulation software used to carry out the simulation model was ProModel. ProModel Simulation Software provides tools to model and simulate manufacturing process. It has a graphical interface, where the modeler can define entities, locations, processing, arrivals, resources, etc. (Harrel, Ghosh, and Bowden 2003; García, García, and Cárdenas 2006). The items listed below must be identified from the real system: Locations: are places used as servers, where entities are processed. Entities: are dynamic objects that are served by locations. Processing: model policies are defined in this part. Resources: used to transport entities among locations, like forklifts, workers, etc. Arrivals: used to define inter arrival time, it could be a constant value or a probabilistic distribution. Attributes: used to add values to entities. Variables: used to save data computed during simulation execution. Subroutines: used to define procedures in order to improve the software functionality. In the following subsections the simulation model is described. 4.1. Construction of the simulation model In this phase of the project, all the elements of the real system that are involved in the product processing were identified, since raw milk arrivals until the bottling filling machines. Because of fluid milk and cream flows are continuous variables, we considered one entity of milk or cream as 1000 liters of milk or cream, respectively.

Figure 1 shows the graphic for the probabilistic distribution fitted for the data corresponding to whole milk in presentation of 1 liter.

4.1.1. Raw milk reception area Arrival times for tanker were shown in table 3, and the tanker contents in table 4. Tankers were defined as resources, and the reception silos, valves, and pipes were defined as locations. See figure 2.

Raw milk and entities 1000 liters

Figure 2: Tankers downloading raw milk. Raw milk is pumped to reception silos, numbered as 29, 30, 31 and 40. The flow of the raw milk is shown in figure 3; it follows the description given in subsection 2.1.

Figure 1: Probabilistic distribution fitted for whole milk of 1 liter.

693

5. RESULTS AND CONCLUSIONS The simulation scenarios were executed 100 times and the gathered statistics are summarized in the following tables. Table 6 shows comparative data about the reception area. In order to get the Efficiency Capacity, time for clean the silos and maintenance time are considered. The table includes Design Capacity (DC), Effective Capacity (EC), Real Capacity (RC), Utilization factor (UF), and Efficiency (Ef). DC, EC, and RC are denoted in thousands of liters.

Valves control

Figure 3: Raw milk pumped toward reception silos (29, 30, 31, and 40).

Table 6: Raw Milk Reception Statistics DC EC RC UF Ef Valve 6,720 5,586 1,803 26.84% 32.28% 1 Valve 6,720 5,586 2,195 32.68% 39.30% 2 Valve 6,720 5,586 2,777 41.34% 49.72% 3 Valve 6,720 5,586 2,027 30.18% 36.30% 4 Valve 6,720 5,586 1,874 27.90% 33.56% 5 Valve 6,720 5,586 1,384 20.60% 24.78% 6 Silo 150 150 56.68 37.79% 37.79% 29 Silo 150 150 57.85 38.57% 38.57% 30 Silo 150 150 63.48 42.32% 42.32% 31 Silo 150 150 73.44 48.96% 48.96% 40

4.1.2. Fluid milk processing area Silos, clarifiers, pumps, pipes, homogenizers, pasteurizers, ultra pasteurizers, and bottle filling machines were defined as locations, with their respective processing time. Figure 4 shows a part of the pasteurized milk area.

Exit to warehouse

Data gathered from fluid milk pasteurization area are shown in table 7.

Figure 4: Pasteurized milk process. 4.1.3. Milk derivative processing area Equipment installed in Cream area and Yogurt area were defined as locations, taking into account their respective processing time. Figure 5 shows the cream homogenization process, and the four storage tanks.

Table 7: Fluid Milk Processing Statistics DC EC RC UF Ef Pasteu 7,560 6,331 2,545 33.68% 40.21% rized milk 21,903 18,343 Ultra 11,048 50.44% 60.23% pasteu rized milk Finally, data obtained from milk derivative processing area are shown in table 8. Table 8: Milk Derivative Processing Statistics DC EC RC UF Ef Cream 1,860 1,626 973 57.95% 59.86% Yogurt 6,762 6,191 2,481 36.70% 40.08% Tables 6, 7, and 8 denote the behavior of the real system, and the capacity that the production plant

Figure 4: Cream homogenization process.

694

provides. Nevertheless, it can be seen that the installed equipment is not used at all. The main obstacle to improve the usage of production lines is the harvesting system, because it is only performed in the morning. Veterinarians of the company declare that it is possible to set a new raw milk harvesting system, where the tankers go to dairy farms twice a day. On the other hand, a good strategy in the elaboration of the Production Master Planning can reduce time wasted in setup changes, and, in consequence, to improve the efficiency of production lines.

dynamic systems and their applications to stochastic manufacturing systems using emergent techniques.

REFERENCES Fare, R., Grosskopf, S., Kokkelenberg, E.C., 1989. Measuring plant capacity, utilization and technical change: a nonparametric approach. International economic review, 30 (3), 655–666. Blackstone, W.H.Jr., 1989. Capacity Management., Cincinnati, OH: South-Western Banks, J., Carson, J.S., Nelson, B.L., and Nicol, D.M., 2005. Discrete-Event System Simulation. USA: Prentice-Hall. Kalos, M.H., Whitlock P.A., 2008. Monte Carlo Methods. : Wiley-VCH. Harrel, C.R., Ghosh, B.K., Bowden, R.O., 2003. Simulation using ProModel. : McGraw-Hill Science/Engineering/Math. García, D.E., García, R.H., Cárdenas, B.L.E., 2006. Análisis de sistemas con ProModel. México:Prentice-Hall. Zeigler, B., Kim, T.G., and Praehofer, H., 2000. Theory of Modeling and Simulation. New York: Academic Press, New York.

Norberto Hernandez-Romero. He received the M.S. degree from Department of Electrical Engineering, Laguna Technological Institute at México, in 2001 and the Ph. D. from Autonomous University of Hidalgo State at México in 2009. Currently, he is a professor of the Advanced Research in Industrial Engineering Center at the Autonomous University of Hidalgo State at Pachuca, Hidalgo, México. His current research interests include system identification, feedback control design, genetic algorithms, fuzzy logic, neural network and its applications.

Juan Carlos Seck-Tuoh-Mora. He received the M.S. and Ph.D. degrees in electrical engineering (option: Computing) from the Research and Advanced Studies Center of the National Polytechnic Institute at Mexico, in 1999 and 2002, respectively. Currently, he is a Professor of the Advanced Research in Industrial Engineering Center at the Autonomous University of Hidalgo State at Pachuca, Hidalgo, México. His current research interests include cellular automata theory and its applications, evolutionary computing and simulation.

Isaias Simon-Marmolejo. He received the M.S. degree in Science in Industrial Engineering, graduated from the Autonomous University of Hidalgo State during the period 2007 to 2009. Currently, he works as a research professor in the School of Ciudad Sahagun Autonomous University of Hidalgo State in Tepeapulco, Hidalgo, Mexico and professor at the Technologic Institute of Pachuca in Pachuca of Soto, Hidalgo, Mexico. Years of experience and collaborative relationships and lines of research include Operations Research, Statistical Analysis of Discrete Event Simulation, Logistics and Systems Engineering.

AUTHORS BIOGRAPHY Joselito Medina-Marin. He received the M.S. and Ph.D. degrees in electrical engineering from the Research and Advanced Studies Center of the National Polytechnic Institute at Mexico, in 2002 and 2005, respectively. Currently, he is a Professor of the Advanced Research in Industrial Engineering Center at the Autonomous University of Hidalgo State at Pachuca, Hidalgo, México. His current research interests include Petri net theory and its applications, active databases, simulation, and programming languages. Gilberto Perez-Lechuga. He holds a Master Degree in Science, in Operations Research by the National Polytechnic Institute at the Mexico City. Later, he obtained the Engineering Doctor Degree in Operations Research (Stochastic Optimization) by the National Autonomous University of Mexico. His current research is directed to the modeling and optimization of complex

695

GPGPU PROGRAMMING AND CELLULAR AUTOMATA: IMPLEMENTATION OF THE SCIARA LAVA FLOW SIMULATION CODE Giuseppe Filippone(a) William Spataro(b), Giuseppe Spingola(c), Donato D’Ambrosio(d), Rocco Rongo(e), Giovanni Perna(f) , Salvatore Di Gregorio(g) (a) (b) (c) (d) (f) (g) (e)

Department of Mathematics and HPCC, University of Calabria, Italy Department of Earth Sciences and HPCC, University of Calabria, Italy

(c)

(d)

[email protected], (a)[email protected], (b)[email protected], [email protected], (e)[email protected], (f)[email protected], (g)[email protected], power of these devices. In fact, the performance of a GPGPU program that does not sufficiently exploit a GPU’s capabilities can often be worse than that of a simple sequential one running on a CPU, such as when data transfer from main memory to video memory results crucial. Nevertheless, GPU applications to the important field of Computational Fluid Dynamics (CFD) are increasing both for quantity and quality among the Scientific Community (e.g., Tolke and Krafczyka 2008, Zuo and Chen 2010). Among the different methodologies used for modelling geological processes, such as numerical analysis, high order difference approximations and finite differences, Cellular Automata (CA) (von Neumann 1966) has proven to be particularly suitable when the behaviour of the system to be modelled can be described in terms of local interactions. Originally introduced by von Neumann in the 1950s to study selfreproduction issues, CA are discrete computational models widely utilized for modeling and simulating complex systems. Well known examples are the Lattice Gas Automata and Lattice Boltzmann models (Succi 2004), which are particularly suitable for modelling fluid dynamics at a microscopic level of description. However, many complex phenomena (e.g. landslides or lava flows) are difficult to be modeled at such scale, as they generally evolve on large areas, thus needing a macroscopic level of description. Moreover, since they may also be difficult to be modelled through standard approaches, such as differential equations Macroscopic Cellular Automata (MCA) (Di Gregorio and Serra 1999) can represent a valid alternative. Several successful attempts have been carried out regarding solutions for parallelizing MCA simulation models (e.g., D’Ambrosio and Spataro 2007). In this research context, the CAMELot virtual laboratory and the libAuToti scientific library represent valid solutions for implementing and automatically parallelizing MCA models on distributed memory machines while, for shared memory architectures, some effective OpenMP parallelizations have been implemented for CA-like models, such as for fire spread simulations, Lattice Boltzmann models or lava flow modeling (Oliverio et al. 2011). However, few examples of GPGPU applications for CA-like models do exist (Tolke 2008) and to our knowledge, none regarding the MCA approach. This paper presents a implementation of a

ABSTRACT This paper presents an efficient implementation of a well-known computational model for simulating lava flows on Graphical Processing Units (GPU) using the Compute Unified Device Architecture (CUDA) interface developed by NVIDIA. GPUs are specifically designated for efficiently processing graphic datasets. However, recently, they are also being exploited for achieving exceptional computational results even for applications not directly connected with the Computer Graphics field. We here show an implementation of the SCIARA Cellular Automata model for simulating lava flows on graphic processors using CUDA. Carried out experiments show that significant performance improvements are achieved, over a factor of 100, depending on the problem size, adopted device and type of performed memory optimization, confirming how graphics hardware can represent a valid solution for the implementation for Cellular Automata models. Keywords: Cellular Automata, Lava flows simulation, GPGPU programming, CUDA. 1. INTRODUCTION High Performance Computing (HPC) (Grama et al. 2003) adopts numerical simulations as an instrument for solving complex equation systems which rule the dynamics of complex systems as, for instance, a lava flow or a forest fire. In recent years, Parallel Computing has undergone a significant revolution with the introduction of GPGPU technology (General-Purpose computing on Graphics Processing Units), a technique that uses the graphics card processor (the GPU – Graphics Processing Unit) for purposes other than graphics. Currently, GPUs outperform CPUs on floating point performance and memory bandwidth, both by a factor of roughly 100. As a confirmation of the increasing trend in the power of GPUs, leading companies such as Intel have already integrated GPUs into their latest products to better exploit the capabilities of their devices, such as in some releases of the Core i5 and Core i7 processing units. Although the extreme processing power of graphic processors may be used for general purpose computations, a GPU may not be suitable for every computational problem: only a parallel program that results suitable and optimized for GPU architectures can fully take advantage of the

696

well-known, reliable and efficient MCA model widely adopted for lava flow risk assessment, namely the SCIARA model (Rongo et al. 2008), in GPGPU environments. Tests performed on two types of GPU hardware, a Geforce GT 330M graphic card and a Tesla C1060 computing processor, have shown the validity of this kind of approach. In the following sections, after a brief description of the basic version of the SCIARA MCA model for lava flows, a quick overview of GPGPU paradigm together with the CUDA framework is presented. Subsequently, the specific model implementation and performance analysis referred to benchmark simulations of a real event and different CA spaces are reported, while conclusions and possible outlooks are shown at the end of the paper.

Refer to Di Gregorio and Serra (1999) for a complete description of the algorithm, besides theorems and applications. 2.1. The MCA lava flow model SCIARA SCIARA is a family of bi-dimensional MCA lava flow models, successfully applied to the simulation of many real cases, such as the 2001 Mt. Etna (Italy) Nicolosi lava flow (Crisci et al. 2004), the 1991 Valle del Bove (Italy) lava event (Barca et al. 1994) which occurred on the same volcano and employed for risk mitigation (D’Ambrosio et al. 2006). In this work, the basic version of SCIARA (Barca et al. 1993) was considered and its application to the 2001 Nicolosi event and to benchmark grids shown. SCIARA considers the surface over which the phenomenon evolves as subdivided in square cells of uniform size. Each cell changes its state by means of the transition function, which takes as input the state of the cells belonging to the von Neumann neighbourhood. It is formally defined as SCIARA = < R, X, Q , P, V >

2.

CELLULAR AUTOMATA AND THE SCIARA MODEL FOR LAVA FLOW SIMULATION As previously stated, CA are dynamical systems, discrete in space and time. They can be thought as a regular n-dimensional lattice of sites or, equivalently, as an n-dimensional space (called cellular space) partitioned in cells of uniform size (e.g. square or hexagonal for n=2), each one embedding an identical finite automaton. The cell state changes by means of the finite automaton transition function, which defines local rules of evolution for the system, and is applied to each cell of the CA space at discrete time steps. The states of neighbouring cells (which usually includes the central cell) constitute the cell input. The CA initial configuration is defined by the finite automata states at time t=0. The global behaviour of the system emerges, step by step, as a consequence of the simultaneous application of the transition function to each cell of the cellular space. When dealing with the modelling of spatial extended dynamical systems, MCA can represent a valid choice especially if their dynamics can be described in terms of local interaction at macroscopic level. Well known examples of successful applications of MCA include the simulation of lava (Crisci et al. 2004) and debris flows (Di Gregorio et al. 1999), forest fires (Trunfio 2004), agent based social processes (Di Gregorio et al. 2001) and highway traffic (Di Gregorio et al. 2008), besides many others. By extending the classic definition of Homogeneous CA, MCA facilitate the definition of several aspects considered relevant for the correct simulation of the complex systems to be modelled. In particular, MCA provide the possibility to “decompose” the CA cell state in “substates” and to allow the definition of “global parameters”. Moreover, the dynamics of MCA models (especially those developed for the simulation of complex macroscopic physical systems such as debris or lava flows) is often “guided” by the “Minimisation Algorithm of the Differences” (cf. Di Gregorio and Serra 1999), which translates in algorithmic terms the general principle for which natural systems leads towards a situation of equilibrium.

where: x

x

x

R is the set of points, with integer coordinates, which defines the 2-dimensional cellular space over which the phenomenon evolves. The generic cell in R is individuated by means of a couple of integer coordinates (i, j  ZKHUH  ” i < imax and ”j < jmax. X = {(0,0), (0, -1), (1, 0), (-1, 0), (0, 1)} is the so called von Neumann neighbourhood relation, a geometrical pattern which identifies the cells influencing the state transition of the central cell. Q is the set of cell states; it is subdivided in the following substates: -

Qz is the set of values representing the topographic altitude (m); Qh is the set of values representing the lava thickness (m); QT is the set of values representing the lava temperature (K°); Qo5 are the sets of values representing the lava outflows from the central cell to the neighbouring ones (m).

The Cartesian product of the substates defines the overall set of state Q: Q = Qz × Qh × QT ×Qo5 x

P is set of global parameters ruling the CA dynamics: -

697

PT={Tvent, Tsol, Tint}, the subset of parameters ruling lava viscosity, which specify the temperature of lava at the vents, at solidification and the “intermediate” temperature (needed for computing lava adherence), respectively;

-

x

Pa={avent, asol, aint}, the subset of parameters which specify the values of adherence of lava at the vents, at solidification and at the “intermediate” temperature, respectively; pc, the cooling parameter, ruling the temperature drop due to irradiation; pr, the relaxation rate parameter, which affects the size of outflows.

per Second), while an Intel Core i7 has barely 52 GFLOPS. In addition, the most interesting aspect still is the elevated parallelism that a GPU permits. For instance, the NVIDIA GeForce 8800 GTX has 16 multiprocessors each with 8 processors for a total of 128 basic cores, while a standard multi-core CPU has few, though highly-functional, cores. Another motivation of GPUs increasing utilization as parallel architecture regards costs. Until a few years ago, in order to have the corresponding computing power of a medium range GPU of today (which costs approximately a few hundred Euros), it was necessary to spend tens of thousands of Euros. Thus, GPGPU has not only led to a drastic reduction of computation time, but also to significant cost savings. Summarizing, it is not misleading to affirm that the computational power of GPUs has exceeded that of PC-based CPUs by more than one order of magnitude while being available for a comparable price. In the last years, NVIDIA has launched a new product line called Tesla, which is specifically designed for High Performance Computing. Supported on Windows and Linux Operating systems, NVIDIA CUDA technology (NVIDIA CUDA 2011a) permits software development of applications by adopting the standard C language, libraries and drivers. In CUDA, threads can access different memory locations during execution. Each thread has its own private memory, each block has a (limited) shared memory that is visible to all threads in the same block and finally all threads have access to global memory. The CUDA programming model provides three key abstractions: the hierarchy with which the threads are organized, the memory organization and the functions that are executed in parallel, called kernels. These abstractions allow the programmer to partition the problem into many sub-problems that can be handled and resolved individually.

V : Q5ĺ Q is the deterministic cell transition function. It is composed by four “elementary processes”, briefly described in the following: - Outflows computation (V). It determines the outflows from the central cell to the neighbouring ones by applying the minimisation algorithm of the differences; note that the amount of lava which cannot leave the cell, due to the effect of viscosity, is previously computed in terms of adherence. Parameters involved in this elementary process are: PT and Pa. - Lava thickness computation (V). It determines the value of lava thickness by considering the mass exchange among the cells. No parameters are involved in this elementary process. - Temperature computation (V). It determines the lava temperature by considering the temperatures of incoming flows and the effect of thermal energy loss due to surface irradiation. The only parameter involved in this elementary process is pc. - Solidification (V). It determines the lava solidification when temperature drops below a given threshold, defined by the parameter Tsol.

3. GPU AND GPGPU PROGRAMMING As alternative to standard parallel architecture, the term GPGPU (General-Purpose computing on Graphics Processing Units) refers to the use of the card processor (the GPU) as a parallel device for purposes other than graphic elaboration. In recent years, mainly due to the stimulus given by the increasingly demanding performance of gaming and graphics applications in general, graphic cards have undergone a huge technological evolution, giving rise to highly parallel devices, characterized by a multithreaded and multicore architecture and with very fast and large memories. A GPU can be seen as a computing device that is capable of executing an elevated number of independent threads in parallel. In general, a GPU consists in a number (e.g., 16) of SIMD (Single Instruction, Multiple Data) multiprocessors with a limited number of floating-point processors that access a common shared-memory within the multiprocessor. To better understand the enormous potential of GPUs, some comparisons with the CPU are noticeable: a medium-performance GPU (e.g. the NVIDIA Geforce GT200 family) is able to perform nearly 1000 GFLOPS (Giga Floating Point Operations

3.1. CUDA Threads and Kernels A GPU can be seen as a computing device that is capable of executing an elevated number of independent threads in parallel. In addition, it can be thought as an additional coprocessor of the main CPU (called in the CUDA context Host). In a typical GPU application, data-parallel like portions of the main application are carried out on the device by calling a function (called kernel) that is executed by many threads. Host and device have their own separate DRAM memories, and data is usually copied from one DRAM to the other by means of optimized API calls. CUDA threads can cooperate together by sharing a common fast shared-memory (usually 16KB), eventually synchronizing in some points of the kernel, within a so-called thread-block, where each thread is identified by its thread ID. In order to better exploit the GPU, a thread block usually contains from 64 up to 512 threads, defined as three-dimensional array of type dim3 (containing three integers defining each dimension). A thread can be referred within a block by means of the built-in global variable threadIdx.

698

While the number of threads within a block is limited, it is possible to launch kernels with a larger total number of threads by batching together blocks of threads, by means of a grid of blocks, usually defined as a twodimensional array, also of type dim3 (with the third component set to 1). In this case, however, thread cooperation is reduced since threads that belong to different blocks do not share the same memory and thus cannot synchronize and communicate with each other. As for threads, a built-in global variable, blockIdx, can be used for accessing the block index within the grid. Currently, the maximum number of blocks is 65535 in each dimension. Threads in a block are synchronized by calling the syncthreads() function: once all threads have reached this point, execution resumes normally. As previously reported, one of the fundamental concepts in CUDA is the kernel. This is nothing but a C function, which once invoked is performed in parallel by all threads that the programmer has defined. To define a kernel, the programmer uses the __global__ qualifier before the definition of the function. This function can be executed only by the device and can be only called by the host. To define the dimension of the grid and blocks on which the kernel will be launched on, the user must specify an expression of the form >, placed between the kernel name and argument list. What follows is a classic pattern of a CUDA application: x x x x x

global memory space of the device (with __device__) or variables that reside in the shared memory space (with __shared__) that are thus accessible only from threads within a block. Typical latency for accessing global memory variables is 200300 clock cycles, compared with only 2-3 clock cycles for shared memory locations. For this reason, to improve performances variable accesses should be carried out in the shared memory rather than global memory, wherever possible. However, each variable or data structure allocated in shared memory must first be initialized in the global memory, and afterwards transferred in the shared one (NVIDIA CUDA 2011b). This means that to copy data in the shared memory, global memory access must be first performed. So, the more his type of data is accessed, the more convenient is to use this type of memory: so, for few accesses it is evident that shared memory is not convenient to use. As a consequence, a preliminary analysis of data access of the considered algorithm should be performed in order to evaluate the tradeoff, and thus, convenience of using shared memory. 4.

IMPLEMENTATION OF THE SCIARA MODEL AND EXPERIMENT RESULTS As previously stated, Cellular Automata models, such as SCIARA, can be straightforwardly implemented on parallel computers due to their underlying parallel nature. In fact, since Cellular Automata methods require only next neighbor interaction, they are very suitable and can be efficiently implemented even on GPUs. In literature, to our knowledge, no examples of Macroscopic Cellular Automata modeling with GPUs are found, while some interesting CA-like implementations, such as Lattice Boltzmann kernels, are more frequent (e.g., Tolke 2008; Kuznik et al. 2010). In this work, two different implementations of the SCIARA lava flow computational model were carried out, a first straightforward version which uses only global memory for the entire CA space partitioning and a second, but more performing one, which adopts (also) shared memory for CA space substate allocation. What follows is an excerpt of the core of the general CUDA algorithm (cf. Section 2.1):

Allocation and initialization of data structures in RAM memory; Allocation of data structures in the device and transfer of data from RAM to the memory of the device; Definition of the block and thread grids; Performing one or more kernel; Transferring of data from the device memory to Host memory.

Eventually, it must be pointed out that a typical CUDA application has parts that are normally performed in a serial fashion, and other parts that are performed in parallel.

// CA loop for(int step=0; step< Nstep; step++) {

3.2. Memory hierarchy In CUDA, threads can access different memory locations during execution. Each thread has its own private memory, each block has a (limited) shared memory that is visible to all threads in the same block, and finally all threads have access to global memory. In addition to these memory types, two other read-only, fast on-chip memory types can be defined: texture memory and constant memory. As expected, memory usage is crucial for the performance. For example, the shared memory is much faster than the global memory and the use of one rather than the other can dramatically increase or decrease performance. By adopting variable type qualifiers, the programmer can define variables that reside in the

// add lava at craters crater (Aread,Awrite); // s1 calc_flows(Aread,Awrite); // s2 calc_width (Aread,Awrite); // s3 calc_temp(Aread,Awrite); // s4 calc_quote (Aread,Awrite);

// swap matrixes copy(Awrite,N,Aread,Substat_N); } cudaMemcpy(A, Aread, size, cudaMemcpyDeviceToHost ); // copy data to Host }

699

In the time loop four basic kernels, calc_flows, calc_width, calc_temp and calc_quote are launched corresponding to the four elementary processes of SCIARA, V1, V2, V3 and V4, respectively, as described in Spingola et al. (2008). The crater() kernel refers to the crater cell(s), which is obviously invoked on a smaller grid that the previous ones. The model was implemented by adopting a system of double matrixes for the CA space representation, one (Aread) for reading cell neighbor substates and a second (Awrite) for writing the new substate value. This choice has proven to be efficient, since it allows to separate the substates reading phase from the update phase, after the application of the transition function, thus ensuring data integrity and consistency in a given step of the simulation. After applying the transition function to all the cell space, the main matrix must be updated, replacing values with the corresponding support matrix ones (swap matrixes phase). In this implementation, a CA step is simulated by more logical substeps where, after crater cells are updated (by means of the crater), lava outflows are calculated according to the V1 elementary process. When all outflows are computed, and therefore all outflow substates are consistent, the actual distribution takes place, producing the new value of the quantity of lava in each cell of the CA. Subsequently, each cell reads from a neighbour cell the associated outflow substate corresponding to the quantity of inflowing lava V2 elementary process). In this phase, the V3 and V4 elementary processes are applied to the new quantity of lava of the cell. At the end of the CA loop, data is copied back to the Host memory by the cudaMemcpy function. Regarding the specific implementation, the first thing to decide on is what thread mapping should be adopted to better exploit the fine-grain parallelism of the CUDA architecture. For example, one might consider using a thread for each row or each column, as occurs in a typical data-parallel implementation (e.g., Oliverio et al. 2011). However, when working in CUDA with arrays, the most widely adopted technique is to match each cell of the array with a thread (e.g., Tolke 2008). The number of threads per block should be a multiple of 32 threads, because this provides optimal computing efficiency (NVIDIA CUDA 2011b) and thus we have chosen to build blocks of size 32 × 16, corresponding to the maximum value (512) of number of threads permitted for each block. What follows is an excerpt for defining the grid of blocks that was considered for SCIARA:

... // invoke kernel functions kernel(...);

… Once that the grid of blocks (and threads) were defined in this simple manner, kernels are managed so that each cell (i, j) of SCIARA is associated to each thread (i, j). This is simply done, for each invoked kernel (i.e., calc_flows, calc_width, calc_temp and calc_quote), by associating each row and column of the CA with the corresponding thread as in this simple scheme: __global__ void kernel(...) { int col = threadIdx.x; int row = threadIdx.y;

blockIdx.x

*

blockDim.x

+

blockIdx.y

*

blockDim.y

+

// memory allocation (shared, global, etc) ... /** transition function for cell[row][col] ** ... }

5. TESTS AND PERFORMANCE RESULTS Two CUDA graphic devices were adopted for experiments: a NVIDIA high-end Tesla C1060 and a Geforce GT 330M graphic card. In particular, the Tesla computing processor has 240 processor cores, 4 GB global memory and high-bandwidth communication between CPU and GPU, whereas the less performing graphic card has 48 cores and 512 MB global memory. The sequential SCIARA reference version was implemented on a 2.66 GHz Intel Core i7 based desktop computer. The sequential CPU version is identical to the version that was developed for the GPUs, that is, no optimizations were adopted in the former version. In practice, at every step, the CA space array is scrolled and the transition function applied to each cell of the CA where lava is present. Many tests have been performed regarding both performance and verification of the accuracy of the results. Regarding performance tests, the best implementation has regarded a version which adopts a hybrid (shared/global) memory allocation. As known, access to a location in shared memory of each multiprocessor has a much lower latency than that carried out on the global device memory. On the other hand, an access to a shared-memory location necessary needs a first access to global memory (cf. Section 3.2). For this reason, an accurate analysis was carried out in determining how much memory access each thread does for each CA substate matrix. This investigation gave rise to a “hybrid'” memory access pattern, where shared memory allocation was adopted for those kernels accessing CA matrixes more than two times. For illustrative purposes, Figure 1 shows how

#define BLOCK_SIZE_X 32 #define BLOCK_SIZE_Y 16 ... int dimX; // CA x dimension int dimY; // CA y dimension dim3 dimBlock(BLOCK_SIZE_X, BLOCK_SIZE_Y); int n_blocks_x = dimX/dimBlock.x; int n_blocks_y = dimY/dimBlock.y; dim3 dimGrid(n_blocks_x, n_blocks_y);

700

shared memory is used in the context of our GPU implementation.

time reduction, significantly outperforming the CPU implementation up to 150× for large grid sizes. Other tests were also performed on a completely global memory version. In this case results, here omitted for brevity, have shown how the use of shared memory can improve performances up to 50%, with respect to the total global memory version. Eventually, to test if single-precision data can be considered sufficient for SCIARA simulations, tests were carried out on the 2001 lava flow event (15000 CA steps) and compared results produced by the GPU version with those produced by the CPU (sequential) version with single precision data (i.e., float type variables), and those produced still by the same GPU version against a double precision CPU implementation (i.e., double type variables). In each case, comparison results were satisfactory, since the areal extensions of simulations resulted the same, except for few errors of approximation in a limited number of cells. In particular, comparing the GPU version with the CPU single-precision version approximation differences at the third significant digit were only for 4% of cells, while differences were less for remaining cells. Differences were even minor compared to the previous case by considering the single precision GPU version and a CPU version which adopts double-precision variables.

Figure 1: Memory mapping of the CA space allocated in global memory with a portion of shared memory. Shaded areas represent portions of neighbouring block areas which need to be swapped at each CA step to ensure data consistency. A first test regarded the simulation of well-known and documented real lava flow event, the Mt. Etna Nicolosi event (Crisci et al. 2004) occurred in July, 2001. Table 1 (first row) reports the first results of tests carried out for this experiment, where the CA space is a 819 × 382 two-dimensional grid. The simulation was carried out for 15000 steps, considering one crater for lava flow emission. In order to further stress the efficiency of the GPU version, further benchmarks experiments were performed by considering four different hypothetical CA spaces, namely 5122, 10242, 20482 and 40962 grids, with cells representing inclined planes, with many craters located over the grid (cf. Table 1 - from second row).

6. CONCLUSIONS This paper reports the implementation of a Macroscopic Cellular Automata model using GPU architectures. As shown, the CUDA technology, in combination with the an efficient memory management, can produce a very efficient version of the SCIARA lava flow simulator. Although results are indeed already satisfactory, future developments can regard further improvements for both increasing performances and implementing more advanced MCA models. The results obtained in this work are to be considered positive and extremely encouraging. As confirmed by the increasing number of applications in the field of scientific computing in general, GPGPU programming represents a valid alternative to traditional microprocessors in high-performance computer systems of the future.

Table 1: Execution times of experiments (in seconds) carried out for evaluating the performance the GPU version of the SCIARA MCA lava-flow model on the considered hardware. The 819 × 382 matrix refers to the 2001 Mt. Etna event. Other grid dimensions refer to inclined planes. N/A (Not Available) data are due to device lack of memory capacity.

CA dim / Device 819×382 5122 10242 20482 40962

Performance results (in seconds) Intel i7 Geforce (sequential) 741 46 677 31.4 2716 99.1 11480 344.5 47410 N/A

ACKNOWLEDGMENTS This work was partially funded by the European Com-mission - European Social Fund (ESF) and by the Regione Calabria.

Tesla 11.8 5.6 21.4 81.1 307

REFERENCES Barca, D., Crisci, G.M., Di Gregorio, S., Nicoletta, F., 1993. Cellular automata methods for modelling lava flow: simulation of the 1986-1987 eruption, Mount Etna, Sicily. In: Kilburn, C.R.J., Luongo, G. (Eds.), Active lavas: monitoring and modelling. UCL Press, London, 12, 291-309. Barca, D., G.M. Crisci, Di Gregorio, S., Nicoletta, F. 1994. Cellular Automata for simulating lava

Timings reported for the considered GPU devices indicate their full suitability for parallelizing CA models. Even if applied to a simple MCA model, performance results show the incredible computational power of the considered GPUs in terms of execution

701

Flows: A method and examples of the Etnean eruptions. Transport Theory and Statistical Physics, 23, 195-232. Crisci, G.M., Di Gregorio, S. , Rongo, R., Spataro, W., 2004. The simulation model SCIARA: the 1991 and 2001 at Mount Etna. Journal of Vulcanology and Geothermal Research, 132, 253-267. D’Ambrosio, D., Rongo, R., Spataro, W., Avolio, M.V., Lupiano, V. , 2006. Lava Invasion Susceptibility Hazard Mapping Through Cellular Automata. In: S. El Yacoubi, B. Chopard, and S. Bandini (Eds.), ACRI 2006, Lecture Notes in Computer Science, 4173, Springer-Verlag, Berlin Heidelberg, 452461. D’Ambrosio, D., Spataro, W., 2007. Parallel evolutionary modelling of geological processes. Parallel Computing 33 (3), 186–212. Di Gregorio, S., Mele, F., Minei, G., 2001. Automi Cellulari Cognitivi. Simulazione di evacuazione, Proceedings “Input 2001”, Seconda Conferenza Nazionale Informatica Pianificazione Urbana e Territoriale, (in Italian) Democrazia e Tecnologia. Di Gregorio, S., Rongo, R., Siciliano, C., SorrisoValvo, M., Spataro, W., 1999. Mount Ontake landslide simulation by the cellular automata model SCIDDICA-3. Physics and Chemistry of the Earth, Part A, 24, 97-100. Di Gregorio, S., Serra, R., 1999. An empirical method for modelling and simulating some complex macroscopic phenomena by cellular automata. Fut. Gener. Comp. Syst., 16, 259–271. Di Gregorio, S., Umeton, R., Bicocchi, A., Evangelisti, A., Gonzalez, M., 2008. Highway Traffic Model Based on Cellular Automata: Preliminary Simulation Results with Congestion Pricing Considerations. Proceedings of 20th European Modeling & Simulation Symposium (EMSS), pp. 665-674. September 17-19, Campora S.G., CS, Italy. Grama, A., Karypis, G., Kumar, V., Gupta, A., 2003. An Introduction to Parallel Computing: Design and Analysis of Algorithms, Second Edition. USA: Addison Wesley. Kuznik, F., Obrecht, C., Rusaouen, G., Roux, J.J., 2010. LBM based flow simulation using GPU computing processor. Computers and Mathematics with Applications, 59, 2380–2392. NVIDIA CUDA C Programming Guide, 2011a. Available from: http://developer.download.nvidia.com/compute/cu da [accessed 27 June 2011] NVIDIA CUDA C Best Practices Guide, 2011b. Available from: http://developer.download.nvidia.com/compute/cu da [accessed 27 June 2011] Oliverio, M., Spataro, W. , D’Ambrosio, D. , Rongo, R., Spingola, G., Trunfio, G.A., 2011. OpenMP parallelization of the SCIARA Cellular Automata lava flow model: performance analysis on sharedmemory computers. Proceedings of International

Conference on Computational Science, ICCS 2011, Procedia Computer Science, 4, pp. 271-280. Rongo, R., Spataro, W., D'Ambrosio, D., Avolio, M.V., Trunfio, G.A., Di Gregorio, S., 2008. Lava flow hazard evaluation through cellular automata and genetic algorithms: an application to Mt Etna volcano. Fundamenta Informaticae, 87, 247–268. Spingola, G. , D’Ambrosio, D., Spataro, W., Rongo, R. Zito, G., 2008. Modeling Complex Natural Phenomena with the libAuToti Cellular Automata Library: An example of application to Lava Flows Simulation. Proceedings of International Conference on Parallel and Distributed Processing Tecniques and Applications, pp. 44-50. July 14-17, 2008, Las Vegas, Nevada, USA Succi, S, 2004. The Lattice Boltzmann Equation for Fluid Dynamics and Beyond, UK: Oxford University Press. Tolke, J., Krafczyka, M., 2008. TeraFLOP computing on a desktop PC with GPUs for 3D CFD. Int. Journ. of Comput. Fluid Dynamics, 22 (7), 443– 456. Tolke, J., 2008. Implementation of a lattice Boltzmann kernel using the compute unified device architecture developed by NVIDIA. Comput. Vis. Sci., 13 1, 29–39. Trunfio, G.A., 2004. Predicting Wildfire Spreading Through a Hexagonal Cellular Automata Model. In: P.M.A. Sloot, B. Chopard and A.G.Hoekstra (Eds.), ACRI 2004, LNCS 3305, Springer, Berlin, 2004, 725-734. von Neumann, J. (Edited and completed by A. Burks), 1966. Theory of self-reproducing automata. USA: University of Illinois Press. Zuo, W., Chen, Q., 2010. Fast and informative flow simulations in a building by using fast fluid dynamics model on graphics processing unit. Build. Envir., 45, 3, 747–757.

702

NEIGHBORHOOD CONCEPT FOR MODELING AN ADAPTIVE ROUTING IN WIRELESS SENSOR NETWORK Jan Nikodem(a) , Maciej Nikodem(a) , Ryszard Klempous(a) , Marek Woda(a) , Zenon Chaczko(b) (a)

Wrocław University of Technology, Poland University of Technology Sydney, NSW, Australia email: {jan.nikodem, maciej.nikodem, ryszard.klempous, marek.woda}@pwr.wroc.pl, [email protected] (b)

ABSTRACT This paper summarizes the research work of the Wireless Sensor Networks group at the Wrocław University of Technology. The group came up with an innovative technique that uses a number of relations to define activities and manage the network communication. The suggested solution, uses the concept of neighborhood as the primary spatial entity in order to demonstrate how the application of relations could support concurrence of both global and local perspectives. Additionally, a case study is provided to show the merits of the new adaptive routing technique for both semi-static and dynamic environments.

1. INTRODUCTION In this paper we present recent results of our work on the formalization of processes description in a distributed systems. We focus on routing in Wireless Sensor Network (WSN) because it is really distributed, moreover, routing as a type of activity, especially associated with information, is a basic processes investigated in WSN. Due to the spatial distribution of nodes and constrained resources, the operation of WSN sensors is focused on local activities and mutual communication. Consequently, our work concentrates on developing the formal methods and techniques necessary to model and evaluate situations in the network, decision processes and implement intelligent behavior while following the general outlines of the network activity. These requirements will entail the investigation of various methods in which intelligent systems could evaluate, interact and self-organize, both individually and cooperatively with other spatial explorers or while interacting with the environment. The Wireless Sensor Network poses stringent communication efficiency requirements in order to sustain its functionality. The dependability of such a network depends on the energy efficiency, node failures as well as the quality of the radio channel in a vicinity. Therefore, when working on routing algorithms we look for various adaptive solutions that are suitable for WSNs

2. RELATED WORKS There is a large number of research publications that consider communication activity in WSN’s that is related mainly to 703 clustering and routing problems. A number of authors have

proposed sensors’ self-configuring [2], self-management [3], [13], [14], adaptive clustering [2], [9], [16] or the concept of adjustable autonomy [6] to efficiently manage data packets in a network. On the other hand there are papers which discuss bio-inspired ideas and tend to isolate some aspects of the natural world for computer emulation. Authors [5] have shown that the communication topology of some biological, social and technological networks is neither completely regular nor completely random but stays somehow in between these two extreme cases. It is worth to mention papers [3], [14], [16] devoted to self-organizing protocols using both random and deterministic elements. Design challenges in building WSN structure can be described using different mapping functions. Consequently, in WSN literature several various models were proposed [5], [2], [9], [16]. These attempts are based on the representation of the network as the constellation of nodes connected with each other. In their research work, Cohn at al. [5] concentrated on using the regions as the primary spatial entities rather than the traditional mathematical and dimensionless points representing nodes. The authors proposed the concept of vague and crisp regions as a qualitative spatial representation and have argued that such a modification would allow for simplification and the use of standard mathematical topology tools. This approach is a basis for such popular concepts of segmentation in multi-hop networks as: region building and clusterization. The researchers who discussed these issues, have proposed various different methods to determine such structures and pointed out benefits and drawbacks of these approaches.

3. BASIC CONCEPT The approach presented in this paper is distinctly different from those mentioned in literature. We propose a relational attempt, based on set and relation theories. We consider three basic relations: subordination, tolerance and collision [6, 8]. These relations correspond to network activity, therefore the fourth relation (neighborhood), corresponding to WSN structure, is involved.

3.1. Traditional network partitioning Network partitioning is a well known idea [7] to solve large and complex problems by segmenting them into smaller and possibly simpler tasks. The most crucial element of such an attempt is to decide how to make a segmentation in

Figure 1. Two methods of WSN partitioning: clusters (left) and regions (right)

order to get a subproblems that can be solved efficiently and, what is more important, can be useful for finding a solution of the original problem. A commonly accepted idea for segmentation of the WSN structure is based on different mapping functions. Hence, in WSN literature [5], [9], [11], [14] several various models were proposed. The most popular concepts of network segmentation for multi-hop networks are regions building, clusterization and neighborhood. Let us come closer to these issues and begin from M ap(X, Y ) expression that can be defined as a collection of mappings of set X onto set Y (surjection). Further, Sub(X) is defined as a family of all X subsets and segment S as a mapping S ∈ {M ap(N odes, Sub(N odes))}

(1)

S(k)|k ∈N odes := {y ∈ N odes | y RS k }

(2)

where

and k is a segment’s main node (segment head). Based on a different segment relation RS further dyadic relations are defined. For example, we can build segments which are both mutually exclusive and collectively exhaustive with respect to the set of all network nodes – N odes. Formally, such segments constitute indexed family of sets S = {Si | i ∈ I} for which following properties are fulfilled:  Si = N odes, (3) (∀i ∈ I)(Si = ∅) ∧ (∀i, j ∈ I | i = j)(Si



Sj = ∅).

(4)

Conditions (3), (4) imply that: (∀y ∈ N odes)(∃! i ∈ I | y ∈ Si )

(5)

where ∃! means ”exists exactly one”. We recall that conditions (3), (4) determine a partition of a set N odes. Consequently, S is a partition of set N odes and the total number of possible partitions of an n-element set N odes is expresses by the Bell number Bn . These numbers satisfy a well known recursion formula and soar in the number of network nodes. In clustering and building regions approaches the emphasis is on determining multiple, potentially useful partitions and letting the algorithm decide. However, this lead to a larger consumption of network resources like energy or channel throughput. In general, research on efficient routing in wireless sensor networks has proceeded along two main approaches: cluster- 704 ing (RC ) or building regions (RR ).

In clustering algorithms (RC ) network partition results in family of subsets called clusters. Clustering algorithm decides whether particular node becomes the cluster head or a regular one. As a consequence a specific type of subsets is created Fig.1(left). Considering the pros and cons of clasterization three are of utmost significance: • it allows to build hierarchical structures with cluster heads and regular nodes, • it reduces mesh communication, places restrictions on regular nodes activity within cluster, • it increases efficiency of multi-hop communication since only cluster heads are responsible for message routing. Nonetheless, clustering results in the existence of exactly one transmission path between any node and the base station Fig.1(left) which is considered as a drawback. The second commonly accepted approach to network partition is based on region (RR ) concept [15]. That solution is based on obvious and regular network segmentation as presented on Fig.1(right). A regions building approach is derived from both the technological limitation of radio communication and multi-path retransmission. First, based on radio link range, the network is partitioned into coronas determined by concentric circles centered in base station. Next, a pie structure is determined using a number of angular wedges centered at the base station. Each pie relates to path routing area. Finally, a regular structure consisted of regions is created. From a global (network) perspective the advantages of traditional network partitioning are evident and clearly seen on Fig.1. There is an obvious trade-off; both partitioning methods i.e. sector building and clustering simplify and clarify the global view, but suffer from the reduced possibility of choice in each node. In some specific situations such segmentation of the network into regions or clusters can be very beneficial, because they simplify communication, but this is not the case in general.

3.2. Network segmentation based on neighborhood In real WSNs, sector building and clustering issues are a mixed blessing: building regions is not practically effective and clustering is not simple at all. Because of that we suggest the use of neighborhoods as the primary spatial entities and show how freak results can be obtained from surprisingly few primitives. Now we attempt to extend network segmentation approach to neighborhood. Based on (1), (2) and substituting RS = RN we define the neighborhood N as follows: N ∈ M ap(N odes, Sub(N odes)).

(6)

Thus, N (k) denotes the neighborhood of node k while N (S) is the neighborhood of set of nodes S defined as: N (k)|k ∈N odes := {y ∈ N odes | y RN k }, N(S)|S ⊂N odes := {y ∈ N odes | (∃k ∈ S)(y RN k)},

(7) (8)

where y RN k means that nodes y and k are in relation RN . There are a number of reasons for eschewing a cluster– based approach to adaptive routing in WSN and indeed simply using neighborhood abstraction supported by the standard tools of mathematical topology. Firstly, clustering is some kind of simplification, that facilitates computation and restricts the set of possible solutions at

Figure 2. Different perspectives of communication: a). within cluster, b). within neighborhood

Figure 3. Two types of WSN structured topology: clusters (right) and neighborhoods (left)

the same time. Concerning the cluster (Fig.2a), there is only one way to send data from node y towards base station while neighborhood (Fig.2b) provides evidently more possibilities. Next, the neighborhood gives a natural way to represent a kind of collective cooperation within some vicinity, that is germane to routing activity. Finally, the neighborhood abstraction is determined by native (mostly technical) constraints – e.g. radio communication range. However, such an approach generalizes the concept of a neighborhood far beyond its intuitive meaning, it also turns out that it is possible to extend neighborhood abstraction to sets (e.g. (8)). Considering native neighborhoods within WSN network one can define an indexed family of sets N = {Ni | i ∈ I}, where I denotes the set of indicies and Ni has the following properties:  (∀i ∈ I)(Ni = ∅) ∧ Ni = N odes, (9)  (10) Nj = ∅). (∃∼ i, j ∈ I | i = j)(Ni

is then used to perform local activities and to choose the best tactics that will be implemented in practice.

The last property can be rewritten for a single node y as  (∀y ∈ N odes)(∃∼ i ∈ I | y ∈ Ni = ∅), (11) where the expression ∃∼ can be translated as: ”there are as many instances as the structure of the network allows for” and this completes the definition of neighborhood. It is worth mentioning that we obtained (11) as a result of relaxing the requirement (5). To put it another way, formula (10) can be seen as a negation of (4). Note, that as a result we obtain a strongly overlapped structure (Fig.3 (left)). If two or more neighborhoods overlap then they share a common region, while this cannot be the case for clusters, which must exactly ’touch’ each other. It means that neighborhoods do not structure a set of WSN nodes into mutually exclusive subsets.

4. NEIGHBORHOODS WITH ADAPTATION ABILITY A neighborhood abstraction, defined by a set of criteria for choosing neighbors and set of common resources to be shared, is very useful in almost all algorithms of routing protocols in WSN. Realizing distributed operation/tasks in which nodes communicate only within vicinity, sensor network draws on concept of neighborhood. It is worth pointing out that the neighborhood relation is of the great significance since the whole activity of every node of WSN is determined 705 by the state of the node and its neighbors. The neighborhood

4.1. Relational attempt to network activity Given our new primitive of the neighborhood, we can now start defining some new relations that exploit this abstraction. Lets introduce the following three new dyadic relations: Subordination π = {; x,y ∈Act | x π y}.

(12)

The expression xπy defines the action x which is subordinated to the action y or action y dominates over action x. T olerance

ϑ = {; x, y ∈ Act | x ϑ y}.

(13)

The expression x ϑ y, states that the actions x and y tolerate each other, Collision

κ = {; x, y ∈ Act | x κ y},

(14)

and finally xκ y means the actions x and y are in collision to one another. The basic properties of mentioned above relations could be formulated succinctly as follows [8]: π ∪ ϑ ∪ κ ⊂ Act × Act = ∅,

(15)

ι ∪ (π · π) ⊂ π,

(16)

where ι ⊂ Act × Act is the identity on the set Action. Moreover: (17) π ∪ ϑ−1 ∪ (ϑ · π) ⊂ ϑ, where ϑ−1 is the converse of ϑ. That is: ϑ−1 = {< x, y >∈ X × Y | y ϑ x}.

(18)

For collision, κ −1 ∪ {π · κ} ⊂ κ ⊂ ϑ , ,

(19)

where ϑ , is the complement of ϑ i.e.: / ϑ}. ϑ , = {< x, y >∈ X × Y |< x, y >∈

(20)

π ∪ ϑ ∪ κ ⊂ Act × Act = ∅,

(21)

ι ∪ (π ◦ π) ⊂ π,

(22)

and where ι is an identity relation on the set Act. Eq. (21) states that all three relations are binary on non-empty set of actions

(Act). Eq. (22) states that subordination is reflexive (ι ⊂ π) and transitive (π ◦ π ⊂ π). Further π ∪ ϑ−1 ∪ (ϑ ◦ π) ⊂ ϑ

(23)

means that: • subordination implies tolerance – if π holds for some x, y ∈ Act then ϑ also holds for these, • tolerance is symmetrical – if xϑy ⇒ yϑx, • subordinated action tolerates all actions tolerated by the dominant – if (xπy ∧ yϑz) ⇒ xϑz. For collision relation we have that κ −1 ∪ {π ◦ κ} ⊂ κ ⊂ ϑ

(24)

where ϑ is the complement of ϑ: / ϑ}. ϑ = {< x, y >∈ X × Y |< x, y >∈

(25)

Eq. (24) states that collision is symmetric and disjoint to tolerance. Moreover, all subordinated actions must be in collision with action being in collision with its dominant.

4.2. Modeling global network strategy The main aim at the WSN, similar as for optimization problems (goal function, drainage function, constraints) related to communication in WSN, is defined globally within the scope of the whole network. Therefore, using relations we determine (globally) initial organization which remains static during a network lifetime. Let us consider the node k and its neighborhood N (k). Any communication activity actk that is performed by node k relates to some members of N (k) and the set of actions actk within neighborhood N (k) can be defined as follows: ActN(k):={actk∈Act|(∃x ∈ N(k))(actx R actk )}.

(26)

The Cartesian product defined as: ISk := ActN (k) × ActN (k) ⊆ π ∪ ϑ ∪ κ,

(27)

(28)

Thus, for a given relation R and a node k we define an intensity quotient within neighborhood N (k) as follows; IRk = Card(Rk )/Card(ISk ),

(29)

where Card(Rk ) is cardinality of set Rk . Let ISW SN be a global interaction space consisting of all feasible actions in WSN and GSW SN be a global strategy defined as a subset of ISW SN . Notice that there is only one set ISW SN while there may be many different ways GSW SN can be chosen. However, for the simplicity let us consider only singleton s = {< IRπ, IRϑ, IRκ >}: s

∗ = {< 0.2, 0.54, 0.07 >} = GSW SN

4.3. Modeling adaptive activity within neighborhoods In the previous chapter we have described how to use three dyadic relations π, ϑ, κ to ensure that all local activities are in the scope of desired global behavior. In WSN systems actions and decisions are taken by nodes based on their knowledge about the network. Due to limited communication range, nodes have to cooperate within the neighborhood to gain this knowledge. ∗ Due to the global strategy GSW SN (30) the activity of each node k is restricted to three subsets of neighbors: Nπ (k)={y ∈ N(k) | yπk},

Nϑ (k)={y ∈ N(k) | yϑk},

Nκ (k)={y ∈ N(k) | yκk}.

(31)

According to (30)

describes interaction space ISk within N (k). Let us now consider a set of possible interactions fulfilled relation R within neighborhood N (k) which can be expressed as: Rk := {y ∈ ActN (k) |< actk , y >∈ ISk ∧ k R y}.

Finally, having a well defined neighborhood and required relational intensity for relations π, ϑ, κ, we are ready to decompose the globally defined goal function and constraints into a uniform locally performed task assigned to each node in the network. It is not an easy task to cast all global dependencies from network area to the neighborhood. Moreover, the fact that neighborhood conditions for the network nodes might be, and usually are, quite dissimilar makes this issue even more difficult. In the paper [12] we have proposed a concept to realize global/local task decomposition based on a Digital Terrain Model. Instead of globally formulated optimization tasks for network, we model 3D surface over the network area. It results in bare drainage surface that gives a basic reference frame which ensures data is send towards the BS. There are many possible surfaces, but for further consideration, let us assume that the drainage surface is based on node to base station hop-distance.

(30)

and initialize all WSN nodes with these intensity quotients as suggested intensity for relations π, ϑ, and κ, within their neighborhoods. This implies that any communication activity within node neighborhood should be limited to this space and each node should retain demanded relational intensity within its neighborhood. This ensures that the local behavior 706 is compliant with the global requirements in a full extent.

C(Nπ (k))/C(Nκ (k))/C(Nϑ (k)) = 0.2/0.54/0.07

(32)

where C(X) denotes cardinality of X. Intensity quotients of relation π, κ, ϑ can determine number of the elements in the each of (31) sets. The decision of which node from the neighborhood belongs to which particular set must be taken locally, but the drainage function based on hop-distance will help to make the correct decision. The aforementioned distance is expressed in terms of number of hops required to reach the base station, moreover every node with distance X+1 can communicate with at least one node which distance to the BS equals X. Eventually any node k knows its hop distance (dish (k)) and distances of all of its neighbors (dish (i) where i ∈ N (k)). Based on this information it is possible to split neighborhood N (k) into three subsets: N< (k) = {y ∈ N(k) | dish (y) < dish (k)}, N= (k) = {y ∈ N(k), y = k | dish (y) = dish (k)}, h

(33)

h

N> (k) = {y ∈ N(k) | dis (y) > dis (k)}, which are mutually exclusive and collectively exhaustive N (k) but they are not partition of a set N (k). When determining the sets (33) any node k can attach elements of sets (31) such that (32) is satisfied and retain a data-flow direction towards the BS. Neighbours that belong to a Nπ (k) set are selected from N< (k) set so it may consist of some nodes i ∈ N< (k). Based on N> (k) a set Nκ (k) is

being established and similarly Nϑ (k) on N= (k) ∪ N> (k), so finally: Nπ (k) ⊂ N< (k); Nϑ (k) ⊂ (N< (k) ∪ N= (k)) Nκ (k) ⊂ N> (k).

(34)

Such an idea of local communication tactics ensures compatibility with globally established strategy and retains a proper data drainage direction towards the BS. This works regardless of the fact that in most cases a node does not know the location of the BS. The problem we now face is to order elements of the sets (31). For this purpose we use two indicators of quality connections: radio link quality indicator (LQI) and Received Signal Strength Indicator (RSSI). To order elements of Nπ (k) we use LQI values while to order the remaining two; Nϑ (k), Nκ (k) we use RSSI. Ordering any of these sets we write down the indicator values in order of magnitude, beginning with the greatest. Each node must locally undertake a decision, which node is the one to pass the packet to. It is vital that the packet from a node k traverses in a direction determined by slope of a bare drainage surface. This means that the next hop node has to be chosen from among the neighbors Nπ (k) which hop distance to the base station is smaller than dish (k). In such a situation we choose the first node (e.g. this with the greatest value of LQI parameter). Because, either the values of LQI or RSSI reflect upto-date conditions in the neighborhood and vary in time, adaptability to current environment conditions within neighborhood is attributed to such techniques. When node detects changes in his vicinity or detects deviations from the normal neighbor behavior, it has wide choice of communication path. As a result, each multi-hop communication between any two nodes may use a different communication path.

5. FINAL REMARKS The research work is still in progress on wireless sensors network and related formalisms based on the theory of sets and relations. Taking advantage of the relational approach, an innovative strategy is applied that involves the entire network, by sending triplets the strategy that is prepared for the whole network, by sending triples of intensity quotients (30). Moreover, the development of relations, its intensity quotients and metric, both globally – in network and locally – within neighborhood, results in routing adaptability. Each node performs some communication activity, based on globally determined balance between subordination π, tolerance ϑ and collision κ. So, this leads rather to governance than to control, what finally causes that particular action strongly depends on varied in time local/neighborhood conditions. The development of relation provided network not only with adaptability but also facilitate communication structure continuous revitalization. The concept of neighborhood is essential and really native to WSN. This is because of radio communication; the radio link range constraint is an origin of neighborhood. Furthermore, there is no doubt that a high cardinality of any neighborhood set and its extensive overlapping makes the neighborhood more attractive than clusters or regions. A wide scope of choices is the reason for success. The 707 approach allows for individual tactic selection (within node’s

neighborhood), that takes into consideration conditions in the node’s vicinity while at the same time retaining global strategy requirements. The neighborhood relation is defined both for a single node of the network and for a group of nodes. It is worth pointing out that the neighborhood relation is of the greatest significance since the whole activity of every node of WSN is determined by the state of a node and its neighbors. The neighborhood is then used to perform local activities and to choose the best tactics that will be implemented in practice. As a result, the native neighborhood was advised as the most suitable form of the local range.

Acknowledgements This work was supported by E.U. Regional Development Fund and by Polish Government within the framework of the Operational Programme - Innovative Economy 2007-2013. Contract POIG.01.03.01-02-002/08-00, Detectors and sensors for measuring factors hazardous to environment - modeling and monitoring of threats.

References [1] Cohn, A.G., Bennett B., Gooday J.M., Gotts N.M., 1997. Representing and Reasoning with Qalitative Spatial Relations about Regions. In: Cohn, A.G., Bennett B., Gooday J.M., Gotts N.M, eds. Spatial and Temporal Reasoning, Dordrecht, Kulwer, 97-134. [2] Cerpa A., Estrin D. ASCENT: Adaptive Self-Configuring Sensor Networks Topologies,IEEE Transactions On Mobile Computing, vol. 3, no. 3, Jul-Sep 2004. [3] C. Chevallay, R. E. Van Dyck, and T. A. Hall. Self-organization Protocols for Wireless Sensor Networks. In Thirty Sixth Conference on Information Sciences and Systems, March 2002. [4] Chaczko Z., Ahmad F.: Wireless Sensor Network Based System for Fire Endangered Areas, ICITA 2005, Sydney, 2005. [5] Cohn, A. G. and Gotts, N. M.: 1994b, A theory of spatial regions with indeterminate boundaries, in C. Eschenbach, C. Habel and B. Smith (eds), Topological Foundations of Cognitive Science. [6] Crandall J.W., Goodrich M.A.: Experiments in adjustable autonomy, in IEEE International Conference on Systems, Man, and Cybernetics, vol.3, Tucson, USA, 2001, 1624-1629. [7] Descartes R., Lafleur L.: Discourse on Method and Meditations, New York: The Liberal Arts Press., 1960. [8] Jaro J.: Systemic Prolegomena to Theoretical Cybernetics, Scient. Papers of Inst. of Techn. Cybernetics, Wroclaw Techn. Univ., no. 45, Wroclaw, 1978 [9] Lin Ch.R., Gerla M.: Adaptive Clustering for Mobile Wireless Networks, IEEE Journal On Selected Areas In Communications, vol. 15, no. 7, Sep 1997 [10] Nikodem, J., 2008. Autonomy and Cooperation as Factors of Dependability in Wireless Sensor Network, Proceedings of the Conference in Dependability of Computer Systems, DepCoS - RELCOMEX 2008, 406-413. June 2008, Szklarska Poreba, Poland [11] Nikodem J.,Klempous R., Chaczko Z., Modelling of immune functions in a wireless sensors network.W: The 20th European Modeling and Simulation Symposium. EMSS 2008, Campora S. Giovanni, Italy, 2008

[12] Nikodem J., Klempous R., Nikodem M., Woda M.: Directed communication in Wireless Sensor Network based on Digital Terrain Model. 2nd International Symposium on Logistics and Industrial Informatics (LINDI), Linz, Austria,[Witold Jacak et all eds., pp. 87-91, Piscataway, NJ : IEEE

Ryszard Klempous holds a M.Sc. in Automation (1971) and Ph.D. in Computer Science (1980) from Wroclaw University of Technology (WUT). Since 1980 he has been an Assistant Professor in the Institute of Computer Engineering, Auto-matics and Robotics, WUT. Senior member of IEEE and NYAS, has already published over 90 papers in Optimization Methods and Algorithms, Simulation and Data Processing and Transmission

[13] Scerri P., Pynadath D., Tambe M.: Towards Adjustable Autonomy for the Real World, Journal of Artificial Intelligence Research, vol.17, 2003 [14] Sohrabi K., Gao J., Ailawadhi V., Pottie G.J.: Protocols for Self-Organization of a Wireless Sensor Network, IEEE Personal Communications, Oct 2000

Marek Woda is an Assistant Professor in the Institute of Computer Engineering, Control and Robotics at Wroclaw University of Technology. In 2001 he graduated at WUT. In 2007, he received Ph.D. degree in Computer Science from the Faculty of Electronics WUT. His research interests focus on multi-agents systems, e-learning, Internet technologies. He participated in international projects sponsored by European Union (e.g. PL96-1046 INCOCOPERNICUS project Multimedia Education: An Experiment in Delivering CBL Materials, VI FP Integrated Project FP6-IST 26600 DESEREC Dependability and Security by Enhanced Reconfigurability). He is the author of about 30 scientific articles and conference papers.

[15] Stojmenovi´c, I.,editor: Handbook of Sensor Networks Algorithms and Architectures, John Wiley and Sons Inc., 2005. [16] Veyseh M., Wei B., Mir N.F.: An Information Management Protocol to Control Routing and Clustering in Sensor Networks, Journal of Computing and Information Technology - CIT 13 (1) 2005, 53-68 [17] Younis O., Fahmy S.: HEED: A Hybrid, EnergyEfficient,Distributed Clustering Approach for Ad Hoc Sensor Networks, IEEE Transactions On Mobile Computing, vol. 3, no. 4, Oct-Dec 2004

AUTHORS BIOGRAPHY

Zenon Chaczko completed a B.Sc. in Cybernetics and Informatics in 1980 and a M.Sc. in Economics in 1981 at the University of Economics, Wrocław in Poland, as well as completed MEng in Control Engineering at the NSWIT 1986 and Ph.D. in Enginering at UTS, Australia. For over 20 years Mr Chaczko has worked on Sonar and Radar Systems, Simulators, Systems Architecture, Telecommunication network management systems, large distributed Real-Time system architectures, network protocols and system software middleware. Mr Chaczko is a Senior Lecturer in the Information and Communication Group within the Faculty of Engineering at UTS.

Jan Nikodem received the B.Sc. in electrical engineering, M.Sc. in artificial intelligence in 1979 and Ph.D. degree in computer science in 1982 from Wroclaw University of Technology (WUT), Poland. Since 1986, he is an Assistant Professor in the Institute of Technical Cybernetics, WUT. Since 2005 in the Institute of Computer Engineering, Automatics and Robotics (ICEAR). His current research is focused on the area of complex and distributed systems, cybernetics, wireless sensor networks and digital data transmission. Maciej Nikodem received a M.Sc. in Computer Science in 2003 and a M.Sc. in Control and Robotics in 2005 from the Wrocław University of Technology in Poland. In 2008 he completed Ph.D. studies in Computer Science at Faculty of Electronics, Wrocław University of Technology. For last 5 years Maciej Nikodem has worked on Countermeasures to Fault Analysis, Boundary Scan Security as well as security aspects of Wireless Sensor Networks. Maciej Nikodem is an Assistant in the Institute of Computer Engineering, Control and Robotics, Faculty of Electronics at WUT.

708

BUSINESS PROCESS SIMULATION FOR MANAGEMENT CONSULTANTS: A DEVS-BASED SIMPLIFIED BUSINESS PROCESS MODELLING LIBRARY Igor Rust(a); Deniz Cetinkaya(b); Mamadou Seck(c); Ivo Wenzler(d) (a)(b)(c)

Systems Engineering Group; Faculty of Technology, Policy, Management Delft University of Technology; Jaffalaan 5, 2628BX, Delft, THE NETHERLANDS (d) Accenture Nederland; Gustav Mahlerplein 90, 1082 MA, Amsterdam, THE NETHERLANDS (a)

[email protected], (b)[email protected], (c)[email protected], (d)[email protected]

ABSTRACT Business process simulation enables analysis of business processes over time and allows to experiment with scenarios (like for instance redesigning business processes) before implementing them into an organization. This research aims at providing an easy way of business process modelling and simulation for management consultants whose core competence is not simulation model development. During the design and development process, management consultants are actively involved following a user-centred design approach. The outcome of this research is a library of DEVS-based business process modelling elements implemented in Java and using the DSOL simulation library to provide the simulation capabilities.

 Q LV WLR \V QWD O D H $Q HP SO LP

([ SH ULP

HQ WD W

LR Q

&R Q

FH SW X

DO P RG

HO LQ

J

 HO RG P Q Q LR LR FW DW UX XO QVW P 6 L FR

Figure 1: Simulation Project Life Cycle. Figure 1 depicts the main phases and products of a simulation study. The organization, for which an analysis is undertaken, is part of the “real world”. First, a conceptual model is developed, often in a graphical form, which contains the essential aspects of the problem situation (Banks 1998). A conceptual model helps to build credibility for and guides the development of a simulation model. Next, a simulation model of the business process is developed based on this conceptual model. The executable model can be manually programmed by the modeller, or constructed through a visual interface (Pidd 2004a). After the development of the simulation model, experiments can be set up and the simulation model is executed by simulation software to analyse the output. A simulation software generally consists of a simulation engine (or simulation executive) and an application program (Pidd 2004b). The engine keeps track of the state changes which occur at some moment in time and reminds the application when a state change is due. How the experiment is set-up and which output parameters are of interest, depend on the business case. Based on the presented outputs, more experiments can be performed or changes may be implemented in business processes. Although the usefulness of business process simulation was proven by many authors and various simulation tools are available, still many consultants

Keywords: business process modelling, business process simulation, component based modelling, DEVS 1. INTRODUCTION To stay competitive and to operate effectively, an organization needs to improve its process efficiency and its quality by adapting its strategy, structure, management and operations to its changing business environment. Management consultants provide expertise and recommendations to improve their clients’ business performance. To support organizational decisions, a good understanding of the business processes is essential. A business process is a series of activities that produces a product or service for a customer. Business Process Modelling (BPM) is the activity resulting in a representation of an organisation’s business processes so that they may be analyzed and improved (Weske 2007). A distinction can be made between static modelling and dynamic modelling of business processes (Bosilj-Vuksic, Ceric and Hlupic 2007). Static modelling tools often provide a graphical process representation, for example simple flowcharts, IDEF0 or BPMN diagrams to depict business processes. Business Process Simulation (BPS) tools, provides the possibility to simulate and evaluate the dynamic behaviour of business processes.

709

The availability of a resource depicts when or for how long a resource is available to commence activities. For instance, a full time employee has an availability of 1 FTE (Full Time Equivalent), which depicts that during a complete working day the resource is available to perform activities. The state of a resource relates to whether a resource is currently busy (or active) with performing a certain activity, or is available for new work. An activity is a piece of work performed by one or more resources which requires a certain amount of time. An activity can be a task or a grouping of task, called a sub-process. A task is a piece of work that cannot be subdivided into smaller pieces of work to be performed by a resource, or is not crucial for the purpose of describing and analyzing a business process. There are three options how an activity can be performed, namely 1) one resource starts and finishes an activity on its own; 2) a resource hands over the entity to another resource who will perform one or more activities; or 3) the amount of work is divided over two or more resources. In the first case, the flow of an entity through activities is called sequential. In the second case, a resource will hand over the entity to another resource that will perform his activities. The third case is called a parallel activity. After the work is split up, two or more resources will perform their activities independently. At some moment the work is synchronized again and some resource will continue performing activities. In a business process decisions are made that influence the choice and order of activities to be undertaken by a resource. A decision can depend on the attribute of an entity (e.g. entity type or state), or the state of the system (e.g. what are other resources doing, how many entities are waiting to be processed, etc.).

and business analysts rely on simple static process mapping methods (Bosilj-Vuksic et al. 2007; Melão and Pidd 2003). Some reasons for the lack of adoption are that much experience is needed to develop valid simulation models and simulation model development is time consuming and costly (Van Eijck and De Vreede 1998). More specifically, there is a lack of business process simulation tools which supports an easy and quick approach of modelling and analysis of business process by consultants and business analysts. This paper presents a business process simulation method to support management consultants to model, simulate and analyze business processes in a well defined manner. Next section provides background information about business process consultancy which is extracted from the interviews with the consultants. Section 3 discusses the design process of our research which is based on a user-centred design approach. Section 4 and 5 present a DEVS component library for business process modelling and its usage. Finally, conclusions and future work are given in Section 6. 2.

SUPPORT FOR BPS BY CONSULTANTS

2.1. The Consultants’ View on Business Processes Harington (1991) defines a business process as a group of logically related tasks that use the resources of the organization to provide defined results in support of the organization's objective. Consultants see business processes more specifically as “a series of activities and decisions which are performed by resources and which influences the flow and state of the entities”. An entity (or passive entity) is an abstract object which can represent anything that undergoes activities in a business process. The entity arrives at an organization, “flows” through the business process(-es), and then leaves the organization. What the entity actually represents is called the entity type. Examples are an order that is received by a company and needs to be processed, an insurance claim or a contract cancellation e-mail. During a business process, an entity undergoes state changes as a result of the activities that are performed by resources. The state of an entity may for instance be “received”, “processed” or “finished”. A resource is responsible for making decisions and performing activities on entities and is considered as the leading part of a business process. A resource can be a human or a machine. Resources are typified by their capability, role, capacity, availability and state. The capability of a resource depicts whether a resource is able or allowed to perform a certain activity. Based on for instance the experience that a resource has, the resource may be able to perform more complicated activities. The collection of all capabilities of a resource is called the resource's role. It is possible that within an organization, multiple resources have the same capabilities and thus share the same role. In that case, a role has a certain capacity: the number of resources that are available to perform a determined set of activities.

2.2. Conceptual Modelling Pidd (1996) defined a model in the context of operations research and management sciences as “an external explicit representation of part of reality as seen by the people who wish to use that model to understand, to change, to manage and to control that part of reality”. In other words, a model can be used as a representation of reality (like for instance an object, idea or an organization), to support someone who wants to understand that part of reality, and possibly wants to make decisions which will influence reality. There exist various modelling languages that support the representation of business processes in a model in a standard and consistent way. Some examples are BPMN, Flow Chart, Gantt Chart, IDEF0, IDEF3, and UML (Aguilar-Saven 2004). Each of these languages has different characteristics (semantics, representation notation, ability to include decomposition and hierarchy of processes, etc.). In order to support consultants with a new modelling approach, a conceptual modelling language should be chosen or developed that is able to represent the consultants view on business processes as described in the previous section. The notation should also be

710

understandable to enable correct interpretation by other consultants, as well as domain experts of the modelled organization.

business process modelling and simulation, we follow a user-centred design (UCD) approach. The main goal of a UCD approach is to increase the likelihood that a designed and developed artefact is found usable by its end-users. User-centred design approach is concerned with incorporating the end-users perspective during the design and development process to achieve a usable system (Maguire 2001). Because management consultants are the end-user of our new business process modelling method, they are placed at the centre of the design process. To incorporate consultants in this research, a series of design and evaluation rounds are held (in the form of workshops) with management consultants of a large international management consultancy firm. These workshops are intended to get (among other things) an understanding of the consultant and his/her view on business processes (as it is already discussed in Section 2.1); to decide upon and evaluate an understandable modelling language of the consultants view and to evaluate the usability of the proposed modelling approach.

2.3. Simulation Model Development As mentioned in the previous section, various languages exist to represent conceptual business process models. However, most of these languages present an abstract way of thinking and don’t provide the possibility to include all details needed to enable direct translation into an executable simulation model, nor direct execution of these models by a simulation engine. Due to the lack of possibilities for a conceptual modeller to include all details in a conceptual model, different simulation models can be developed based on the same conceptual model. If for instance the simulation model is developed by someone who was not part of the conceptual modelling stage, the risk increases that a final simulation model does not represent the business processes as was intended by the conceptual modeller. Cetinkaya, Verbraeck and Seck (2010) concluded in recent research that there is a large semantic gap between the conceptual modelling stage and the simulation model construction stage. With regard to the actual development of an executable simulation model, various formalisms exist to support the formalization of simulation models, like for instance Discrete Event System Specification (DEVS) (Zeigler, Praehofer and Kim 2000) and Petri Nets (Peterson 1981). Developing an executable simulation model using one of these formalisms requires a deep understanding of the underlying concepts, as well as programming experience. Reusing parts of simulation models or reusing even complete simulation models is suggested to decrease the complexity and time needed to develop models. Different forms of reuse are: code scavenging (reusing existing code of a simulation model), function reuse (reusing predefined functions that provide specific functionalities), model reuse (reusing a complete simulation model for a different situation) and component reuse (reusing an encapsulated simulation module with well-defined interfaces) (Pidd 2002). Usage of components for simulation modelling is considered to be fruitful concept to increase the efficiency of hierarchical model construction (Cetinkaya, Verbraeck and Seck 2010).

4.

A DEVS COMPONENT LIBRARY FOR BPM

4.1. Modelling Elements The modelling representation that was the outcome of the design research process is based on the Business Process Modelling Notation (BPMN). BPMN is an industry-wide standard for modelling of business processes and was chosen because of various reasons, like: 1) the way consultants see business processes, corresponds closely to how business processes are described in the BPMN specification; 2) many management consultants have experience with modelling of business processes through the use of “swimlane diagrams" (a method that resembles BPMN to some extent, but more simplified); 3) BPMN was used for conceptual modelling of business processes in past simulation projects and was found to be useful by the involved consultants; and 4) the BPMN is becoming a standard language to model business processes throughout industries, which increases the likelihood that clients are already familiar with the notation and the models. A set of modelling elements were determined (see Figure 2), which allow modelling of business processes by consultants as how they actually perceive business processes (as mentioned in Section 2.1). A resource (or group of resources which share the same capabilities) is represented by a Swimlane, which is a graphical modelling element that can contain activities (like tasks and sub processes) and decision elements (gateways). The arrival of entities in a business process is represented by a Start Event. The Start Event is placed outside a Swimlane, to depict that arriving entities are coming from outside the organization (or business unit) and no resources are busy at that arrival time. The End Event represents the end of a business process, namely when an entity leaves the organization, and is also placed outside a Swimlane.

3. DESIGN PROCESS When we consider business process simulation from a consultant's point of view, the three main activities that he/she is interested in performing, are: 1) developing the business process model (conceptual model) and specifying the model parameters; 2) experimenting with a simulation model; and 3) interpreting the results. How the translation of a conceptual model into an executable simulation model can be done is important, but also irrelevant to the consultant (as long as the simulation model leads to results as how the consultant intends it to do). Because our goal is to support consultants with

711

Figure 3: State Diagram for Task Element. The supplementary components are developed to support some of the needed simulation functionality as discussed by the consultants. For instance, when an entity enters a swimlane, the entity is placed in a queue and a resource is requested. When available, the resource is attached to the entity after which the entity leaves the swimlane entry. For this purpose, the Swimlane Entry component was designed. When an entity leaves a swimlane, the resource occupied with the entity should be made available for other (possibly already waiting) entities. This is done by the Swimlane Exit component. To organize the assignment of resources to waiting entities, a partial de-central approach was chosen, namely by implementing a Resource Manager component which is part of each Swimlane. The Resource Manager (RM) receives signals from amongst other the Swimlane Entry and Swimlane Exit that a new entity arrived and was placed in an entry queue, or that an entity is leaving a Swimlane. Based on a certain rule as specified by the modeller, the RM evaluates the states of all resources and queues within a Swimlane, and directs if possible a resource to a queue (by sending a signal to the appropriate queue containing information about the available resource and the destination queue).

Figure 2: Overview of the BPM Elements. After the entity enters through a Sequence Flow a Swimlane, a resource “picks up” the entity in a prespecified manner (e.g. on a random basis, following some pattern or based on a certain priority rule) and performs one or more activities. A Task modelling element is an activity which represents work that is performed by a resource and consumes a certain amount of time. Sub-process modelling element is included to support hierarchical modelling. Parallel Gateways are included to enable activities to be performed in parallel by multiple resources and are used in pairs: one gateway is used at the start of a parallel activity. It duplicates an entity and forwards the entities to the activities that are performed by different resources in parallel. A second Parallel Gateway is used to synchronize a parallel activity again after both activities are completed. Exclusive Gateways provide the functionality to resemble business decisions. The flow of an Entity is directed based on the evaluation of a condition. This condition can be either the evaluation of an entityspecific attribute (e.g., entities of a specified type/state move in one specified direction, other entities move in another direction), or based on probability (e.g., 70% of the arriving entities move in one direction, the other 30% move in the other direction).

4.3. Implementation with DEVSDSOL DSOL, which stands for “Distributed Simulation Object Library”, was selected to provide the simulation and execution functionalities (Jacobs 2005). DSOL is a proven multi-formalism simulation library which can be considered as a generic purpose simulation tool. It is written in the Java programming language and has been used effectively in various simulation projects. DSOL also supports execution of simulation models based on the DEVS formalism through the DEVSDSOL library (which is compatible with hierarchical DEVS) (Seck and Verbraeck 2009). The choice for DSOL was made because of its flexibility and functionality regarding simulation, and its support for the DEVS formalism. Each DEVS component has been implemented in Java and these components are executable with DEVSDSOL simulation library. Some implemented components are BPMNStartEvent, BPMNEndEvent, BPMNTask, BPMNExclusiveGateway, BPMNParallelGateway, etc. Figure 4 shows the DEVS internal transition function of the Task element.

4.2. Formalization of Modelling Elements We use DEVS to specify our simulation models. In DEVS, a system can be represented by two types of models: atomic and coupled models. Atomic DEVS models describe the behaviour of a system, whereas coupled models describe the composed structure of a system. Atomic models can be integrated into coupled models, and coupled models can be integrated in higher level coupled models. This way, a model is decomposed in a hierarchical manner. The suggested BPM elements and some supplementary elements are matched to DEVS components. For every element a state-diagram was developed and validated. Figure 3 shows the state diagram for Task element. Since a Swimlane represents either a resource or a group of resources, a Task needs to check for waiting entities.

712

Figure 5: Sample Business Process Model.

Figure 4: Sample Code from Task Component In order to provide a higher level abstraction mechanism to our library, we applied the model driven development framework presented in (Cetinkaya, Verbraeck and Seck 2011). Next section gives brief information about the framework and then explains how it is performed in this work. 5. APPLYING THE MDD4MS FRAMEWORK MDD4MS is a model driven development framework for modelling and simulation. The framework suggests an M&S life cycle with five stages (Problem Definition, Conceptual Modelling, Specification, Implementation and Experimentation), metamodel definitions for different stages, model to model (M2M) and model to text (M2T) transformations for the metamodels and a tool architecture for the overall process. MDD4MS presents a sample prototype implementation which is developed in Eclipse. The MDD4MS prototype provides: • • • • • •

Figure 6: A sample rule from BPMN_2_DEVS.atl. coupled components that uses the implemented classes for BPM modeling elements in our library. In other words, visual business process models, drawn by the BPMN editor, are transformed to executable Java code and they can be simulated with DSOL.

a BPMN metamodel and a BPMN editor, a DEVS metamodel and a DEVS editor, a DEVSDSOL metamodel and a DEVSDSOL editor, a BPMN 2 DEVS M2M transformation, a DEVS 2 DEVSDSOL M2M transformation, a DEVSDSOL 2 Java M2T transformation. Figure 7: Auto generated DEVS Model.

In this study, we used the BPMN editor to draw our business process models. A sample business process model is shown in Figure 5. Since the MDD4MS prototype provides generic model transformation rules for BPMN, we rewrote some rules for BPMN2DEVS M2M transformation. In this way, we directly transformed the visual modelling elements to the components that we implemented in our library. For example, Figure 6 shows the rule to transform a Swimlane to a coupled model. We added the part that generates a Resource Manager with one input and one output port for each Swimlane. The auto generated DEVS model via model transformations is shown in Figure 7. Once we have the DEVS model, the MDD4MS prototype automatically generates the DEVSDSOL model and the java code for

6. DISCUSSION This work proposed a new modelling approach for consultants to model and analyse business processes based on a proven theory, industry-wide standards and active end-user involvement during the design process. A library of DEVS-based BPMN modelling elements is implemented with Java that uses the DSOL simulation library to provide the simulation capabilities. It provides an easy way of dynamic modelling for consultants with limited knowledge of simulation model development. As a future work, the credibility of our approach will be evaluated.

713

Weske, M., 2007. Business Process Management: Concepts, Languages, Architectures (p. 368). New York: Springer-Verlag. Zeigler, B. P., Praehofer, H. and Kim, T. G., 2000. Theory of Modeling and Simulation (Second Edi.). Academic Press.

REFERENCES Aguilar-Savén, R.S., 2004. Business Process Modelling: Review and Framework. International Journal of Production Economics, 90(2), 129-149. Banks, J., 1998. Handbook of Simulation - Principles, Methodology, Advances, Applications, and Practise (p. 849). New York: John Wiley & Sons, Inc. Bosilj-Vuksic, V., Ceric, V. and Hlupic, V., 2007. Criteria for the evaluation of business process simulation tools. Interdisciplinary Journal of Information, Knowledge, and Management, 2, 73– 88. Cetinkaya, D., Verbraeck, A. and Seck, M. D., 2010. Applying a Model Driven Approach to Component Based Modeling and Simulation. Proceedings of the 2010 Winter Simulation Conference (pp. 546553). Baltimore, MD. Cetinkaya, D., Verbraeck, A. and Seck, M. D., 2011. MDD4MS: A Model Driven Development Framework for Modeling and Simulation. Proceedings of the Summer Computer Simulation Conference 2011. Den Haag, NL. Harrington, H., 1991. Business Process Improvement: The Breakthrough Strategy for Total Quality Productivity. New-York: McGraw Hill. Jacobs, P. H. M., 2005. The DSOL Simulation Suite. Thesis (PhD). Delft University of Technology. Maguire, M., 2001. Methods to support human-centred design. International Journal of Human-Computer Studies, 55(4), 587-634. Melão, N. and Pidd, M., 2003. Use of business process simulation: A survey of practitioners. Journal of the Operational Research Society, 54(1), 2-10. Peterson, J. L., 1981. Petri Net Theory and the Modeling of Systems (p. 290). New Jersey: Prentice Hall. Pidd, M., 1996. Tools for Thinking: Modelling in Management Science. Chichester, John Wiley & Sons, Inc. Pidd, M., 2002. Simulation Software and Model Reuse: A Polemic. Proceedings of the 2002 Winter Simulation Conference. (p. 772-775) Pidd, M., 2004a. Computer Simulation in Management Science (5th ed., p. 311). West Sussex, England: John Wiley & Sons, Inc. Pidd, M., 2004b. Simulation Worldviews – So What?. Proceedings of the 2004 Winter Simulation Conference, 2004. England: John Wiley & Sons, Inc. Seck, M.D. and Verbraeck, A., 2009. DEVS in DSOL: Adding DEVS Operational Semantics to a Generic Event-scheduling Simulation Environment. Proceedings of the Summer Simulation Multiconference 2009. Istanbul, Turkey Van Eijck, D. T. T. and De Vreede, G.-J., 1998. Simulation support for organizational coordination. Proceedings of the 1998 Hawaiian Conference of Systems Sciences (Vol. 1, p. 633–642). Los Alamitos: IEEE Computer Society.

AUTHORS BIOGRAPHY IGOR J. RUST is an M.Sc. graduate in Systems Engineering, Policy Analysis and Management and received his degree in 2011 from Delft University of Technology. In 2011 he received his B.Sc. at the faculty of Technology, Policy and Management, also from Delft University of Technology. His B.Sc. thesis was nominated for the Dutch Logistics Thesis Award 2011. His interests include discrete event simulation and mathematical programming. His e-mail address is DENIZ CETINKAYA is a Ph.D. student at Delft University of Technology. She is in the Systems Engineering Group of the Faculty of Technology, Policy and Management. She received her M.Sc. in Computer Engineering from the Middle East Technical University, Turkey in 2005. She received her B.Sc. with honours in Computer Engineering from the Hacettepe University, Turkey in 2002. Her research focuses on component based modelling and simulation. Her e-mail address is . MAMADOU D. SECK is an Assistant Professor in the Systems Engineering Group of the Faculty of Technology, Policy, and Management of Delft University of Technology. He received his Ph.D. degree from the Paul Cezanne University of Marseille and his M.Sc. and M.Eng. degrees from Polytech Marseille, France. His research interests include modelling and simulation formalisms, dynamic data driven simulation, human behaviour representation and social simulation, and agent directed simulation. His e-mail address is . IVO WENZLER is part of Accenture’s Global Consulting Experts Group and was until recently a Senior Executive within Accenture’s Talent and Organization Service Line. For one day per week he also holds a position of Associate Professor at Delft University of Technology where he teaches a master’s course in simulation game design. His e-mail address is .

714

RESEARCH ON CO-SIMULATION TASK SCHEDULING IN CLOUD SIMULATION PLATFORM Chen Yang(a), Bo Hu Li(b), Xudong Chai(c) (a)

Beijing University of Aeronautics and Astronautics, Beijing, China Beijing University of Aeronautics and Astronautics, Beijing, China (c) Beijing Simulation Center, Beijing, China

(b)

(a)

[email protected], (b)[email protected], (c)[email protected]

which realizes the dynamic share and reusability, collaborative interoperability, dynamic optimization for simulation execution, of different resource. Simulation grid has resource management and task scheduling for certain degree, but due to the heterogeneity of OS and software environment, the large variance of performance, of nodes in simulation grid, in addition to unstableness of network, the simulation grid can not effectively and quickly execute the large-scale simulation, and the effect of co-simulation task scheduling methods is decreased severely with little advantage. Moreover, simulation grid can not show efficient support for fine granular resource (e.g. CPU, storage, software in nodes) share, multi-user, faulttolerance, etc. For example, in simulation grid, distributed computing nodes have different kinds and versions of OS and software, which will lead to the limited nodes the task can be scheduled to, such as the federates programmed and compiled to run in Windows OS is hard to scheduled to execute in Linux servers without any adaption. Due to the unsolved problems in simulation grid, People in Beijing Simulation Center introduce the notion of “cloud computing”, and further with integration of virtualization, pervasive computing, and high performance computing technology, propose a networktized M&S platform-Cloud Simulation Platform (CSP) to enhance the ability of simulation grid. CSP employs virtualization technology to encapsulate the computing resource, software, simulation models, etc as virtualized simulation resource, masking the heterogeneity of resources. Based on encapsulated resource, virtualization technology can provide enabling technology for dynamic setting up of simulation execution environment, or even simulation system, which will benefit the co-simulation task scheduling and federates’ deployment. Co-simulation task scheduling referred to in this paper focuses on one single task scheduling, in which several subsystems collaboratively run to accomplish the task, and the scheduling here is mainly about the resource scheduling process to construction virtualization-based simulation system. The scheduling in simulation runtime is not discussed in this paper.

ABSTRACT In order to address the problem of single collaborative simulation task scheduling, considering the characteristics of simulation resource encapsulated with virtualization technology, the author first proposed the unified model describing the co-simulation system for co-simulation task, which is the basis of task scheduling; Secondly, based on the unified model, the virtualization-based supporting system, in cloud simulation platform (CSP), of dynamic construction of the co-simulation system were introduced. Thirdly, the scheduling procedure, namely the dynamic construction of the co-simulation system, was discussed. Finally, the primary application example and conclusion was presented. Keywords: cloud simulation platform, HLA federation, co-simulation task scheduling, virtualization technology 1. INTRODUCTION Recently, as an effective tool for understanding and reconstruction of the objective world, M&S theory, methodology, and technology have been well developed, form its own systematic discipline, and their application area are increasingly expanded. At the same time, they are developed toward “digitization, virtualization, networking, intelligence, integration, and collaboration”, which are considered as the characteristics of the modern trend. As the application field of M&S continues to expand, the size and complexity of simulation application system are greatly increased, which poses a severe challenges to M&S technology. High level architecture provides a general framework and corresponding software engineering method “FEDEP” for developing large-scale distributed simulations, which can promote the reusability and interoperability of simulation model. However, HLA does not inherently take into account of resource management and task scheduling for co-simulation, especially the simulation resource is statically bound together with federates before simulation start, in which case the automatic scheduling is lacked. The combination of M&S and grid computing gives birth to simulation grid,

715

Considering the requirement of co-simulation task in CSP, the author first proposes the top-level description model of collaborative simulation system for task, and then the supporting system for cosimulation task scheduling, and the whole procedure of scheduling are then presented. Lastly, the conclusion and further work are given. Without special declaration, the scheduling object in this paper refers to HLA-based co-simulation. 2.

based description of federate. FuncDesc denotes the semantic description of federate. Definition 3. Intera= is defined as the model for interactions between federates. In which Src refers to source federate generating information, Dest denotes the destination federate for interactive information, and Info is the corresponding information to be exchanged. Definition 4. Env= defines the running environment of each federate, including the requirement of OS and software sets. EnvDepend( FMi )=, in which OS refers to the type of operating system such as Windows XP, and SoftSet denotes the software listDŽ Definition 5. Comm= defines the interaction requirement for communication capability of network between federates. Definition 6. Comp= defines the computation requirement of each federate. Definition 7. StopCrit is defined as the condition for stopping simulation running. It can be the times of simulation execution, or the time of simulation running, or the bound of variables in simulation. In fact, the process of dynamic construction of virtualization-based co-simulation system is the realization of the co-simulation task scheduling process. So, the supporting system for scheduling is given as follows.

THE TOP-LEVEL DESCRIPTION MODEL FOR CO-SIMULATION SYSTEM

Co-simulation system is the execution entity for fulfilling the co-simulation task. The efficiency of its running directly determines the effect of simulation, which is the reason of co-simulation task scheduling. For the purpose of high efficient scheduling, we first build the unified top-level model for describing cosimulation system. The model contains the essential properties of co-simulation system, which determines the constitution of federation and interoperability between federates. The model provides solid foundations for dynamic construction of federation, which is the process of the scheduling. Firstly, HLA-based co-simulation system can be described as follows by unified model: Definition 1. Co-simulation system = Definition 2. Fed= is defined as the model of federation, where FM= is defined as the model for describing federates. And |Fed|=Num defines the number of federates in federation. DomainOnt is domain ontology-

3.

THE VIRTUALIZATION-BASED SUPPORTING SYSTEM FOR DYNAMIC CONSTRUCTION OF CO-SIMULATION SYSTEM

Core simulation services

Simulation portal

System toplevel modeling

Simulation task management

Simulation resource management

Automatic federate generation

Requirement parsing for federate

Simulation task scheduling

File deployment

VM control

Simulation run control

Remote desktop connection

Results evaluation & visualization

Simulation resource virtualization

Simulation resource library

Virtualized encapsulating

Physical computing nodes

Node monitoring agent

VMs VM VM VM

Computing resource

Software resource

Model resource

Data resource

VMM

Figure 1: the framework of the supporting system CSP employs virtualization technology to encapsulate agent, the core simulation services in CSP can support the computing resource, software, simulation models, the dynamic construction of co-simulation system. etc as virtualized simulation resource, which is stored in CSP together with the virtualized simulation simulation resource library. Using encapsulated resource constitute Cloud Simulation System (CSS). resource, based on the state information of each The supporting system in CSS includes the logical physical computing node collected by node monitoring functional modules shown in Figure 1. It consists of

716

three main parts: simulation portal, core simulation services, and simulation resource library. • Simulation portal Simulation portal is the entrance point of simulation activities for users. It supports the simulation activities based on internet or desktop, in which users can submit co-simulation task, acquire the results, etc. • Core simulation services Core simulation services are composed of services: system top-level modeling, simulation task management, simulation resource management, automatic federate generation, requirement parsing for federate, simulation task scheduling, file deployment, Virtual machine control, simulation run control, remote desktop connection, results evaluation and visualization, simulation resource virtualization, etc. System top-level modeling service can provide the top-level modeling service for users based on internet or desktop, support to decompose the complex multidisciplinary tight coupled system into several subsystems, and help to describe the models of federation and interactions between federates. Simulation task management service enables the management of co-simulation task. Specifically, the services can support distributed simulation workers to accomplish the simulation task collaboratively. Simulation resource management services support semantic-based simulation resource searching services, and also the downloading, uploading, revising, deleting, registering, publishing service, etc of simulation resource with certain permission. Automatic federate generation services can generate the framework of federate according to the description model of federation. Moreover, it provides user-friendly interface to customize the framework, then user can finish functional entity development in the federate on the basis of the framework. Requirement parsing for federate services can parse and acquire the configuration file of the federate, which includes the semantic description of the federate running environment. Simulation task scheduling services provide the function to monitor the running state of physical machine, and use certain method to choose suitable physical computing nodes for the execution of virtual machines. File deployment services give support to remotely deploy simulation related files to specified file path of the computer with certain IP address. These files together with the running environment constitute the executable simulation system. VM control services lend support to the operation of the virtual machine: start, shutting down, suspending, and resuming the virtual machine. Further, it provides the user to access the virtual machines remotely by desktop connection for performing simulation activities. Simulation run control services support the execution management of simulation federation, which includes creating federation execution, unified start of federate execution, monitoring federates’ state,

synchronizing the logical time of federate. And the function of pausing, resuming, resigning, destroying federation execution, etc is also supported. Simulation resource virtualization services include services of creating different kinds of virtualized simulation resource. Here, virtualized simulation resource mainly refers to virtual machines, into which the simulation resources are encapsulated. • Simulation resource library Simulation resource library is employed to store and manage the virtualized simulation resource, especially the management of VM pools, such as the searching of suitable VMs in library with semantic information. 4.

THE PROCEDURE FOR THE DYNAMIC CONSTRUCTION OF CO-SIMULATION SYSTEM The dynamic construction of co-simulation system is composed of three stages: system modeling stage, simulation model development stage, and simulation model deployment and assembling stage. Strictly the former two stages should not be included. But the former two stages are the basis for the scheduling, and for the aim of easy understanding, we list them out. The three stages will specify the content of each factor in unified model of simulation system. We do not pay much attention to the details of realization, and just focus on the principles in the procedure of federation dynamic construction. The steps are as follows: • System top-level modeling Through the requirement analysis of the simulation task, the decomposition of complex research object is a must. The simulation system can be decomposed into several sub-systems in different domain. The decomposition principle is like this: after decomposition, the sub-systems each should have tight coupling inside and loose coupling with entities outside in order to wipe the bottleneck of communication brought by the irrational decomposition. Using the system modeling services in simulation portal, based on graphical interface, users finish the decomposition of system, and build the description model of interactions between sub-systems. Then, according to certain semantic template, users should accomplish the description of federates (semantic-based domain ontology and function description). Users referred here are mainly chief simulation technology officers. Then, they create the co-simulation task by simulation task management services, with which they upload the files generated in system modeling process. In this step, the factors Fed, Intera, Num, and StpCrit in unified description model are instantiated. • Searching for simulation models based on the description of federates System high-level modeling gives the description of different domain federates, including ontology based domain description and function description. Using

717

simulation task management services, simulation practitioners in different domains first get the federate description. Then according to the description, practitioners search for the professional models of entities using simulation resource management services (or develop professional models on their own). The models are downloaded. However, this is not enough, because normally, the professional models cannot be used directly in HLA-based collaborative simulation. • Automatic federate generation services for HLA-based federate development The professional practitioners employ the system top-level model to generate the framework of each federate using automatic federate generation services. Then based on the federate framework and the professional models, the complete federate which is ready for simulation execution can be developed. Finally, the professional practitioners build the requirement description of federate in these aspects: the running environment, computation and communication requirement, by referring to the description of professional models. In this step, the factors Env, Comm, and Comp in unified description model are instantiated. • Submitting the co-simulation task The practitioners who are responsible for the subsystem development, upload federates and description files to corresponding containers in simulation task management services. The top-level model of co-simulation system, federate models and respective description models together constitute the basis of co-simulation system. After checking that all the subsystems have been well finished, the chief simulation technology officers will submit the simulation task using simulation task management services. • Parsing the description files of the simulation task The requirement parsing for federate services can acquire the running environment of each federate from the configuration files of respective federate, in which the requirement of executing federate is contained, namely: EnvDepend( FMi )= ˈ where FMi denotes federateˈOS refers to the type of operating system, SoftSet represents the software list neededDŽFor the requirement of the whole federation, Env contains the requirement information of all federates. The running environment information of each federate is used to searching for virtual machine image for each federate. • Searching for the virtual machine image based on the information of federate running environment The virtual machine image pool, which is a part of simulation resource library, is a collection of virtual machine images. The virtual machine image pool is located on shared storage. Users can create virtual machine images via simulation resource virtualization services, and upload them to the virtual machine image

pool, at the same time, register and publish the semantic description of them via simulation resource management services. Each virtual machine image is a kind of resource that can be shared and reused with different running environment. So, a large number of virtual machine images in the pool can meet the demand of most users. If not, users can create their own ones based on their special requirement. Simulation resource management services can search virtual machine pools for suitable one in accordance to the semantic description of each federate’ requirement. If there exist more than one virtual machines in the result, then users can click the virtual machines listed to check the detailed information to choose the most suitable ones. • Start virtual machines on suitable physical computing nodes The simulation scheduling services support the state monitoring service of physical computing nodes, such as the CPU utilization, the available Memory, and the bandwidth, the network delay between them. Further, these services can help gather the statistics of historic monitoring information and forecast the load of each physical computing node. In order to realize the load forecasting function, the exponential smoothing algorithm is employed to forecast the load of physical computing nodes, then select suitable nodes to start virtual machines on. The simulation task scheduling services can optimally choose several physical nodes using different algorithms to guarantee enough computing and fast communication capability. Or users can search for physical nodes according to their configuration description, check the historical statistics of monitoring information of them, and then choose several ones manually. VM control services can support the control function of virtual machines on physical nodes, for example start, powering off, suspending, and resuming virtual machines. VM control services first automatically configure the CPU, memory, storage, and network of virtual machines. And then start the virtual machines in chosen physical nodes. • Remote connection to get the desktop of virtual machines After the construction of virtual computing environment, the following demands show that the remote desktop connection services are needed. 1) Before the simulation system running, special requirement leads to changing the configuration of simulation software, or even the OS. 2) According to the demand of interactions between users and simulation system, users need the desktop of virtual machines to implement more sophisticated control on simulation process. 3) In order to guarantee the correctness of simulation results, not only the state monitoring of simulation system is need, but also whether the exceptions are thrown should be pay attention to, because some kinds of exceptions will not lead to the

718

crackdown of simulation system, but will affect the correctness of simulation results. The remote desktop connection services can give support to getting the virtual desktop of virtual machines. After the start of virtual machines, based on these services, users can use the virtual desktop just like it is the local desktop of the physical machine. In other words, users can start and configure simulation software, and build the running environment of federates. Users can also check whether federates have thrown exceptions. • Deploying model files and start the execution of simulation system to accomplish the dynamic construction of simulation system File deployment services can support deployment of files to specified path of machines with designated IP address. Simulation system consists of simulation models and simulation running environment. After the dynamic construction of simulation running environment, the only need is to deploy and configure simulation models (federates). Then the simulation run control services are utilized to create federation execution, and then start federates deployed in virtual machines to join the federation execution. Other simulation run control services such as monitoring the state of federation, control the execution of federation are also provided to uniformly perform the simulation experiment.

Figure 2: system top-level modeling

Figure 3: the decomposition of aircraft landing gear simulation system

5. APPLICATION EXAMPLE The aforementioned virtualization-based dynamic construction of co-simulation system has been primarily applied in multidisciplinary virtual prototype, largescale system collaborative simulation, and high performance simulation areas. This paper presented an application example of aircraft landing gear virtual prototype collaborative simulation system. The main steps are as follows: • System top-level modeling Using desktop virtualization technology, the virtual user interface of software can be acquired through select the desired software, which is shown in Figure 2. Simulation practitioners first get the requirement of simulation task, and then build the system top-level description model. Specifically, the aircraft landing gear simulation system can be decomposed into undercarriage control model, undercarriage multi-body dynamics model, undercarriage hydraulics model, etc, as shown in Figure 3. Each model here is one federate in the federation. And the interactions between these subsystems are built. Then the HLA FOM files and the ontology-based and function description files of professional models are generated. Simulation practitioners use the simulation task scheduling services to create the simulation task by submitting the system top-level description files, as shown in Figure 4.

Figure 4: creation of co-simulation task Completing the development of federates on the basis of searching for domain simulation models Domain simulation practitioners can obtain the system top-level description model files via the simulation task management services. The suitable models are listed by searching the domain models in simulation resource library. Then according to the detailed semantic description of listed domain models, domain simulation practitioners download the ones which meet the requirement. With the help of automatic federate generation services, all federates are then developed using these domain models. Finally, referring to domain model descriptions of required running environment, computing and communication capability, the requirement of federates in co-simulation system are demonstrated as follows in table 1. •

No.

1

719

Table 1: The requirement of federates Simulation Softwa- Comp- Performmodel re uting ance (federate) Enviro- Enviro- Demand nment nment Fluid Fluent Redhat CPU Simulation 8core, Model

memory 4GB 2

Undercarriage Control Model

Matlab

Windo-

CPU

ws

1core, memory 1GB

3

Undercarriage Hydraulic Model

Easy5

Windo-

CPU

ws

1core,

Figure 5: submitting simulation task and dynamic building of simulation running environment • The dynamic construction of co-simulation system After the creation of simulation running environment, the chief simulation technology officers can acquire the virtual desktop of virtual machines by remote desktop connection services. The virtual desktop of virtual machine with “matlab” software inside is shown in Figure 6. The files of each federate are then deployed to certain path in virtual machines. With the help of remote virtual desktop, users can configure OS and software. Here we just let the local RTI component point to central component with certain IP address. Finally, simulation run control services are used to start the co-simulation system and monitoring its running state, as shown in Figure 7.

memory 1GB 4

Undercarriage Multi-body Dynamics Model

Adams

Windo-

CPU

ws

1core, memory 1GB

5

3D Model

CATIA

Windo-

CPU

ws

2core, memory 2GB

In which, Gigabit Ethernet and 80GB storage space is fixed for each node. • Submitting the simulation task and constructing the simulation running environment Federates developed by domain simulation practitioners are submitted using simulation task management services. The chief simulation technology officers submit the simulation task. The supporting system for dynamic construction of co-simulation system in CSP will parsing the requirement information of each federate, search for five suitable virtual machines images (realized using Xen middleware), in which different simulation software are included. CSP will select five physical computing nodes simply using weight sorting method fusing the factors of computation and communication capabilities. Then the virtual machine images are deployed to physical computing nodes according to the same weight sorting method and started with the demand configuration of federates. Figure 5 is the interface of simulation task submitting and dynamic building of simulation running environment.

Figure 6: remote virtual desktop

Figure 7: monitoring and controlling the cosimulation system 6. CONCLUSION AND FURTHER WORK Through introducing virtualization technology, the author presents the supporting system and procedure for simulation task scheduling, and the primary application example. The primary application shows that: virtualization-based dynamic construction of federation can address the constrained scheduling problem caused by the tight coupling between the simulation system and physical computing resources, which will make the co-

720

simulation task automatically scheduled to certain degree. Future work includes as follows: 1) Further research on the semantic-based unified description model and share mechanism of virtualized simulation resource. 2) Further research on high performance collaborative simulation technology to support high efficient execution of simulation system. 3) Further research on the scheduling method when facing a large number of co-simulation tasks.

distributed computing, HLA-RTI, theory and practice of modeling, etc. Xudong Chai was born in 1969. He is a researcher and deputy director at Beijing Simulation Center of Second Academy of Aerospace Science & Industry Co. and council members of Chinese System Simulation Association and National Standardization Technical Committee. His research interests include automatic control and simulation. Bo Hu Li was born in 1938. He is a professor at School of Automatic Science and Electrical Engineering, BeiHang University, and Chinese Academy of Engineering, and the chief editor of “Int. J. Modeling, Simulation, and Scientific Computing”. His research interests include multi-disciplinary virtual prototype, intelligent distributed simulation and cloud manufacturing.

ACKNOWLEDGMENTS This paper is supported by National Defense Key Lab, the National 973 Plan (2007CB310900) and National Defense Pre-research Foundation of China. And authors should like to express the sincere thanks to all colleagues for their help and valuable contribution. REFERENCES Bo Hu Li, et al., 2009. A network modeling and simulation platform based on the concept of cloud computing --- “Cloud Simulation Platform”, Journal of System Simulation, 12 (17), 5292-5299. Hai Jin et al., 2008. Computing system virtualization--theory and application (in Chinese). Tsinghua press. Bo Hu Li, Xudong Chai, Baocun Hou, 2009. Cloud Simulation Platform. Proccedings of the 2009 International Summer Simulation Conference, Turkish, Istanbul. Z. Li, W. Cai, S. J. Turner, and K. Pan, 2007. Federate migration in a service oriented HLA RTI, Procs of international Symposium on Distributed Simulation and Real-Time Applications, 113-121 Bo Hu Li, Xudong Chai, Yanqiang Di, et al., 2005. Research on service oriented simulation grid. International Symposium on Autonomous Decentralized Systems (ISADS 05), pp. 7 – 14. Bo Hu Li, Xudong Chai, Baocun Hou, 2006. Research and Application on CoSim (Collaborative Simulation) Grid. The Proceeding of MS-MTSA’06. Bo Hu Li, Xudong Chai, Wenhai Zhu, 2004. Some focusing points in development of modern modeling and simulation technology. Journal of System Simulation, 16 (9), 1871-1878(in Chinese). IEEE, 2000. Standard 1516 (HLA Rules), 1516.1 (Federate Interface Specification) and 1516.2 (Object Model Template). S.A.Herrod, 2006. The Future of Virtualization Technology, Keynotes of ISCA 2006, http:// www.ece.neu.edu / conf / isca2006 / docs /Herrodkeynote.pdf. AUTHORS BIOGRAPHY Chen Yang was born in 1987. He received his B.S. degree in Beihang University. He is currently a Ph.D. candidate of Beihang University, Beijing, China. Research focuses on advanced distributed simulation,

721

AN OPTIMAL NON-BLOCKING DISPATCHING IN FREE-CHOICE MANUFACTURING FLOWLINES BY USING MACHINE-JOB INCIDENCE MATRIX Ivica Sindiþiü (a), Stjepan Bogdan (b), Tamara Petroviü (b) (a)

(b)

ABB, Bani 72, 10000 Zagreb, Croatia LARICS - Laboratory for Robotics and Intelligent Control Systems, Department of Control and Computer Engineering, Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia (a)

[email protected] , (b)[email protected]

ABSTRACT In this paper we extend a method for resource allocation in particular class of flexible manufacturing system, namely, free-choice multiple reentrant flowlines (FMRF), which is based on MJI matrix. The proposed method is a solution to the problem on how to allocate jobs to resources and how to allocate resources to jobs. A solution is in a form of repeatable resource sequence over the set of resources available for particular choice job. As addition proposed in this paper, the solution is enhanced by the procedure that provides an optimal utilization of resources based on operation price and balanced use of all resources. Although efficiency of the proposed methods have been demonstrated on examples involving manufacturing workcells, the method can be used for other discrete event systems as well, as long as the system under study belongs to free-choice multiple reentrant flowlines class.

of various clustering methods [6]. In [17] we proposed construction of machine-job incidence matrix (MJI) which can be obtained from MPI and DSM matrices. Efficient procedures for determination of simple circular waits (CWs) [7, 8] as well as other important structural properties (which are responsible for stability in the sense of absence of deadlock), such as critical siphons and critical subsystems [9, 11], based on MJI, are presented in [13]. It should be noted that MJI matrix can be straightforwardly transformed in matrix model described in [9]. Generally, scheduling requires a) allocation (dispatching) of available resources to predetermined operations (tasks), and b) definition of sequences in order to provide stable behavior of the system. Usually, supervisory controller not only stabilizes the system (in a sense of deadlock and bounded buffers) but in the same time optimizes some performance criteria. Herein we extend a method for resource allocation in particular class of FMS, namely, free-choice multiple reentrant manufacturing systems (FMRF), which is based on MJI matrix, described in details in [17]. A solution is in a form of repeatable resource sequence over the set of resources available for particular choice job. As addition, proposed in this paper, the solution is enhanced by the procedure that provides an optimal usage of resources based on price and balanced use of all resources.

Keywords: dispatching, manufacturing systems, optimal control 1. INTRODUCTION The first step in the supervisory controller design is modeling of the system and investigation of its structural properties. There are many approaches to modeling and analysis of manufacturing systems, including automata [1], Petri nets [2, 10], alphabetbased approaches, perturbation methods [3], control theoretic techniques, expert systems design, and so on. One way to model relations between tasks in an FMS is in form of Steward sequencing matrix [4], also referred to as design structure matrix (DSM). DSM is a square matrix containing a list of tasks in rows and columns. The order of tasks in rows or columns indicates the execution sequence. Although very useful in production planning, DSM lacks of information related to the resources required for execution of tasks. This aspect of an FMS is captured by the resource requirements matrix [5], also know as the machine-part incidence matrix (MPI). Each column of MPI represents one resource, while rows represent part types processed by the system. The most common usage of MPI is in the field of manufacturing cells design by implementation

2.

SYSTEM DESCRIPTION AND PROBLEM FORMULATION We make the following assumptions that define the sort of discrete-part manufacturing systems: No preemption – once assigned, a resource cannot be removed from a job until it is completed, Mutual exclusion – a single resource can be used for only one job at a time, Hold while waiting – a process holds the resources already allocated to it until it has all resources required to perform a job. Furthermore, we assume that there are no machine failures. Multiple reentrant flowlines (MRF) class of systems, investigated herein, has the following properties: a) each part type has a strictly defined sequence of operations, b) each operation in the system requires one and only one resource with no two

722

consecutive jobs using the same resource, c) there are no choice jobs, d) there are no assembly jobs, e) there are shared resources in the system.

assigned. That is, several resources might be capable and available to perform a specific job (MRF property c is not valid, i.e. there are jobs with choice). We define k

R( J i ) as a set of resources that could be allocated to

2.1. System description Let 3 be the set of distinct types of parts produced (or customers served) by an FMS. Then each part type Pk  3 is characterized by a predetermined sequence of job operations J k

^J

k k k k 1 , J 2 , J 3 ,..., J Lk

`

choice job

An example of FMRF workcell is given in Figure 1. with J = {RP1, BP, MP, RP2} and R = {M1, M2, B, R}. From buffer (job BP), part proceeds to machine M1 or machine M2 (choice job MP). Hence, vector representation of resources that could be allocated to choice job MP is rMP = [1 1 0 0 ]T and R(MP)=sup(rMP)={M1, M2}.

with each operation

employing at least one resource. (Note that some of these job operations may be similar, e.g. J ik and J jk with i z j may both be drilling operations.) We uniquely associate with each job sequence J k the

2.2. Problem formulation Since the system contains shared resources and choice jobs, the scheduling problem discussed in the paper is twofold: i) for a given choice job J ik  J k define

operations of raw part-in, J ink , and completed productk . out, J out

Denote the system resources with R

^ri `in 1 , where

allocation sequence of resources in R( J ik ), and ii) for a given shared resource rsRs with resource job set J(rs), define dispatching policy. Both solutions, allocation sequence and dispatching policy, should be such that overall system is stable in a sense of deadlock. A solution of problem i) offers an answer on how to allocate resources to jobs. For that purpose we propose a result in a form of repeatable resource sequence over the set of resources available for particular choice job. On the other hand problem ii) is related to the number of active jobs in a particular parts of FMRF systems, called critical subsystems. In the chapters that follow we show why critical subsystems are important, how they can be determined from MJI matrix, and how their content (number of active jobs) can be controlled. In fact, solution to problem ii) describes how to allocate jobs to resources.

riR can represent a pool of multiple resources each capable of performing the same type of job operation. In this notation R k  R represents the set of resources utilized by job sequence J k . Note that R ‰ R k and k3

J

‰ J

k3

k

J ik  J k .

represent all resources and jobs in a

particular FMS. Since the system could be re-entrant, a given resource r  R k may be utilized for more than one operation J ik  J k (sequential sharing). Also, certain resources may be used in the processing of more than one part-type so that for some {l, k}3 3, lzk, l k R ˆ R z ‡ (parallel sharing). Resources that are utilized by more than one operation in either of these two ways are called shared resources, while the remaining are called non-shared resources. Thus, one can partition the set of system resources as R Rs ‰ Rns , with Rs and Rns indicating the sets of shared and non-shared resources, respectively. For any rR we define the resource job set J(r). Obviously, J (r ) 1 (> 1) if rRns (rRs). Resource r, with its job set J(r), comprise resource loop L(r), L(r ) r ‰ J (r ) . We define a job vector v : J o  and a resource vector r : R o  that represent the set of jobs and the set of resources corresponding to their nonzero elements. The set of jobs (resources) represented by v (r) is called the support of v (r), denoted sup(v) (sup(r)); i.e. given v = [v1 v2 … vq]T, vector element vi >0 if and only if job vi sup(v). In the same manner, given r = [r1 r2 … rp]T, vector element ri >0 if and only if resource risup(r). Usually, index i is replaced with job (resource) notation, for example, rMA stands for the component of resource vector r that corresponds to resource MA. The definitions of job and resource vectors imply that the job and resource sets should be ordered. MRF class is a special case of FMRF - systems with jobs that do not have predetermined resources

Figure 1: An example of FMRF class of FMS 3.

MACHINE-JOB INCIDENCE MATRIX (MJI)

As we stated in Introduction, DSM is a square matrix containing a list of tasks in rows and columns with matrix elements indicating an execution sequence. The second matrix used for the system description is MPI. It captures relations between resources and parts processed by the system. Since the sequence J

723

k

^J

k k k k 1 , J 2 , J 3 ,..., J Lk

`

represents part Pk processing

order, by combining DSM and MPI matrices, we get a general form of machine-job incidence matrix ȁ for an FMRF system [17]. In case job i is performed by resource j, matrix element (i, j) is equal to ‘1’, otherwise is ‘0’. For an MRF system, each operation in the system requires only one resource (there are no choice jobs), hence, exactly one element ‘1’ would appear in each row of MJI matrix. On the other hand, column representing shared resource comprises multiple entries of ‘1’.

J

1 1 1 2

J  J L11 ȁ=

J12  J L22

 J1m  J Lmm

R1

R2



Rq

0/1

0/1



0/1

0/1

0/1



0/1





0/1



0/1

0/1

0/1



0/1



0/1



0/1



0/1

 

0/1



0/1



0/1



0/1



0/1

ȁ



0/1

0/1

In this Chapter we are interested to determine repeatable resource sequence over the set of resources available for execution of each choice job in the system. The sequence should prevent conflicts and deadlocks simultaneously. In the analysis that follows we consider one part-type regular FMRF with each buffer capable of holding one part at a time. Let MJI matrix of the system is given as

  

T

ȁ=

M1 0 0 1 0

M2 0 0 1 0

B 0 1 0 0

R 1 0 0 1

It can be seen that robot R is shared resource as the corresponding column has two elements equal to ‘1’. As we demonstrate in [17] one of the benefits provided by MJI matrix is reduction of computational complexity in FMRF system analysis and simulation. 4.

0 0 1 0 1 0

0 0 0 0 1 0

0 0 0 0 1 0

M1 M 2 M 3 M 4 M 5

ª¬ 1 ȁ T 2 ȁ T ... m ȁ T º¼ (1) For the system given in Figure 2. MJI attains the following form:

RP1 BP MP RP2

0º J 1 0 »» JB1 0» J 2 » 0 » JB 2 0» J 3 » 0 » JB 3 1 0 1 0 0 0 0 0» J 4 » 0 0 0 0 0 0 0 1 » JB 4 » 0 0 0 0 0 0 0 0¼ J 5 0 0 1 0 0 0

0 1 0 0 0 0

0 0 0 1 0 0

0 0 0 0 0 1

B1

B2

B3

(2)

B4

From the matrix we see that a part production requires sequence of 5 jobs, and system is composed of 5 machines and 4 buffers. For choice jobs J2, J3 and J4, we have R(J2)={M2, M3}, R(J3)={M3, M4, M5}, and R(J4)={M2, M4}. All machines, except M5, are shared resources. There are many part routes that complete the required job sequence - to mention just few of them: ı1={M1ĺM2ĺM3ĺM2ĺM1}, ı2={M1ĺM3ĺM5ĺM4ĺM1}, ı3={M1ĺM3ĺM3ĺM4ĺM1}, and so on. We partition set of part routes, denoted Ȉ, into two disjoint sets, Ȉ = ȈR U ȈNR, with ȈR comprising all part routs with multiple use of resources allocated to choice jobs, and ȈNR containing all part routes without multiple use of resources. In our example ı1 and ı3 belong to ȈR, while ı2 is an element of ȈNR. It is obvious that conflicts might occur in ıi  ȈR since same resource is used for execution of more than one job. In ı1 resource M2 is used for jobs J2 and J4. On the other hand part routes in ȈNR are inherently conflict free. We define resource sequences as combination of several routes from ȈNR. As a result, the structure of repeatable resource sequences over the set of resources available for execution of choice jobs would, by itself, provide not only conflict free but also deadlock free behavior of the system. The number of all possible part routs, N = | Ȉ|, is

Machine-job incidence matrix can be defined separately for each part type in an FMS.In that case overall MJI matrix can be written as: ȁ

ª1 «0 « «0 « «0 «0 « «0 «0 « «0 « ¬1

MJI AND RESOURCE SEQUENCING

defined as N

n

m

i 1

j 1

– (¦ ȁ i , j )

where

n

and

m

correspond to the number of rows and columns of ȁ , respectively. In our example one has N = 1·1·2·1·3·1·2·1·1=12. For each part route ıi we define

As already mentioned, in addition to the assumptions made at the beginning of Chapter II, a general class of FMRF systems has the following nonrestrictive capabilities: i) some jobs have the option of being machined in a resource from a set of resources (allocation of jobs), ii) job/part routings are NOT deterministic, iii) for each job there exists a material handling buffer (routing resource) that routes parts.

MJI sub-matrix ȁ in a way that in case of multiple entries ‘1’ in row (choice job), sub-matrix comprises only the one that corresponds with resource belonging to the part route. For ı2={M1ĺM3ĺM5ĺM4ĺM1} MJI sub-matrix attains a form i

724

ȁ2

ª1 «0 « «0 « «0 «0 « «0 «0 « «0 « ¬1

0 0 0 0 0 0 0 0 0

0 0 1 0 0 0 0 0 0

0 0 0 0 0 0 1 0 0

M1 M 2 M 3 M 4

0 1 0 0 0 0 0 0 0

0 0 0 1 0 0 0 0 0

0 0 0 0 0 1 0 0 0

0º J 1 0 »» JB1 0» J 2 » 0 » JB 2 0» J 3 » . 0 » JB 3 0» J 4 » 1 » JB 4 » 0¼ J 5

M 5 B1

B2

B3

B4

0 0 0 0 1 0 0 0 0

S J2

S J3

S J4

M1

M2 M3

form of

ȁ 2 from

0º J 2 1 »» J 3 . 0 »¼ J 4

M4 M5

Now, let us suppose that resource allocation policy requires that resources M3 and M2 should be used repetitively, one after the other, for execution of job J2. This repeatable resource allocation sequence can be written in a form of matrix ª0 0 1 0 0 º S J2 « » ¬0 1 0 0 0 ¼ , M1

ª0 0 1 0 0 º «0 0 1 0 0 » « » «¬0 1 0 0 0 »¼

ª0 0 0 1 0 º «0 0 0 0 1 » « » «¬0 0 1 0 0 »¼ ª0 1 0 0 0 º « » «0 1 0 0 0 » . «¬0 0 0 1 0 »¼

In order to get a better insight in the system behavior, Figure 2. presents how parts pass through the line. It should be noted that sequences are executed in a way that new allocation of resources (new step) is performed after all jobs are finished and parts reside in the buffers. First part enters the system and it is processed by M3 (J2M3-‘1’ in the first row of SJ2), than, it proceeds to job J3 on M4 (J3M4-‘1’ in the first row of SJ3), and than to job J4 on M2 (J4M2--‘1’ in the first row of SJ4). The second part enters the system and it is processed by M3 (J2M3-‘1’ in the second row of SJ2), than, it proceeds to job J3 on M5 (J3M5-‘1’ in the second row of SJ3), and so on. Conflict occurs for the first time at k+2 as J4 on part 1 is planned for M2 while, in the same time, the third part, that just entered the system, requires the same machine (J2 on M2). This example clearly demonstrates that repeatable resource sequences can lead the system in conflict. In [17] it has been shown how to determine conflict-free set ‚ ^S j ` . Furthermore, we proved that usage of

Having defined sub-matrices, search for sequences in Ȉ is in fact search for all MJI sub-matrices that are characterized by single entry ‘1’ in each row and multiple or no entry ‘1’ in each column. It is easy to show that there exists a procedure of complexity O(N) for calculation of such matrices. Resource sequences are related to choice jobs, for this reason, we introduce reduced form of MJI submatrices, denoted ȁ*i , such that encompass only rows corresponding to those jobs, and without columns related to the buffers. Reduced previous example is given as ª0 0 1 0 ȁ*2 «« 0 0 0 0 «¬ 0 0 0 1

ª s1J 2 º « 2 » « sJ 2 » « sJ3 2 » ¬ ¼ ª s1J 3 º « 2 » « sJ 3 » « sJ3 3 » ¬ ¼ ª s1J 4 º « 2 » « sJ 4 » « sJ3 4 » ¬ ¼

conflict-free set of resource allocation sequences not only resolves possible conflicts in the system, but also has direct consequence on the system stability, i.e. it provides deadlock-free behavior of the system.

M2 M3 M4 M5

or in general form

Sj

ª s1j º « 2» «sj » . «» « Z» ¬« s j ¼»

(3)

Figure 2: Presentation of parts passing through the line.

Resource allocation sequence matrix should be defined for each choice job in the system. Our goal is to find set ‚ ^S j ` (where w = ‚ equals to the number

4.1. Sequence optimization In order to determine an optimal sequence we introduce cost matrix ȁ W . The cost matrix, with the form

of choice jobs in the system) such that system has no conflicts and it is deadlock free. In [17] it has been proven that elements of such set, i.e. sequence matrices, are formed of rows of MJI sub-matrices. As an example, let us examine 3-step sequence for system (2), defined by the following resource allocation sequence matrices,

identical to MJI matrix ȁ , captures cost of execution of ith job by using jth resource. By using cost matrix we can extract cost of each MJI sub matrix as n

C

p

n

¦¦ (ȁ i 1 j 1

725

p W ij

) ,

(4)

where ȁ W p

ȁ W ˜ ( ȁ p )T , ȁ p  6 NR , p 1..k , is

mixed integer linear programming problem which can be solved using standard algorithms.

an MJI submatrix. If one assumes that sequence matrices are comprised of Z rows (sequence of Z repeatable steps), where each MJI submatrix ȁ be

Zp

used

Z

i.e.

times,

p

4.2. Case study The proposed method has been tested on the system presented in [15] and [16]. Although, this example has only two choice jobs and it is comprised of MRF and FMRF sub-systems, it has been chosen so that the proposed method can be compared with various control techniques already implemented on this particular manufacturing system. The system’s PN model is shown in Figure 3. The system has 3 part types, P1, P2 and P3, 4 machines M1-M4, and 3 robots R1-R3. Part routes for P1 and P2 are predetermined (MRF), while P3 has choice jobs (FMRF). All resources, except for M1, are shared (only utilization of shared resources will be optimized).

will

k

¦Z

p

,

p 1

Zp  , 0 d Zp d Z ,

then the total cost generated

by those sequences is k

¦Z C

C tot

p

p

.

(5)

p 1

It is clear that minimization of the total cost, defined as (5), is trivial problem – one should use only ȁ

p

with

p

the smallest C in order to achieve minimal cost. However, in that case it might happen that utilization of the system resources would be highly unbalanced or even some resources would not be used at all. Hence, the cost function should be extended with a relation that captures resources utilizations. For each MJI sub-matrix one can define a resource utilization vector as

up

1T ˜ ȁ p ,

(6) where 1mx1 is vector with all elements equal to 1. As a result resource utilization vector ui is a binary vector with jth element equal to 1 if corresponding resource participates in execution of the sequence containing rows of ȁ sub-matrix. Finally, an integer vector that represents overall usage of the system resources is determined as i

k

u

¦Z u p

p

.

(7)

Figure 3: PN model of the system used for the case study [15].

p 1

Now, the second objective, balanced usage of the system resources, can be defined in the following form

(1  H ) ui (1  H ) d d , i, j 1..q, (1  H ) u j (1  H )

The goal is to find the optimal sequence that includes all shared resources in the system for Z 6 and H 0.2 .

(8)

Such value of İ gives 0.8 d

where ui and uj are ith and jth component of u, İ is design parameter such that 0 d H  1 , and q is the number of resources that should be balanced (for q = m all resources in the system shall be included in optimization). Fully balanced utilization of resources is achieved if ui

uj

d 1.2 , i.e. it is required

cM11=6, cM31=4, cM22=5, cM42=9. Three MJI sub-matrices have been used for construction of sequences such that u1= [1 1 0]T, u2= [0 0 1]T and u3= [0 1 1]T, where components correspond with resources M2, M3 and M4. This gives overall usage of shared resources as

might be very difficult (in some cases even impossible) to obtain, which depends on the system structure and executed sequence. Hence, by introducing parameter İ one is able to relax rigorous balancing requirement – for İ § 1 the system could become unbalanced, while for İ = 0 one requires full balance of the system resources exploitation. Minimization of (5) by varying Z p , p 1,..., k

Z

uj

that usage of resources is balanced. The following costs have been used in optimization:

1, i, j 1..q . However, this goal

under conditions (8) with predefined İ and

ui

[Z1 Z1  Z3 Z2  Z3 ]T . Optimization yields to the following values: Z1 3 , Z2 2 , Z3 1 with total usage of resource within the u

Z1u1  Z2u 2  Z3u 3

sequence equal to u = [3 4 3]T as it is shown in Figure 5. Results obtained by simulation with MJIWorkshop, software tool presented in [12], are shown in Figures 5

is a

726

[3] Y.C. Ho, X.R. Cao, Perturbation Analysis of Discrete Event Dynamic Systems, Kluwer Academic Publishers, 1991. [4] D.V. Steward, The Design Structure System: A Method for Managing the Design of Complex Systems, IEEE Transactions on Engineering Management, 28, 1981, pp. 71-74. [5] A.Kusiak, J. Ahn, Intelligent Scheduling of Automated Machining Systems, Computer-Integrated Manufacturing Systems, 5, 1, 1992, pp. 3-14. [6] T.R. Browning, Applying the Design Structure Matrix to System Decomposition and Integration Problems: A Review and New Directions, IEEE Transactions on Engineering Management, 48, 3, 2001, pp. 292-306. [7] M.D. Jeng, F. DiCesare, Synthesis Using Resource Control Nets for Modeling Shared-Resource Systems, IEEE Trans. Rob. Autom. RA-11, 1995, pp. 317–327. [8] R.A. Wysk, N.S. Yang, S. Joshi, Detection of Deadlocks in Flexible Manufacturing Cells, IEEE Trans. Rob. Autom, 7, 6, 1991, pp. 853í859. [9] S. Bogdan, F.L. Lewis, J. Mireles, Z. Kovacic, Manufacturing Systems Control Design: a matrix based approach, Springer, 2006. [10] M.V. Iordache, P.J. Antsaklis, Supervisory Control of Concurrent Systems: A Petri Net Structural Approach, Birkhauser, Boston, USA – 2006. [11] M.C. Zhou, M.P. Fanti, Deadlock resolution in computer-integrated systems, Marcel Dekker/CRC Press, New York 2005. [12] I. Sindiþiü, T. Petroviü, S. Bogdan, Modeling and Simulation of Manufacturing Systems based on MachineJob Incidence Matrix, Proc. Int. Conference on Mathematical Modelling, Vienna, 2009. [13] T. Petroviü, S. Bogdan, I. Sindiþiü, Determination of Circular Waits in Multiple-Reentrant Flowlines based on Machine-job Incidence Matrix, Proc. European Control Conference, Budapest, 2009. [14] S. Lee, D.M. Tilbury, Deadlock-Free Resource Allocation Control for a Reconfigurable Manufacturing System With Serial and Parallel Configuration, IEEE Trans. on SMC-part C, 37, 6, 2007, pp.1373-1381. [15] ZhiWu Li, MengChu Zhou; Two-Stage Method for Synthesizing Liveness Enforcing Supervisors for Flexible Manufacturing Systems Using Petri Nets, IEEE Transactions on industial informatics, Vol. 2, No. 4, November 2006, pp. 313-325 [16] I.Sindiþiü, S.Bogdan, T.Petroviü; Dispatching in Freechoice Multiple Reentrant Manufacturing Flowlines by using machine-Job Incidence Matrix; 6th IEEE Conference on Automation Science and Engineering – CASE 2010, p617-623, Toronto, Canada, 22-24. August 2010 [17] I.Sindiþiü, S.Bogdan, T.Petroviü; Resource Allocation in Free-choice Multiple Reentrant Manufacturing Systems Based on Machine-job Incidence Matrix;, IEEE Transactions on Industrial Informatics, Vol. 7, No. 1, 2011, pp. 105-114.

and 6. It can be seen from Figure 6 that system is deadlock free, i.e. flow of the parts is uninterrupted.

Figure 5: Utilization of M2, M3 and M4 within the sequence.

Figure 6: Utilization of all resource in the system. 5.

CONCLUSION

In a finite-buffer flexible manufacturing systems, any dispatching policy for interrupted part flow has to essentially take into account the composition of the interconnection between jobs and resources. The proposed optimal non-blocking dispatching policy is based on machine-job incidence matrix (MJI), obtained from Steward sequencing matrix and Kusiak machinepart incidence matrix, and explained in details in [17]. Since FMRF systems contain shared resources and choice jobs, a solution to allocation of resources to jobs is determined in a form of repeatable resource sequence over the set of resources available for particular choice job. Obtained sequences not only stabilize the system but provide an optimal utilization of resources based on price and balanced use of all resources. Efficiency of the proposed method has been demonstrated on an example involving multi-part type manufacturing system. REFERENCES [1] W.M. Wonham, Supervisory Control of Discrete Event Systems, Lecture notes, 2005. [2] T. Murata, Petri nets: properties, analysis and applications, Proc. IEEE, 77, 4, 1989, pp. 541–580.

727

SIMULATION OF VASCULAR VOLUME PULSATION OF RADIAL INDEX ARTERY Pichitra Uangpairoj(a), Masahiro Shibata(b) (a,b)

Department of Bio-science and Engineering, College of Systems Engineering and Science, Shibaura Institute of Technology, Japan. (a)

[email protected], (b)[email protected]

ABSTRACT

pressure-volume relationship from many in vitro tests

This paper presents an application of finite element

(Cox 1978a-c; Carew, Vaishnav and Petal 1968). They

simulation with the analysis of arterial stiffness. The

investigated the volume change of blood vessel by

influences of the intravascular pressures on arterial wall

increasing the intravascular pressure and the transmural

which behaves like hyperelastic material was investigated

pressure. The intravascular pressure was measured by

by using Mooney-Rivlin hyperelastic constitutive model

using pressure transducer. Meanwhile, in vivo tests, the

with the finite element solver of LS-DYNA. The results

pressure-volume relationship can be obtained from

were obtained in the forms of nonlinear pressure-diameter

photoelectric

relationship. Moreover, the diameter variation of arterial

(Kawarada et al. 1986). PPG system is compatible with

model corresponds to the pulsatile blood pressure. But the

the clinical application. It is a non-invasive measurement

distensibility of artery reduces when the level of pulsatile

system and easy to use. To evaluate the arterial blood

blood pressure increases. These numerical results are

volume change, the other tissues are considered to be

expected to clarify an assessment of the arterial stiffness

incompressible. The venous system is collapsed by the

using photoelectric plethysmograph.

exertion of external pressure. At the same time, the

plethysmographic

(PPG)

technique

changes of arterial blood volume can be controlled by the Keywords: arterial stiffness, vascular volume change,

decrease in external pressure and the increase in

finite element simulation

transmural pressure. With these assumptions and the Lambert-Beer's Law, Kawarada et al. (1986) and Ando et

1. INTRODUCTION

al. (1991) could investigate arterial elasticity by detecting

Arterial stiffness is associated with the development of

the arterial volume change at any changes in transmural

cardiovascular risk factors. It is one of the indices which

pressure from DC signal of PPG.

of

In Biomechanics, arterial stiffness associates with the

cardiovascular system in both of research and clinical

mechanical properties of arterial wall which consists of

applications.

three main layers; tunica intima, tunica media and tunica

are

used

to

diagnose

the

pathophysiology

Arterial stiffness is typically investigated by monitoring

adventitia. It is believed that elastin and smooth muscle

the arterial motion in the circumferential direction. The

cell in media layer assist artery to resist high loads in the

instantaneous

circumference

circumferential direction. At the same time, the thick

corresponds to the arterial pressure pulse which can be

bundles of collagen fibres in adventitia also contributes

seen in the form of pressure-diameter relationship or

significantly to protect artery from overextension and

change

of

vessel

728

rupture when artery is exerted by force from blood

15,360 hexahedron solid elements by using LS-Prepost

pressure. With the structure of arterial wall, it makes

version 2.1. All elements were assumed to be the constant

artery exhibits hysteresis under cyclic loading, stress

stress solid element in order to avoid volumetric locking

relaxation under constant extension and creep under

effect as shown in Figure 1.

constant loads. This behavior can identify artery to be viscoelastic. However, arteries are frequently considered simply as hyperelastic

material and all inelastic

phenomena are neglected. Therefore, the constitutive models of arterial wall have been developed regarding the hyperelasticity and the distribution of collagen fibres which reflects an orthotropic property of arterial wall in both of microstructure (Bischoff, Arruda and Grosh 2002, Zhang et al. 2005) and

macrostructure (Gasser, Ogden

and Holzapfel 2006). Figure 1: The geometrical model of radial index artery

With the high performance of current computers, the constitutive models for arterial wall and finite element method have been widely implemented to observe the

2.2 The finite element method

responses of arterial wall to various types of load. These

The commercial explicit finite element solver of

observations have been utilized in many clinical

LS-DYNA 971 was employed to investigate the responses

application (Xia, Takayanagi and Kemomochi 2001;

of arterial wall on various types of pressure which are

Zhang et al. 2005, 2007; Zhao et al. 2008).

given in the following form:

In this study, the responses of arterial wall to the loads z

have been observed by coupling anisotropic hyperelastic

Intravascular pressures which represented mean

constitutive models with the finite element method. These

blood pressure were assigned to be 10, 20, 30, 40, 50,

numerical results are expected to clarify the assessment of

70, 90, 100, 120, 140, 160 and 180 mmHg at the

the arterial stiffness using PPG.

inner wall of the tube as the ramped loads. The pressure raised from 0 to the maximum pressure of

2. MATERIAL AND METHODS

each level within 100 ms. This was done to obtain

2.1 The finite element model of artery

the pressure-diameter relationship of the artery. z

A radial index artery was considered in order to apply this

Pulsatile loads which represented pulsatile blood

study with PPG application. A tube with three layers of

pressures were given in the form of sinusoidal

arterial wall; tunica intima, tunica media and tunica

function to simplify the arterial pressure pulse.

adventitia, was combined into one-layer to simplify ഥ ൅ ƒ’ •‹༌ሺʹπˆ–ሻ ሺ–ሻ ൌ 

anatomical structure of artery. The outer diameter of the

(1)

tube was 1.54 mm (Bilge et al. 2006). The ratio of total wall thickness to outer diameter was 0.189 mm

ഥ where P(t) is the instantaneous pressure (mmHg), 

(Holzapfel et al. 2005). The length of the tube was

is mean level of the pressure pulse (mmHg). The

specified to be 10 mm. This model was discretised into

ഥ was assigned to be 70, 100 and 120 level of 

729

mmHg which represented low blood pressure,

(4)

‫ܝ‬ൌ‫ܝ‬

normal blood pressure and high blood pressure conditions, respectively. Pamp is the amplitude of

where u is the displacement. When x+ = x-The contact discontinuity:

pulsatile pressure (mmHg) which relates to systolic pressure and diastolic blood pressure. In this study,

‫ ۼ‬ή ሺો൅ െ ોെ ሻ ൌ Ͳ

the systolic/diastolic blood pressures were 90/50,

(5)

120/80 and 140/100 mmHg for low blood pressure, normal blood pressure and high blood pressure

The Cauchy stress in the momentum balance equation

conditions, respectively. f is frequency (Hz) and t is

related to the strain energy function through the

time (s). Pulsatile pressures were also applied at the

constitutive equation.

inner wall of the tube to investigate the responses of

Constitutive Equations for incompressible hyperelastic

arterial wall on pulsatile pressure. z

material can be expressed as follows:

External pressure was assume to be zero in this simulation.

ોൌ

These specifications were considered as traction boundary conditions

of

boundary value

μ‫܅‬ μ۴

െ ‫݌‬۷

(6)

where  is the strain energy function, F is the

problem.

deformation gradient tensor, p is the Lagrange multiplier

Meanwhile, the displacement boundary condition was

and I is the identity tensor.

applied at the annulus of the tube which constrained

The strain energy function for arterial wall was defined

movement in all direction.

by a Mooney-Rivlin hyperelastic constitutive model as

The numerical models were based on solving the

follow:

momentum balance equation and boundary conditions ܹ ൌ ‫ Ͳͳܥ‬ሺ‫ ͳܫ‬െ ͵ሻ ൅ ‫ ͳͲܥ‬ሺ‫ ʹܫ‬െ ͵ሻ ൅ ‫ Ͳʹܥ‬ሺ‫ ͳܫ‬െ ͵ሻʹ

which are given as follows:

൅‫ ͳͳܥ‬ሺ‫ ͳܫ‬െ ͵ሻሺ‫ ʹܫ‬െ ͵ሻ ൅ ‫ Ͳ͵ܥ‬ሺ‫ ͳܫ‬െ ͵ሻ͵

(7)

The momentum balance equation: where C10, C01, C20, C11 and C30 are hyperelastic ‫ ׏‬ή ો ൅ ρͲ ܎ ൌ ρͲ ƒ

(2)

coefficients used for artery. I1 and I2 are invariants which can be expressed as

where ો is the Cauchy stress, F is the deformation gradient, ρͲ is the mass density, f is the body force and a

‫ ͳܫ‬ൌ ‫ݎݐ‬۰

(8)

is the acceleration. ͳ

‫ ʹܫ‬ൌ ሾ‫ ʹͳܫ‬െ ‫ݎݐ‬ሺ۰ ʹ ሻሿ

The traction boundary condition:

ήોൌ‫܂‬

ʹ

(9)

where B is the left Cauchy-Green deformation tensor.

(3)

The hyperelastic coefficients of arterial wall which were reported by Loree et al. (1994) are shown in Table 1

where ‫ ܂‬is traction force and N is the unit normal vector. The displacement boundary condition:

730

increases with transmural pressure nonlinearly.

Table 1: hyperelastic coefficients of arterial wall

Artery

C10

C01

C20

C11

C30

This result confirms that the simulation model of artery

[KPa]

[KPa]

[KPa]

[KPa]

[KPa]

behaves similarly to the real artery, it is also suitable to

708.416

-620.042

0

2827.33

0

apply this model with the investigation of the influence of pulsatile pressure on arterial wall.

2.3 Analysis of results

100

Pressure (mmHg)

In this simulation, the change in diameter of artery was obtained from the average movement of radial position of the inner wall at every 1 mm length in axial direction using post-processing of LS-DYNA software.

90 80 70 60 50 40

3. RESULTS AND DISCUSSIONS

0.5

1

1.5

2

time (s)

200 180 160 140 120 100 80 60 40 20 0

(a) 1.33 1.32

Diameter (mm)

Pressure (nnHg)

0

1.31 1.3 1.29 1.28 1.27 1.26

1.24000 1.26000 1.28000 1.30000 1.32000

1.25

Diameter (mm)

1.24 0

0.5

Figure 2: Pressure-diameter relationship of artery

1

time (s)

1.5

2

(b) Figure 2 shows the relationship between intravascular

Figure 3:

pressure and diameter of arterial tube which was obtained

Diameter-time relationship

by varying the level of intravascular pressure as in section

condition.

2.1. This curve shows that using Mooney-Rivlin

130

hyperelastic

120

generates

non-linear

Pressure (mmHg)

model

pressure-diameter relationship. At the lower level of intravascular pressure, 10-70 mmHg, the distensibility of arterial model is higher than the upper level pressure, 90-180 mmHg. This shows that arterial model is stiffer at higher pressure level. This result corresponds to the

(a) Pressure-time relationship and (b) of low blood pressure

110 100 90 80 70

relationship between the transmural pressure and volume

0.00

0.50

1.00

elastic modulus of rabbit artery from the experiment of

time (s)

Kawarada et al. (1986) that the elastic modulus of artery

(a)

731

1.50

2.00

Diameter (mm)

1.34

the amplitude of pulsatile pressure, the amplitude of

1.32

diameter variation in high blood pressure condition is the

1.30

lowest one (1.14% of mean diameter). Meanwhile, the

1.28

amplitude of diameter variation in normal blood pressure

1.26

condition (1.28% of mean diameter) is also lower than the

1.24

amplitude in low blood pressure condition (1.62% of 0.00

0.50

1.00

time (s)

1.50

2.00

mean

These

are

because

the

arterial

distensibility decreases when the level of intravascular

(b) Figure 4:

diameter).

pressure

(a) Pressure-time relationship and (b)

increases

which

correspond

to

the

pressure-diameter relationship.

Diameter-time relationship of normal blood pressure

The investigation of the responses of arterial diameter

160 150 140 130 120 110 100 90 80

on pulsatile pressure is just only the first step to apply

Pressure (mmHg)

condition.

finite element method with the analysis of arterial elasticity. To apply the simulation results with the analysis of arterial elasticity using PPG. The further simulations need to be carried on for the future work. For example, 0.00

0.50

1.00

1.50

the pulsatile pressure should be modified to be similar to

2.00

the real pulsatile blood pressure. Moreover, the influence

time (s)

Diameter (mm)

(a)

of external pressure (or cuff pressure for PPG) on diameter of arterial model also need to be studied.

1.33 1.32 1.31 1.30 1.29 1.28 1.27 1.26 1.25 1.24

REFERENCES Ando, J., Kawarada, A., Shibata, M., Yamakoshi, K. and Kamiya, A., 1991. Pressure-volume relationships of 0.00

0.50

1.00

time (s)

1.50

finger arteries in healthy subjects and patients with

2.00

coronary atherosclerosis measured non-invasively by

(b) Figure 5:

photoelectric

plethysmography.

Japanese

Circulation Journal 55: 567-575.

(a) Pressure-time relationship and (b)

Bilge, O., Pinar, Y., ሷ zer, M. A. and G‘ሷ vsa, F., 2006. A

Diameter-time relationship of high blood pressure

morphometric study on the superficial palmer arch

condition.

of the hand. Surg Radiol Anat 28: 343-350. Figure 3, 4 and 5 show the responses of arterial

Bischoff, J. E., Arruda, E. A. and Grosh, K., 2002. A

diameter on pulsatile pressure in low blood pressure,

microstructurally based orthotropic hyperelastic

normal blood pressure and high blood pressure conditions,

constitutive law. Transactions of the ASME 69:

respectively. The results show that arterial diameters vary

570-579.

with pulsatile pressure in all conditions. The difference of

Carew, T. E., Vaishnav, R. N. and Patel, D. J., 1968.

arterial response in each blood pressure condition is the

Compressibility of the arterial wall. Circulation

amplitude of diameter variation. Within the same level of

Research 23: 61-68.

732

Cox, R. H., 1978a. Passive mechanics and connective

microstructural hyperelastic model of pulmonary

tissue composition of canine arteries. Am J Physiol

arteries under normo- and hypertensive conditions.

234: H533-541.

Annals

of

Biomedical

Engineering

33(8):

1042-1052.

Cox, R. H., 1978b. Differences in mechanics of arterial

Zhang, Y., Dunn, M. L., Hunter, K. S., Lanning, C., Ivy,

smooth muscle from hindlimb arteries. Am J Physiol

D. D., Claussen, L., Chen, S. J. and Shandas, R.,

235: H649-656. Cox, R. H., 1978c. Regional variation of series elasticity

2007. Application of a microstructural constitutive

in canine arterial smooth muscle. Am J Physiol 234:

model of the pulmonary artery to patient-specific

H542-551.

studies: validation and effect of orthotropy. Journal of Biomechanical Engineering 129: 193-201.

Gasser, T. C., Ogden, R. W. and Holzapfel, G. A., 2006. Hyperelastic modelling of arterial layers with

Zhao, A. R., Field, M., Digges, K. and Richens, D., 2008.

distributed collagen fibre orientations. J R Soc

Blunt trauma and acute aortic syndrome: a

Interface 3: 15-35.

three-layer finite element model of the aortic wall. European Journal of Cardio-thoracic Surgery 34:

Holzapfel, G. A., Gasser, T. C. and Ogden, R. W., 2000. A

623-629.

new constitutive framework for arterial wall mechanics and a comparative study of material models. J Elasticity 61: 1-48.

AUTHORS BIOGRAPHY

Holzapfel, G. A., Sommer, G., Gasser, T. C. and Regitnig,

Pichitra Uangpairoj received the B.Sc. (2007) in Food

P., 2005. Determination of layer-specific mechanical

Technology

properties

with

Engineering from Suranaree University of Technology,

nanotherosclerotic intimal thickening and related

Thailand. She is the PhD student in the Department of

constitutive modeling. Am J Physiol Heart Circ

Bioscience and Engineering, Shibaura Institute of

Physiol 289: H2048-H2058.

Technology, Japan. Her current interests include the

of

human

coronary

arteries

and

M.E.

(2010)

in

Mechanical

applications of numerical simulation in arteries.

Kawarada, A., Shimazu, H., Yamakoshi, K. and Kamiya, A., 1986. Noninvasive automatic measurement of

Masahiro Shibata received the BS in Applied Physics

arterial elasticity in human fingers and rabbit

and his PhD in Biomedical Engineering from Hokkaido

forelegs using photoelectric plethysmography. Med.

University, Japan. Since 2008 he has been with the

& Biol. Eng. & Comput 24: 591-596.

Department of Bio-Science and Engineering, Shibaura

Loree, H. M., Grodzinsky, A. J., Park, A. J., Gibson, L. J.,

Institute of Technology, where he is a Professor of

Lee, R. T., 1994. Static and circumferential

System Physiology. His research interests include oxygen

tangential

dynamics in microcirculation.

modulus

of

human

atherosclerotic

plaques. Circulation 27: 195-204. Xia, M., Takayanagi, H. and Kemmochi, K., 2001. Analysis

of

multi-layered

filament-wound

composite pipes under internal pressure. Composite Structures 53: 483-491. Zhang, Y., Dunn, M. L., Drexler, E. S., McCowan, C. N., Slifka, A. J., Ivy, D. D. and Shandas, R., 2005. A

733

CODING TCPN MODELS INTO THE SIMIO SIMULATION ENVIRONMENT Miguel Mujica(a), Miquel Angel Piera(b) (a,b)

Autonomous University of Barcelona, Faculty of Telecommunications and Systems Engineering, 08193, Bellaterra, Barcelona (a)

[email protected], (b)[email protected]

ABSTRACT The coloured Petri net formalism has been widely used by scientific community to perform not only research and behavioural analysis of models but also as a simulation tool to carry out systems' analysis. Its characteristics allow a better understanding of the causal relationships present in systems. On the other hand discrete event system simulation software has evolved in order to reduce the efforts needed in simulation projects; the developers have improved the developing paradigm, graphical interface and analysis tools among other aspects. Unfortunately when dealing with big projects the simulation paradigms hinder the understanding of the causal relationships of systems. In this paper a way to integrate the timed coloured Petri net modelling formalism with the SIMIO simulation software is presented in order to overcome this problem.

improved if some underlying logic is added in order to achieve a better understanding of the modelled system. The timed coloured Petri net formalism (TCPN) has characteristics that allow modelling true concurrency, parallelism or conflicting situations present in industrial systems (Jensen 1997, Moore & Gupta 1996, Mušiþ, 2009). Unfortunately when the TCPN formalism is used with the purpose of systems analysis it lacks of tools that can be used to perform statistical analysis by a user who is not expert in the TCPN field (industrial engineers, process engineers, managers, etc). Furthermore if the model or the results are going to be presented to decision makers within a firm the graphical representation results difficult to understand when the people is not familiar with the formalism. In the best cases the token game is the one that can be obtained which is the case of CPNtools (www.daimi.au.dk/~cpntools/) or Petrisimm (http://seth.asc.tuwien.ac.at/petrisimm/).

Keywords: Discrete event systems, scheduling, decision support systems, timed coloured Petri nets, Simulation, Simio.

SIMIO simulation software has the processes paradigm that allow to code the TCPN firing rules using simple steps thus the casual relationships between the generated events can be governed by the TCPN semantic rules. The advantage of integrating TCPN models to govern some activities of the SIMIO model is that it is possible to develop models using the formalism and at the same time take advantage of the characteristics and the graphical potential that SIMIO possesses. The article is organized as follows. Section 2 presents the TCPN modelling formalism, section 3 discusses some characteristics of SIMIO; section 4 describes a way to code TCPN models using some elements of SIMIO for governing some events of the modelled system. Section 5 discusses briefly the graphical and analytical capabilities of SIMIO and section 6 gives the conclusions of the article.

1. INTRODUCTION Simulation is a very well recognized methodology which possesses a high descriptive level. Most commercial simulators have several modules to analyze data, implement the relationship between elements and to perform simulation experiments. In recent years discrete-event system simulation developers have put focus on improving the characteristics of the simulation software in order to reduce the efforts needed to develop a simulation project. Some simulators such as PROMODEL(www.promodel.com) or Witness (www.lanner.com) have developed graphical modules in order to improve the graphical representation of systems, but unfortunately its original source code remains the same. SIMIO simulation software (www.simio.com) was developed by the creators of the very well known ARENA (www.arenasimulation.com) software. They have developed a novel simulation program that improves several aspects of the simulation software; they combine the processes approach with the object oriented paradigm (OOP), 2D-3D visualisation among other features that makes it a very powerful tool. With the use of the OOP the developing phase of the simulation project is improved taking advantage of the characteristics of this approach (encapsulation, inheritance, polymorphism). The developing and analysis characteristics present in SIMIO can be further

2.

TIMED COLOURED PETRI NETS Coloured Petri Nets (CPN) is a simple yet powerful modelling formalism, which allows the modelling of complex systems which present an asynchronous, parallel or concurrent behaviour and can be considered discrete event dynamic systems (Jensen & Kristensen 2009). The formalism allows developing models without ambiguity and in a formal way. It is possible to model not only the dynamic behaviour of systems but also the information flow, which is an important

734

characteristic and very useful in industrial systems modelling and decision making.

EXPR denotes the mathematical expressions associated to the elements of the formalism (variables, colours, logic conditions) where the syntax can vary when coding the formalism in a programming language. The TYPE[e] denotes the type of an expression e  EXPR, i.e. the type of values obtained when evaluating e. The set of free variables in an expression e is denoted VAR[e] and the type of a variable v is denoted TYPE[v].

In order to investigate the KPI’s (Key Performance Indicators) at which the industrial systems operate under different policies, such as scheduling, resource occupancy, costs and inventory among others it is convenient to extend CPN with a time concept. This extension is made by introducing a global clock for the model, time stamps for the entities and a time delay for the model transitions; the nets that use this extension are known as timed coloured Petri nets(TCPN). When using TCPN the global clock represents the model time, and the time stamps describe the earliest model time at which the entities of the model, graphically represented by dots (tokens), can be used for the transition evaluation process (Jensen 1997). A token is ready if the correspondent time stamp is less than or equal to the current model time. If the token is not ready, it can not be used in the transition enabling procedure.

The formalism can be graphically represented by circles which represent the place nodes and rectangles or solid lines that represent the transition nodes. The place nodes are used to model resource availability or logic conditions that need to be satisfied. The transition nodes can be associated to activities of the real system. The developed models can be analyzed with the help of available academic software such as CPNtools or PetriSimm. These types of software programs are commonly used by the scientific community in order to carry out analysis of the model to verify behaviour or systems' performance. This analysis can be performed through simulation of the system making use of the token game or performing state space analysis (Christensen et al. 2001, Mujica & Piera 2011, Mujica et al. 2010).

The transitions of the TCPN are used to model the activities of the real system and a time delay is attached in order to simulate time consumption of a certain activity. It is a common convention to use the sign @ to denote time in the elements of the model. When it is attached to transitions, it specifies the time consumption.

3. SIMIO SIMULATION SOFTWARE The most common discrete event system (DES) simulators have been coded using the activity scanning or the event scheduling approach (Shannon 1997). These approaches have been developed to reproduce the behaviour of systems whose states change in discrete instants of time. Most of the commercial tools are very efficient in modelling DES using one of these approaches or even combinations of them. With the development of fast CPU processors, powerful graphic cards, efficient statistical analysis tools etc., simulation has become more important than before. Some years ago most of the effort during a simulation project was put on the development phase of the model but with the development of more efficient software programs this effort has been considerably reduced. The available tools allow analysts to spend more time in the analysis phase of the simulation project. In order to reduce the lead time of the simulation project some developers have invested time in implementing new paradigms in their simulation products; this is the case of SIMIO. This software has been coded merging the OOP together with the processes paradigm in order to reduce the number of blocks needed to develop complex models. Based on these programming paradigms the developers coded SIMIO as a collection of objects that are instantiated from classes. These classes were designed using the principles of abstraction, encapsulation, polymorphism, inheritance and composition (Pegden 2007). Making use of the few available objects it is only necessary to add new functionalities (processes) to the original ones in order to have additional behaviour or logic (inheritance) or even overriding the original one.

The formal definition of TCPN is the following one. Definition 1. The non-hierarchical TCPN is the tuple: TCPN = (P, T, A,™, V, C, G, E,D, I) where 1. P is a finite set of places. 2. T is a finite set of transitions T such that P ˆ T = ‡ 3. A Ž P u T ‰ T u P is a set of directed arcs 4. ™ is a finite set of non-empty colour sets. 5. V is a finite set of typed variables such that Type [ X ]  ™ for all variables X  V. 6. C: P o ™ is a colour set function assigning a colour set to each place. 7. G: T o EXPR is a guard function assigning a guard to each transition T such that Type [G(T)] = Boolean. 8. E: A o EXPR is an arc expression function assigning an arc expression to each arc a, such that: Type [E(a)] = C(p) Where p is the place connected to the arc a 9. D: T o EXPR is a transition expression which assigns a delay to each transition. 9. I is an initialization function assigning an initial timed marking to each place p such that: Type [I(p)] = C(p)

735

x Steps:

Since the objects in SIMIO follow the encapsulation principle their implementation is sealed from the outside world. The composition principle allows building new classes combining the existing ones. These characteristics allow great flexibility when developing a model. One important aspect of the simulation project is the implementation phase which depends strongly on how the results are presented to decision makers.

TRANSFER NODE

x Decide x Search x Destroy x Transfer x Assign Elements: x

Station

Other objects can be also added after the development of the model in order to improve the visual aspect of the model or to make the simulation more detailed. The model of Figure 3 will be used to illustrate the implementation of the TCPN rules into SIMIO. 2’(1,1,1)+2’(2,3,1,)+2’(4,3,1)

Figure 1. Example of a SIMIO model in 2D

P1

1’(Ta,X,Op)

In this sense, the graphical interface of SIMIO has been developed in such a way that it is very easy to have very good-looking results in short time. It can switch between 2D and 3D depending on which kind of task is being performed (development or validation). Figure 1 presents a typical 2D view of the SIMIO model and Figure 2 the 3D view of the same model.

e1

e1: IF Ta=1&Op=1 THEN 1’(1,3,2) ELSE [X=Y]

T1

IF Ta=2 & Op=1 THEN 1’(2,1,2) ELSE IF Ta=4 & Op=1 THEN 1’(4,1,2)

1’(Y)

1’(Y)

P2

1’(1)+1’(2)+1’(3)

Figure 3. TCPN model 4.1. Modelling the Place nodes The place nodes are modelled using the Station element which has been developed with the objective of storing entities. The stations are Elements that need to be defined in the Definitions area of the software. Figure 4 illustrates the station definition.

Figure 2. SIMIO model in 3D view. The switch between both views is as easy as clicking the mouse. It is important to mention that SIMIO comes with a basic graphical library but it can be extended with graphical models from the GOOGLE 3D warehouse (http://sketchup.google.com/3dwarehouse/?hl=en). Figure 4. Station definition

4.

INTEGRATING THE TCPN MODELS IN THE SIMIO ENVIRONMENT Due to the dual developing paradigm used in SIMIO (process/object) it is possible to extend the functionality of the SIMIO objects with a sequence of steps which fit the purpose of coding the TCPN rules. These rules can be easily implemented using some elements of SIMIO such as the following ones.

In this figure Place1 and Place2 represent the place nodes of Figure 3. Graphically a Station is not related to any predefined object of the Standard Library in SIMIO, but its graphical visualization can be performed associating a detached QUEUE to the Station element. The entities of SIMIO can have any 3D appearance which allows exploiting the graphical potential of SIMIO. Figure 5 gives an example of a detached queue associated to the place (Place1) in the SIMIO model.

Objects: x x x

SOURCE SINK SEPARATOR

736

Decide, Search, Assign, and Destroy steps such as the one depicted in Figure 6:

Figure 6: Evaluating the restrictions

Figure 5. A detached queue for the Place1

x Decide1: the Decide step works like an IF..THEN..ELSE instruction in any programming language. The Decide Type attribute (Figure 7) must be put in condition based in order to state the expressions that need to be satisfied. The attribute Expression is used to specify the logic conditions that must be fulfilled by the entities of the TCPN model. The expression for the example model can be written in the following way.

The QUEUE is just added into the facility area and the association is performed in the Properties window: Queue State| Place1.Contents. Depending on the type of model the resource availability can also be modelled using another predefined object which can be used as a resource such as the SERVER or WORKSTATION; moreover with the use of a predefined object the graphical representation can be enhanced by taking advantage of the object’s behaviour. In the example presented here the two place nodes will be modelled making use of two Station elements.

Place1.Contents>0&Place2.contents>0

4.2. Defining the Token Colours The entities that represent the tokens are generated using the SOURCE object and they are destroyed using the SINK object or the Destroy step in the Processes area of SIMIO. The entities in SIMIO can also be extended with attributes. These attributes work as the fields of a record in any simulation language and they are associated to the entities in the States Definitions area. The states in SIMIO have been conceived as variables whose values change during the simulation run. They can be modified to include as many attributes as colours present in the token. The states can be of different types: Real, Integer, Boolean, Date, or Strings.

Figure 7: Decide condition The Decide1 is used to verify that both stations (place nodes) have at least the number of entities (tokens) imposed by the arc weights. A similar expression can be used if one place is modelled using a SERVER or any other object of SIMIO (Resource is the object’s name).

4.3. Modelling the TCPN Restrictions The dynamics of the TCPN models are governed by the input/output flux of tokens that takes place when a transition is fired. The Processes area of SIMIO is used to code the TCPN logic which evaluates the constraints to unchain the processes modelled by the transitions.

PLACE1.Contents>0&Resource.ResourceState==0 x Search1: The Search is used to perform searches within the list of elements of the stations or queues under particular restrictions. If the arc expressions have known information such as constant values, this condition is stated using the match condition attribute of the Search step. When the Search step has found one entity that satisfies the restrictions imposed by the arc, this entity goes out from the found side of the step and continues through the rest of the flowchart to check the conditions of the remaining place node. The latter is performed in the Search1 of the flowchart of Figure 6. If the Search1 founds an entity that satisfies the restriction, it leaves the Search step through the found side of the step. If it does not found any entity that fulfills the

In order to enable a transition the following conditions must be fulfilled: x

the number of tokens in the input place nodes are greater than or equal to the arc weight x the colours of the correspondent tokens must have the particular value that is stated in the arc inscription x The colour binding must satisfy the boolean expression stated by the Guards The evaluation of the restrictions imposed by the arcs and guards can be performed using a combination of

737

restrictions then it leaves the step and goes to the Destroy1step. If it finds an entity that satisfies the restriction, it flows through the Found side of the search step and continues to the Assign4. x Assign4: is used to bind the value of the entity to the variable which will be evaluated by the Guard (X variable) x Search2: This step is used to perform searches in the Station2 (Place2) under the restrictions imposed by the value of the variable (X=Y). If the search obtains one entity that satisfies the restrictions it flows through the found side of the step and continues with the animation activities. If no entity is found then the original one continues to the Decide2 step. x Decide2: This step checks whether the last search step found or not an entity that satisfied the restrictions, if not it sends the entity to perform the cycle again testing another combination of entities from Station1 and Station2.

4.5. Time Consumption The time consumption is modelled in a straightforward way using the time attributes available in every SIMIO object (these attributes can be found in the properties window). In most of the objects it is possible to associate time to the activities performed by the objects (entering the object, exiting the object, seize, delay etc). In the case of a predefined object such as the SERVER the time consumption is modelled by the delay or any other time-consuming activities available in the object (Figure 9). The advantage of using these properties is that it is possible to model activities which consume time in a deterministic or stochastic way.

4.4. Sending the entities to the facility In order to animate the capturing of resources, the entities can be sent to the facility window where the animation can be performed depending of the resource that is used (colours of tokens from place P2). Figure 8 illustrates the flowchart section where the successful entity is sent via transfer step to the correspondent node within the facility window based upon the attribute value of the correspondent token. In this example the attribute value can have three different values depending on the resource used (1,2,3) therefore three possible outcomes are included in the flowchart (two different Decide steps).

Figure 9: Defining the time consumed by an activity 4.6. Attaching a Transition to the SIMIO model Finally, in order to govern the flow of entities in the SIMIO model, it is necessary to attach the logic (coded in the process section of SIMIO) of the TCPN to an event of the SIMIO model. All the objects in SIMIO have the Add-On Process Triggers which enumerate all the possible events associated to the object. These triggers are used to call the user-defined processes when a particular event occurs. The processes (transition evaluations) can be called at almost any point of the SIMIO model within the facility area.

Figure 10. Adding the transition Figure 10 illustrates the Properties window associated to an object. In this case Transition1 is called every time the entity enters the object in run-time mode. When this event happens, the entity behaviour is governed by the logic defined by the flow diagram of the process used to model Transition1.

Figure 8. Sending the entities to the correspondent nodes The Decide steps are used only to evaluate the colour attribute that specifies the kind of resource being used and based on the result of the evaluation the entities are sent to particular locations within the Facility window.

4.7. Modelling the output arc The last element to be modelled in the example of Figure 3 is the output arc whose output function assigns the values of the colours for the output tokens.

The continuous evaluation of Transition T1 is performed in the facility window making use of a SEPARATOR object after each Server has been released. It makes a copy of the entity that comes from the Server and it is sent to the node where the transition T1 is evaluated.

738

The model can be constructed starting with the available objects in SIMIO and afterwards download the figures that will represent the actual objects in the system. Figure 12 shows a view of a manufacture model that has been graphically enhanced using the Google 3D Warehouse models. 6. CONCLUSIONS A way of implementing TCPN models making use of the analysis and graphical potential of SIMIO software has been presented. This approach can be used for developing simulation projects taking advantage of the characteristics of the TCPN formalism and the graphical and analytical tools available in SIMIO. In addition the GUI allows the user a better understanding of the real system and it allows giving a better aesthetic appearance if the simulation model is used as a tool for decision making.

Figure 11. Coding the output function This function is coded in SIMIO in a simple way (Figure 11) making use of nested combinations of Decide-Assign steps in order to evaluate the values of the attributes of the entities and afterwards based upon those evaluations the new ones are updated with the Assign steps before sending the entities (via a Transfer step) to the facility window. 5. ANALYSING THE SYSTEM Once the TCPN rules have been defined within the Process area and attached to an event of the SIMIO model the simulation can be performed in the typical way.

REFERENCES Christensen, S; Jensen, K; Mailund, T; Kristensen, L.M., 2001."State Space Methods for Timed Coloured Petri Nets", in Proc. of 2nd International Colloquium on Petri Net Technologies for Modelling Communication Based Systems, pp. 3342, Berlin, 2001. Jensen K; Kristensen L.M.; 2009.”Coloured Petri Nets: Modelling and Validation of Concurrent Systems”, Springer,2009. Jensen K.; 1997.”Coloured Petri Nets: Basic Concepts, Analysis Methods and Practical Use”, Vol.1 Springer-Verlag. Berlin, 1997. Kelton, W.D.; Smith, J.S.; Sturrock, D.T.; Verbraeck, A.; 2010.“Simio & Simulation: Modeling, Analysis, Applications”, McGraw-Hill, Boston, 2010. Moore, K.E.; Gupta, S.M.;1996." Petri Net Models of Flexible and Automated Manufacturing Systems: A Survey", International Journal of Production Research, Vol. 34(11), pp. 3001-3035, 1996. Mujica, M.A.; Piera M.A.; 2011. "A Compact Timed State Approach for the Analysis of Manufacturing Systems: Key Algorithmic Improvements", International Journal of Computer Integrated Manufacturing, Vol.24 (2), February 2011. Mujica, M.A.; Piera, M.A.; Narciso M.; 2010. “Revisiting state space exploration of timed coloured petri net models to optimize manufacturing system's performance", Simulation Modelling Practice and Theory, vol.8(9), pp. 1225-1241, Oct. 2010. Mušiþ G., 2009.“Petri Net Based Scheduling Approach Combining Dispatching Rules and Local Search”, in Proceedings of the I3M2009 Multiconference, Tenerife, Spain, 2009. Pegden, D., 2007."SIMIO: A new simulation system based on intelligent objects", in Proc. of the 39th winter simulation conference, 2007. Shannon, R. E. , 1997.“Systems Simulation, the Art and Science”, Englewood Cliffs, N. J., Prentice Hall, 1975.

5.1. Experimenting with the TCPN/SIMIO model As it has been mentioned, one advantage of integrating CPN models in SIMIO is that it is possible to use the analytical tools that are integrated within SIMIO. Using those tools it is possible to obtain KPI's that allow the decision makers to evaluate the best engineering decisions. This analysis is performed making use of the Experiment tool (Kelton et al. 2010) which allows performing experiments (replications) of the model in order to gather information from the model. After performing the experiments a report or a matrix called Pivot Grid are generated and they can be used to analyze and filter the information obtained from the experiment. Performance indicators such as resource utilization, average number of entities in the stations, average processing time, etc. are the kind of information that can be obtained from the analysis performed. 5.2. Improving the graphical view of the model One great attribute of SIMIO is the graphical interface. Since it can switch easily from 2D to 3D and the graphical models can be downloaded directly from GOOGLE 3D Warehouse the resulting models can be graphically improved without effort.

Figure 12. Final view of a model

739

Developing a Simulation Training Tool from a Medical Protocol Catherine M. Banks, Ph.D., John A. Sokolowski, Ph.D., Virginia Modeling, Analysis and Simulation Center Old Dominion University, 1030 University Blvd, Suffolk, VA 23435, USA [email protected] [email protected]

do, but who are learning something new. That retraining capability is core to this project, Developing a Simulation Training Tool from a Medical Protocol. This paper describes the methodology, modeling paradigms, and programming development to create a web-based simulation tool to train anesthesiologists and surgeons on a patient blood management medical practice. Part 2, Why This Tool, Why This Protocol, Who to Train discusses why the tool is needed, who will benefit from this medical practice, and who is the targeted trainee audience. Part 3, Who to Train, What to Train describes the trainee and the philosophy and the practice of patient blood management techniques. Part 4, How to Build It speaks to the methodology, modeling paradigms, and computer programming steps taken to develop the web-based tool. Part 5, Future Work, provides concluding comments and further tool development opportunities. It explains how the methodology, design, and tool itself adheres to the unequivocal fundamentals of modeling and simulation (M&S): verification and validation.

ABSTRACT This paper discusses the multidisciplinary effort for the development of a web-based simulation training tool that incorporates a medical protocol of patient blood management for a surgical procedure. The significance of the type of simulation tool development lies in the fact that medical simulation is able to execute training in a multiplicity of modes, it can house large digital libraries for a breadth of experiences, and it can accommodate a repetition of exercises to reinforce learning. This simulation training tool is built upon engineering and mathematical modeling. The tool is populated with simulations are drawn from actual patient case studies. The targeted trainees (users) are skilled anesthesiologists and surgeons in need of an initial introduction to this medical protocol via an expedient means to train. The tool is developed for web-based access with continuous simulation capability and hands-on exercises.

1.

INTRODUCTION

2.

Simulation training facilitates a permission to fail environment whereby the practitioner is taught in a multiplicity of modes [1]. It can house a large digital library of case studies that allows for random access to training scenarios and it can accommodate a repetition of exercises. All simulation training facilitates learning from errors and this is especially important in the field of medical training as the training takes place on a patient image, not the patient himself. The breadth of training enables a medical practitioner to methodically move from novice to master. In sum, simulation training incorporates both fundamental tasks and new tasks that serve to advance exposure to various patient cases and and develop expertise. It can support retraining that is, training experts who know what to

WHY THIS TOOL, WHY THIS PROTOCOL, WHO TO TRAIN

There is a strong case endorsing the use of M&S in training among medical institutions and research centers—the sheer need for a larger body of health-care professionals who are trained in an effective and expedited manner leads that discussion.1 Additionally, anesthesiologists and surgeons expert in the field of bloodless surgery have informed the developer community that a tool of this sort does not exist in a form they prefer. Essentially, 1

As of 2008 the American College of Surgeons certification requires three student categories who are to be taught using simulation

740

operative. These three phases lay the foundation for the patient blood management practice.

we believe a foremost, unmet need in medical simulation tools exists. Moreover, the medical subfield of patient blood management as a whole is growing exponentially and simulation training instrument development is wide-open and timely. Findings in medical literature have proven that many once excepted treatments, including blood transfusions, often carry more risk than benefit. Blood transfusions carry the risk of blood-borne illnesses including HIV, Hepatitis B, Hepatitis C, Human Lymphocytotrophic Virus, Cytomegalovirus, West Nile, sepsis, and others. One study showed that transfusion of greater than 4 units of blood increased the risk of peri-operative infection by a factor of 9.28 [2]. Furthermore, the cost in dollars of transfusion is greater than once thought. One study compared the various costs of one unit of blood ranging from $522 to $1183 [3]. The hospitals cited in that study had an annual expenditure on blood and transfusion related activities limited to surgical patients ranging from $1.62 to $6.03 million per hospital. Conversely, by implementing a blood management program Englewood Hospital (Englewood, New Jersey) was able to reduce blood use by 42% over the course of 4 years. Doing so has significantly lowered first and foremost patient morbidity and mortality and it has affected hospital financial costs [4] [5]. This tool serves as a training instrument for anesthesiologists and surgeons in the field of patient blood management techniques. It was apparent that the targeted audience, anesthesiologists and surgeons, possessed an elevated proficiency requiring a sophisticated tool that would be readily accessible and user-friendly for very busy medical professionals. The tool should comprise exercises that require the anesthesiologist and surgeon to make medical decisions relating to blood management during the three phases of patient management with each phase containing appropriate decision points.

3.

3.1 The Blood Management Philosophy The blood management philosophy can be expressed in terms of three pillars executed throughout three phases of patient care [6]. Below is a brief explanation of each pillar. Pillar 1 Optimize Formation of Blood Cellular Components (haemopoiesis) This is done by producing or encouraging conditions in the body to generate healthy levels of blood cellular components. Pillar 2 Lessen Blood Loss This ranges from identifying and managing the risk of blood loss to the mechanical aspects of surgery to avoiding secondary hemorrhage. Pillar 3 Channel / Optimize Patient’s Tolerance of Anemia Important to this pillar is the realization of patient’s actual blood loss with his tolerable blood loss. Various aspects of these three pillars are employed during three designated phases of patient care: Phase 1 Pre-operative – During this phase simple measures can be taken to satisfy Pillar 1 such as detecting and treating anemia. Pillar 2 might include obtaining the autologous blood donation for hemodilution. Pillar 3 might see the implementation of a patient-specific management plan using appropriate blood conservation modalities. In short, the preoperative phase is premised on patientreadiness for the surgery. Phase 2 Intra-operative – Timing of the operation is key to addressing Pillar 1 and that requires optimizing the formation of blood cellular components. During the surgery meticulous care is taken, aka the mechanics of surgery, to satisfy the requirement of Pillar 2 in lessening blood loss. Pillar 3 focuses on optimizing the

WHAT TO TRAIN

As with research required for any modeling task, the modelers needed to have a reasonable comprehension of the blood management philosophy vis-à-vis standard medical surgical procedure during the three surgical phases: pre-, intra-, and post-

741

3) there would be an end-of-the session assessment whereby the trainee could compare his decisions / actions to the actual case study.

patient’s ability to tolerate anemia and in intra-operative phase this can be done through appropriate ventilation and oxygenation. In the intra-operative phase minimizing bleeding is core to the patient blood management philosophy.

4.

HOW TO BUILD IT

Developing a tool premised on a patient blood management protocol requires diverse modeling skills. For example, included in the model is the mapping soft or fuzzy data (human factors) such as patient subjective data and procedural decision-making on the part of the surgeon. Grounding the soft data characterization is the mathematical modeling used to accurately chart physiological changes of the case study based on a range of things like patient response to a procedure and unexpected or inadvertent bleeding. Model development began by employing a system dynamics modeling paradigm as a means of crafting a visual representation of the factors and their correlative and/or causal relationships. This enabled the developers to take a holistic approach to tool design. This modeling methodology was followed by significant mathematical modeling to ensure precise measurements of patient vital signs like blood pressure, blood volume (loss), and other physiological reactions such as heart rate, oxygen saturation (SPO2), and respiratory rate.

Phase 3 Post-operative – This phase is a critical period for the patient. Pillar 1 necessitates much care be taken to note drug interactions that can increase anemia. Pillar 2 calls for monitoring postoperative bleeding and avoiding secondary hemorrhage. Pillar 3 concludes patient blood management by determining any postoperative anemia prescriptions This phase pays close attention to blood composition and volume. 3.2 Individualized Strategies and Decisions To develop the prototype tool it was decided that 12 case studies would be sufficient for providing a range of experiences. These case studies facilitate an opportunity to execute the trainee’s individualized strategy / decisions for each procedure. These cases would be elective surgeries (as opposed to urgent or emergent) because elective surgeries offer greater opportunity for patient / physician interaction and preparation for the surgery in the pre-operative phase. Information necessary to fully represent (model) a de-identified case study would include: the pre-operative note, any blood management orders/forms, and the post-operative note. Incorporating as much patient data as available would result in a complex characterization of the procedure. With an understanding of the blood management philosophy, introduction to the blood management techniques employed, and the determined number and variety of patient case studies needed to provide a breadth of training experience, the developers made a few broad, preliminary assumptions on what the tool would entail: 1) it will be constructed as a web-based interface (for accessibility) and it will contain real patient case studies with an assessment capability (for lessons learned) 2) the simulations will unfold in real-time (for decision-making experience)

4.1 System Dynamics Modeling The developers started by mapping out a system dynamics model to provide an explanation and validation of the tool. And because the model is all about blood management, the dependent variable would be blood volume. Independent variables include drugs, fluids, surgery, ventilation or oxygenation – all factors that could affect blood volume. As with any system dynamics model the first step is to craft a causal loop diagram. Mapping the causal loops led to representing the relationships between and among variables needed into stock and flow representations. These representations establish the feedback loops which serve to indicate the dynamic system that captures how the body functions from a mathematical standpoint. Figure 1 is the initial attempt at the system dynamics model.

742

programming tasks and web interface for the prototype tool. 4.3

Programming Design It was important that the programming of the simulation precisely follow the three blood management pillars as they overlapped the three phases of the procedure. This began by way of a scripting interface. The approach would be a decision tree structure with artificial intelligence (AI) and computer science integration. The tool would employ parallel systems: one with trainee outcomes and one with actual case study outcomes and the training would take place in a time-compressed fashion with explanations and updates forcing decision points. This process began with the use of opensource software as the development environment. It followed a game development environment with drop-box choices. The initial mapping of design called for a walk-through of the training experience. As a result the tool begins with: 1) a Welcome screen that requires a case study selection to initialize the system 2) next, a review of the case study folder of all pertinent data 3) this is followed by information such as labs, x-rays, studies, physical exam, vitals and an assessment plan that suggests what the problem may be and how to proceed After the training scenario has been initialized, the blood management pillars in the Preoperative Phase are presented in the form of dropdown boxes with choices of no surgery or move directly to surgery. If no surgery is selected, the trainee must answer why and then take steps to prepare the patient for surgery. In each case study the trainee will also experience and need to act on patients with evident potential problems or patients with undetected potential problems of which he should explore. The Operative Phase display includes what is typically presented on the OR monitor: heart rate, blood pressure, pulse rate, breathing rate. These values will be pre-assumed but with abnormal bleeding needs that require corresponding change. The interface would also need a means to display the electro-cardiogram (ECG) and X-ray readings. These would also be done in a preprocessed manner according to the case study so that image is named then displayed. Blood Loss is visualized as a tank with a bar that goes down with volume loss in realtime.

Figure 1 Initial System Dynamics Model 4.2 Mathematical Modeling All modeling and simulation includes some aspect of mathematical modeling. This is especially true for those who think in terms of computer models and computer simulations. The developers employed mathematical models and computations to support specific graphs such as the heart rate, oxygen saturation (SPO2), and blood loss. In short, the application of mathematical modeling for this exercise is to accurately represent the patient’s initial response to hemorrhage as reflected in the vital signs. It would also monitor the steady state response to hemorrhage reflected in the vital signs as well as graded responses to variable amounts of blood loss coupled with proportionate responses to corrective measures to the blood loss (such as fluid infusion). This modeling resulted in the use of a variety of equations such as that representing pulse oxygen: Exp[-x^2]*.65 + Exp[-(x - 2)^2]*0.4 + .2 The system dynamics model proffered the design of the tool in that it set out what should be represented in model via a visual representation of all factors needed. The additional effort put forth respective to the mathematical modeling was taken to ensure a high level of fidelity in the learning experience. The final undertaking in the tool development project was to formulate the

743

Finally, in the Post-operative Phase the training tool reflects the same exercises as those provided in the pre-operative phase. Figure 2 below is a screen shot of the web interface.

3. Develop a training tool software that is scalable and extendable The entry barrier was kept to a minimum by designing the tool as a web-based service, with potential trainees being able to access the service via web browsers, e.g., Internet Explorer, Mozilla Firefox, Google Chrome, etc. This also precluded the installation of proprietary software for the trainee. The software is developed using a standard programming language, C#, available as part of Microsoft’s .NET environment. For an interactive user interface, JavaScript was chosen as a language for producing visual artifacts, which range from showing patient demographic information to displaying patient medical status via ECG waveforms, and capturing user input, which includes prescription choices, etc. JQuery, a free-to-use software released under MIT License, is used for conforming to cross-browser JavaScript requirements. The interactive flows of the training span the three operative phases. These flows are driven by the information specified in the case studies. As a result, depending on the case study selected the specifics of the trainee experience are updated. The tool also allows specifying a large number of case studies, each of which is captured using a data model defined as part of the software build. That data model for capturing case studies can be used for specifying new case studies or modifying existing case studies. Once case study models are registered with the tool, the tool makes them available for training purposes requiring no change either to the software or the deployed tool. This approach to drive the software behavior via case study models meets the extensibility goal and the capability to store a large number of such models meets the scalability goal.

Figure 2 Screenshot of the Web Interface The prototype tool has proven a viable mechanism for training blood management philosophy and techniques to the specified audience, trained anesthesiologists and surgeons. In actuality, the tool more literally serves to re-train with a view to better blood management practice to be used in the pre-, peri-, and post-operative phases of patient care. As a process or philosophy the developers have taken great measures to integrate and align the training experience with blood management techniques by linking this information with existent surgical decision making and practice. The medical body of literature details a knowledge gap in practice and this tool serves as a way to lessen that gap [16]. The tool design is an XML model that incorporates the comprehensive system dynamics model and the mathematical models. It mirrors the specific procedures used in the patient blood management practice by dividing the training into pre-, peri-, and post-operative phases as these phase act out the various aspects of the 3 practices. The encoded case studies drive the behavior of the tool without the need to re-implement the software. As a result the tool is designed in a scalable fashion and allows the registration of a large number of case studies. The software that enables this tool is designed with three primary goals. 1. Keep the entry barrier for the medical professional to a minimum 2. Keep development and associated maintenance costs of the software to a minimum

5.

CONCLUSION

The initial testing of the tool proved it to operate / function with great ease moving from decision point to decision point and phase to phase. The ability to progress at will or allow the simulation to provide error comments requiring corrective actions allows the physician to train at his own discretion. The assessment values at the close of the exercise allow the trainee to evaluate his own performance. It is significant to note the multidisciplinary nature of the development of this tool. Expertise was called on from engineering, computer science, and social science disciplines. Integral was the role of medical subject matter experts in to include MDs, RNs (registered nurses), CRNAs (certified registered

744

nurse anesthetists) who shared in the development of the interface, explanation and application of the blood management practice, and testing and evaluation of the tool itself. Medical simulation training is fast becoming a necessary modality in healthcare education. This tool is the means to providing a comprehensive, effective, and time-sensitive learning experience.

6.

REFERENCES [1] Kyle RR, Murray WB, eds. Clinical Simulation: Operations, Engineering, and Management. Amsterdam: Elsevier, 2008. [2] Dunne Jr, MD, Tracy JK, Gannon C, Napolitano IM, (2002), “Perioperative anemia: an independent risk factor for infection, mortality, and resource utilization in surgery.” Journal of Surgical Research Feb;102(2):237-44.

FUTURE WORK

As mentioned at the outset, simulation training for medical professionals is an acknowledged means of effective and efficient training as it provides a learning environment with depth and breadth. And although medicine is an evidence-based exercise, the blood management is relatively novel and goes contrary to standard practice. For a variety of reasons, many skilled medical practitioners are uninformed, unable, or unwilling to engage blood management techniques. The purpose of this training tool is to facilitate that learning exercise for closing the learning gap through simulation re-training of new information and new technique. The fact that blood management practices have a dramatic impact on hospital finances and patient outcome itself serves as valid reasons why this area of medical simulation needs to be developed [7]. Future work along the lines of blood management simulation training can take many directions. First, there are divergent paths in the operating room setting as the medical professionals there, the anesthesiologist, the surgeon, and the nurses, have different roles and as such there can be a mental disconnection between / among them. Developing a tool that represents the roles of each to reflect the knowledge of who carries out different tasks and how to train to those tasks as a multidisciplinary team. A second potential training scenario is a tool that focuses on the mechanical aspects of surgical procedure in the intra-operative phase to include such things as precision with surgical incision, ANH and cell-salvage implementation, patient autologous blood donation for hemodilution (how much and when).

[3] Shander A, et.al, (2010). “Activity Based Costs of Blood Transfusions” Transfusion 50 (4): 753-765. [4] Moskowitz DM, Klein JJ, Shander A, et al, (2004). “Predictors of transfusion requirements for cardiac surgical procedures at a blood conservation center.” Annals of Thoracic Surgery; 77: 626-634. [5] Thomson A, Farmer S, Hoffmann A, Isbister J, Shander A, (2009). “Patient Blood management- a new paradigm for transfusion medicine?” ISBT Science Series 4, 423-435. [6] Shander A., et.al. Perioperative Blood Management: A Physician’s Handbook. 2nd edition. Bethesda: AABB, 2009. [7] Seeber P, Shander A. Basics of Blood Management. Hoboken: Wiley-Blackwell Publishers, 2007.

AUTHOR BIOGRAPHY Catherine M. Banks PhD, is Research Associate Professor at VMASC. Her focus is on qualitative research among the social science disciplines to serve as inputs into various modeling paradigms: game theoretical, agent-based, social network, and system dynamics. Dr. Banks’ research includes models representing humans and human behavior to include the translating / mapping of data for quantitative representations, modeling states and their varied histories of revolution and insurgency, political economy and state volatility, and medical simulation. She has authored and edited books and journal articles on these topics and is contributor and coeditor of Modeling and Simulation in the Medical and Health Sciences (Wiley Publication to be released April 2011).

745

John A. Sokolowski PhD, is the Executive Director for VMASC, supervising 50 researchers and staff with an annual funded research budget of $10 million. He supervises research and development in Transportation, Homeland Security, Defense, Medical M&S, Decisions Support, Business & Supply Chain, and Social Science (real-world) M&S applications. He is contributor and co-editor of Modeling and Simulation in the Medical and Health Sciences (Wiley Publication to be released April 2011).

746

MODEL SYNTHESIS USING A MULTI-AGENT LEARNING STRATEGY Sebastian Bohlmann(a), Arne Klauke(b), Volkhard Klinger(c), Helena Szczerbicka(d) (a)(d)

Department of Simulation and Modeling, Leibniz University Hannover, 30167 Hannover, Germany (b)(c) Department of Embedded Systems, FHDW Hannover, 30173 Hannover, Germany

(a)

[email protected], (b)[email protected], (c)[email protected], (d)[email protected]

The verification strategy is based on a set of data called Input sequencesሺšͳ ሻ– ǡ ǥ ǡ ሺš ሻ– ǡ – ‫ א‬Գ, ሻ ሺ› Sequences and Output Sequences ͳ – ǡ ǥ ǡ ൫›Œ ൯ ǡ – ‫ א‬Գ, – which are related to the Input Sequences by a functional relationship ˆǣ Թ ื ԹŒ (formula 0), illustrated in figure 2.

ABSTRACT In this paper we give an overview of our multi-agent based model identification framework. We are identifying functional relationships in process data. We do this by using multi-agent based heuristic algorithms. Moreover we give a proof of concept concerning the abilities and performance of our system. Keywords: model synthesis, agent-based evolutionary computation 1. INTRODUCTION Manufacturing systems are one of the largest application areas for modelling and simulation. In particular the pulp and paper industry is one instance of a large-scale production processes (Bohlmann and Klinger, 2007). We have brought up a framework for modelling and simulation of those process environments in (Bohlmann, Klinger and Szczerbicka, 2009), (Bohlmann, Klinger and Szczerbicka, 2010b) and (Bohlmann, Klinger and Szczerbicka, 2010c). This paper focuses on the identification procedure. We consider time series extracted from process data. These time series are subdivided in input and output series. The problem treated here, is to find a functional relationship between the input and output series. At the beginning it is unknown which of the input series are actually used in that relationship. Our approach to solve this problem is a multi-agent based learning strategy (Bohlmann, Klinger and Szczerbicka, 2010b). In figure 1 the system identification overview is shown. It consists of two basic steps, the preprocessing and the multi-agent based optimization. The process data input (PData) is used to generate an appropriate process model (Law and Kelton, 2000). To verify this identification procedure we have to evaluate the different steps very carefully not only to its technically correct function but on its performance behaviour. ݂ͳ ሺሺ‫ ͳݔ‬ሻ‫ ݐ‬ǡ ǥǡ ሺ‫ ݉ݔ‬ሻ‫ ݐ‬ሻ ݂݆ ሺሺ‫ ͳݔ‬ሻ‫ ݐ‬ǡ ǥǡ ሺ‫ ݉ݔ‬ሻ‫ ݐ‬ሻ

ൌ ‫ڭ‬ ൌ

Figure 1: Function block view

Figure 2: Input/ Output sequence In figure 3 an example for m=10 and j=1 is shown. The problem we are solving is to identify this function f, only knowing values of ሺšͳ ሻ– ǡ ǥ ǡ ሺš ሻ– (thin lines) and ሺ›ሻ– (thick line) for a limited set  ‫ ؿ‬Գ of time indices, which may differ for each sequence. In this paper we are treating only problems with j=1. Our approach for this challenge is formed by the identification framework used for process model identification and it uses the data management framework presented in (Bohlmann, Klinger and Szczerbicka, 2010c).

ሺ‫ ͳݕ‬ሻ‫ ݐ‬ǡ ‫ א ݐ‬Գ ൫‫ ݆ݕ‬൯‫ ݐ‬ǡ ‫ א ݐ‬Գ

(0)

747

To achieve this, at first the data from each sensor is linearly interpolated and then smoothed by a convolution. After that the new data sequences are equalized with the original values from the sensors. Finally equidistant values from each sensor are picked and combined to data samples, which describe the whole system, as desired. 2.3 Prefiltering the Data The environment built for the evolutionary algorithm, explained in the next section, needs a predetermined number of data samples. In general the number of data samples delivered by the resampling unit is too large. Moreover it may contain redundant data samples, containing no information. This may happen if the state of the real system does not change for a time period. In this unit the samples, which contain the most information, i.e. these with the highest entropy, are chosen. This is implemented in different prefilter modules.

Figure 3: Input and output data series The pre-processing is followed by a Multi-Agent-based Learning Strategy (section 3.1.2) using memetic evolutionary algorithm. In the next sections the identification framework is explained in detail. It consists essentially of two parts (see Figure 1): A data pre-processing unit and our evolutionary algorithm. Finally some examples, proving the functioning of the framework, are discussed.

1. 2.

2. DATA PREPROCESSING In this section the preprocessing of the raw data is explained in detail. figure 4 presents the basic function blocks.

3.

4.

Figure 4: Data Proprocessing 2.1 Data Factory The data used for the identification process is either extracted from various sensors in a real system or they may be synthetic. In the first case the data is usually collected over a long time period and saved in a data archive. This raw data then runs through different preprocessor units, explained below, to build usable data series for the identification framework. In the second case the raw data is produced synthetic using random data streams and is then passed to the preprocessing units.

No Prefilter: The simplest way is to choose just the last 729 samples. Random Prefilter: The samples to be passed onto the planets are chosen randomly. Weighted Random Prefilter: The samples to pass on are chosen randomly, but with different probabilities. This probability corresponds to the angle between the input values of the current sample and its predecessor and successor. k-means Prefilter: In this method we are using a cluster algorithm to choose the samples to be passed on (Kanungo, Mount, Netanyahu, Piatko, Silverman and Wu, 2002). The data set delivered by the resampling unit is subdivided in blocks of a fixed size. In each of these blocks we build a fixed number of clusters and only the centres of these clusters are passed to the planets. We have decided to use k-means clustering because one can choose the number of centres from start. Furthermore the clusters generated by this algorithm are formed spherical, what seems to be the most suitable form for our purpose.

2.4 Presorting the Data After the samples are selected they need to be arranged on the planets in a useful way. The simplest method is to keep them in their current order. But there are concepts, which can improve the system behaviour. One approach to order the samples in a more useful way is the so called TSP Filling (Travelling Salesman Problem). The samples are arranged in a way that approximately minimizes the sum of the distances of neighbouring samples, with respect to a chosen metric. This concept can be generalized by not just taking the direct neighbours into account, but the next n neighbours in both directions for each sample.

2.2 Resampling the Data In this initial unit the weaknesses of the data-recording by the sensors are remedied. The sensors distributed over the system are usually not working periodic or even synchronously. Moreover they might sometimes measure values, which are obvious wrong. These error values are simply removed from the data set. The actual task in this module is to produce a time series of equidistant data samples, each of which describes the state of the whole system at a moment.

748

3. MULTI-AGENT-BASED OPTIMIZATION The algorithm uses an evolutionary approach to find the functional relationship in the process data. We have created an environment which offers aliments to the creatures living in it. These creatures own a genotype, which they are trying to adapt to their location in the environment by building children with a changed genotype. A creature which is well fitted to its location has a better ability to absorb aliments from it. Aliments are used to perform various evolutionary operations to build a child. The genotype passed to the child can be mutated, crossed with the genotype of another creature, and enhanced in several ways. The creatures can move within their environment and interact with other creatures. The environment is representing the data set, the local view is just a subset of it. The creatures are software agents and their genotype is a model function, approximating the functional relationship. We can rate how well an agent is fitted to its location by calculating the error of his model function using the local data and a search metric. In the following section we will give an overview of the system architecture, depicted in figure 5.

Figure 5: Planet Surface 3.1.1 Agent Environment The environment of the agents consists of areas. Each area contains one data sample and can hold one agent. Moreover areas can be linked to other areas, called neighbours. These neighbours serve for two purposes. At first the agent held by one area can be moved onto one of the areas neighbouring areas. Secondly the links between the areas are used to build the set of data samples for the local learning. The areas are aggregated in planets. The surface of a planet is a torus, represented by a quadratic field of areas. This surface is build in a recursive pattern of squares containing nine elements, filled meander like. This method leads to the planet size ͻ͵ ൌ ͹ʹͻ and is illustrated in figure 6. Each area on a planet is linked to its four neighbours to the left, right, top and bottom. Finally the planets build the universe, which is controlled by the universe supervisor. All planets are controlled by so called planet supervisors. Some of the areas on each planet are marked as beam areas. When a certain number of iterations have passed, the planet supervisor sends copies of all agents, which are placed on a beam area to a randomly chosen area on a randomly chosen planet.

3.1 System architecture overview The architecture is organized in four stages, the data processing and local and global agent environment. Each is arranged in functional levels, for data management, agent behaviour control, supervision, execution and synchronization, regarding the overall management. The basic level is called MPU (Multi Processor Unit) and represents the system thread representation. The optimization starts with the initialization of the preprocessed data, managed in the data processing stage. The splitter in the supervision level supplies the evaluation units in the other stages with data samples. To illustrate the agent based algorithm, we give a detailed description of the environment, the individuals live in, of the individuals themselves and of their behaviour.

749

Figure 6: System Architecture

750

are done we decide whether the agent may live for one more iteration or not. If the energy level is not positive the agent is removed. If the agent had a child, it is placed onto the area. If the energy level is positive the agent tries to move to a randomly chosen neighbouring area. When the chosen area is empty, the agent just moves. If the area is already occupied by another agent, the procedure depends on the agent’s energy level. If it is not high enough to perform a cross operation, the agent does not move. Else a cross operation is performed. If the cross operation yields a new agent it is saved as the child of the original agent and the agents do not move. After the move operation, the individual tries to build a child, if none is present. Depending on the agent’s energy level, the agent performs a Replication, Mutate or Enhance operation to build his child. In the next step the agent may perform a Learn operation, if he is not adult (without loss of energy), or if his energy is high enough it is decided randomly if the agent is allowed to learn or not. Now the individual performs a short learn operation, if its age is appropriate. Finally, if the agent has moved in this iteration and owns a child, he tries to place it on his last area.

3.1.2 Software Agents The software agents are operating in the environment described before. Each agent owns a model function and is placed onto an area. The model function is stored in a tree representation and is build using elementary operations like sin, +, *, / the variables x1,..,xm and a set of parameters (Schmidt and Lipson, 2007). This concept is shown in figure 7. Moreover agents may build a child, to pass on their information. We have implemented in the following evolutionary operations: Replication: The individual produces a copy of himself. This is the most expensive operation. Cross: Two individuals interchange randomly chosen parts of their model functions. Mutate: A part of the model function is replaced by a randomly build function. Enhance: The structure of the model function is simplified, if possible. Short Learn: The parameters of the model function are fitted to the local learn data using a simple, but fast algorithm. Learn: A more sophisticated algorithm is used to calibrate the parameters of the model function. This operation is implemented in a memetic coprocessor. All of these operations have an energy effort, which is subtracted from the agents energy, if the operation is performed. Furthermore the agents can measure the error of their model function using different kind of search metrics.

3.2 Optimization stages In the last section the agents have been introduced. Here we map this agent based algorithm on the system architecture. According to the planet setup, there is the global stage (universe) and the local stage (one planet). The fourth stage is formed by memetic coprocessors, assigned to the planets. These stages form separate execution loops, which run in parallel. The listing below gives an overview of these three loops. 1.

Figure 7: The tree representation of the model function 3.1.3 Agent Behaviour Each iteration starts with an update of the agents properties. The age of the individual is increased. After that the energy level is recalculated. This happens as follows: The individual calculates the error of its model function using one of the search metrics described in section 3.4. If this error is too big the energy level is not increased. Else, the individual is allowed to absorb the energy offered by the area it is placed on and a value depending on the complexity of its model function is subtracted from its energy level. When these operations

2.

751

Global Stage: This loop is used to manage the elite population. Agents nominated in the local loop are passed via the MPI (Message Passing Interface, synchronization level) to the EVA Executor (execution level) and then to the Judge (supervision level), which decides, if the agent is added to the elite population (agent level). The agents in the elite population are evaluated on the full data set (data level). If one of them fulfils the termination criterion the algorithm stops. The universe processor (MPU level) returns the ranking of the best individuals if the algorithm terminates. Local Stage: In this stage the agents are generated and put onto the planets. Once they are placed on an area, the agents start their life cycle, described in subsection 3.1.3. In the data level the agents evaluate their model function, using a search metric and a subset of the challenge data set, called local learn data. In the MPI (synchronization level) agents are exchanged between the planets.

3.

mean value or the maximum over all samples. They might perform these steps using not all, but a randomly chosen subset of their local learn data, in these cases the metric is called partial.

Memetic Stage: In this loop agents, which were passed from the Operators (agent level, local stage) to the memetic unit (execution level), are optimized by more expensive algorithms, like downhill simplex (DHS) (Nelder and Mead, 1965). To evaluate the error of the model function, a small test data set is used. The agents are then passed back to the EVA Executor in the execution level of the local loop.

4. PROOF OF CONCEPT To proof the basic usability of our framework we executed 8 different experiments with different problem complexities to be solved by the agent system. Each experiment is repeated 30 times to reduce statistical deviation. For each run we take the runtime beginning with the first agent generation and ending at the first successful detection. A successful detection is defined by a mean absolute error of 0.001 over the whole dataset. Each run has a maximal runtime of 1800 seconds. If no valid solution is found in time, the execution is aborted and counted as an unsuccessful execution. To proof the parallelisation concept we repeated 8 times 30 experiments with 1 execution core (one Planet) and 8 cores.

3.3 The Elite Population There are two opportunities for an individual to get nominated for the elite population. The first is when an individual is removed from the planet and has lived for at least 200 iterations. Furthermore an individual is nominated for the elite population after every 250 iterations. The universe supervisor chooses the best 25 individuals from all nominated agents. The elite population is used to check the termination criterion: The model functions in the elite population are evaluated not just using the local learning set, but the data from the whole universe. If one of them has an error under a predefined border the algorithm terminates and returns this model function.

4.1 Experimental Setup This section specifies the used configurations and dependencies. All experiments are executed on a Dell PowerEdge R815 with in total 4 AMD Opteron 6174 processors (each providing 12 cores with respectively 128KByte L1-cache, 512 KByte L2-cache and common 12 MByte L3-cache) and an overall RAM configuration of 128 GByte. For the parallelization evaluation this platform provides a scalable hardware environment. The challenge to be solved by the agent system is generated synthetically. Because randomness in the generation has huge impact on the detection system the same dataset is used in all 30 runs per experiment. We generated a five dimensional time series input stream

3.4 Search Metrics The agents are able to use different search metrics to estimate the error of their model functions. For each data sample in the local learn set, which is build using the neighbouring areas, the agent calculates the difference between the original output and the output calculated by the model function. For this step the agent can use the euclidean or the absolute distance. In the next step these values are aggregated by building the 1000

f7 f8

f6 detection time in s

100

f4

f5 10

f3 f2 f1 1 1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

iteration ordered by detection time

Figure 8: Measured ordered runtime

752

20

21

22

23

24

25

26

27

28

29

30

with ͳͲͷ samples each. Further a result stream with the same length is generated using one of the 8 functions in formula (1) to (8). To measure the quality of the result a metric calculating the mean absolute error is used. ˆͳ ሺšͳ ǡ ǥ ǡ šͷ ሻ ൌ šͳ ȉ šʹ

(1)

Ɏ ˆʹ ሺšͳ ǡ ǥ ǡ šͷ ሻ ൌ •‹ ቀšͳ ൅ ቁ Ͷ ˆ͵ ሺšͳ ǡ ǥ ǡ šͷ ሻ ൌ Ͷʹ ȉ šͳ ȉ šͳ ൅ ʹʹ ȉ šʹ

(2)

ˆͶ ሺšͳ ǡ ǥ ǡ šͷ ሻ ൌ •‹ሺšͳ ȉ šʹ ൅ ͵ሻ ൅ Ͷ

(4)

ˆͷ ሺšͳ ǡ ǥ ǡ šͷ ሻ ൌ š͵ ൅ šͶ ൅ ͶǤ͵ ȉ •‹ሺšͳ ൅ ͹ሻ ൅ ͳǤ͸ͷ

(5)

ˆ͸ ሺšͳ ǡ ǥ ǡ šͷ ሻ ൌ šͳ ȉ šͳ ൅ šʹ ȉ šʹ ൅ š͵ ȉ š͵ ൅ ʹʹ

(6)

ˆ͹ ሺšͳ ǡ ǥ ǡ šͷ ሻ ൌ െ͵Ǥʹ͵ ȉ •‹ሺšͳ ሻ ȉ •‹ሺšʹ ሻ ൅ ʹǤͶ͵ Ɏ ȉ •‹ ቀšͳ ൅ ቁ Ͷ ˆͺ ሺšͳ ǡ ǥ ǡ šͷ ሻ ൌ െͳʹʹ ȉ šͳ ൅ ʹǤ͵ ȉ šʹ ൅ ͲǤʹ ȉ šͶ ൅ š͵ ȉ •‹ሺͲǤͳ ȉ šͶ ሻ

(7)

do not lead to a better system performance on the used machine because the other system components (see figure 6) consume the remaining system resources. We conclude that also the agents only use heuristics for interaction and learning the combined execution is target-oriented.

(3) f1 f2 f3 f4 f5 5.

SUMMARY Modeling and simulation of non formalized system behavior still is a grand challenge for science and engineering. As we demonstrated it is possible to implement a machine learning system to help modeling specialists to gain knowledge form the data produced by the original system. In this scenario it is required to formulate the produced recommendations in a human comprehensible form. Differential equations (and simple equations, as in this concept paper) are one possible knowledge representation. And in difference, from knowledge e.g. learned by a neuronal net, knowledge is not encapsulated. Engineers have a huge tool kit to continue processing such a result. As done for a simulation system in (Bohlmann, Klinger and Szczerbicka, 2009) such a agent based modeling support system can easily connected to real word data sources and could be helpful to enhance or generate complex models for simulation environments (Zeigler, Praehofer and Kim, 2000). As a result the complexity to model a complex process is simplified by using the analytic strength of the modeling engineer and the knowledge compression strength of an agent based machine learning environment.

(8)

4.2 Results Figure 8 illustrates the results for the parallelised agent system with 8 planets. Each line type represents the different successfully ended runs for one of the generator functions ordered by runtime. The y-scale is logarithmic. As expected the mean runtime is higher if the function is more complex. This value can also be found in table 1. Some of the data lines do not have 30 samples, because not all runs resulted in a valid solution. For function 1 to 6 the agents evolution always leads to the correct structure. Especially for the last two functions the highest runtimes increase strongly.

f1 f2 f3 f4 f5 f6 f7 f8

Table 2: Statistics for parallelization Mean x1 Mean x8 Speedup Detection Improveme nt 5,6 s 1.8 s 309 % 100 % 6,5 s 1.8 s 356 % 100 % 28,0 s 5.0 s 561 % 500 % 82,0 s 9.9 s 826 % 230 % 59,0 s 11.7 s 505 % 272 %

Table 1:Statistics for different functions Mean Std. Dev. Detection Rate 1.8 s 0.5 s 100 % 1.8 s 0.6 s 100 % 5.0 s 11.2 s 100 % 9.9 s 23.9 s 100 % 11.7 s 6.4 s 100 % 20.6 s 31.0 s 100 % 25.3 s 121.6 s 97 % 73.4 s 84.4 s 37 %

6. FURTHER WORK The further work has two key aspects of activity: Increase the parallelization to be able to use more agents. As mentioned before the memetic co-processor cores use the majority of our machine resources. All used memetic algorithms are suited for SIMD coprocessors and would scale the system to about 40 planets. The second work to be done is to reduce the number of problem specific parameters by the help of control loops. Finally the framework is written as generic as possible. There are only few dependencies e.g. the problem has to be dividable into local challenges. So we like to formulate solvers for different known problems in the area of modeling and machine learning.

This effect caused a deadlocked evolution, if the overall diversity of the agents is low. When the overall number of agents and the separation by multiple planets is increased this probability decreases. This is indicated if we compare the total number of detections and the speedup as done in table 2. Here the functions five to eight are not listed because the detection rate is too low to acquire adequate measurements for a non-parallelized system. Speedup for complex challenges (3-5) is effective. As positive detection rates increase and detection runtime decreases the positive effect is higher than the expected factor 8. At the moment more planets

753

REFERENCES Bohlmann, S. and Klinger, V. (2007) Modellbildung für kontinuierliche Produktionsprozesse in der Papierindustrie. Forschungsberichte der FHDW Hannover (ISSN 1863-7043), 08:1–20.

Schmidt, M. and Lipson, H. (2007) Comparison of tree and graph encodings as function of problem complexity. In GECCO ’07: Proceedings of the 9th annual conference on Genetic and evolutionary computation, pages 1674–1679, New York, NY, USA. ACM.

Bohlmann, S., Klinger, V., and Szczerbicka, H. (2009) HPNS - a Hybrid Process Net Simulation Environment Executing Online Dynamic Models of Industrial Manufacturing Systems. In Proceedings of the 2009 Winter Simulation Conference M. D. Rossetti, R. R. Hill, B. Johansson, A. Dunkin, and R. G. Ingalls, eds.

Zeigler, B. P., Praehofer, H., and Kim, T. G. (2000) Theory of Modeling and Simulation: Integrating Discrete Event and Continuous Complex Dynamic Systems. Academic Press, San Diego, USA, 2 edition. AUTHORS BIOGRAPHY SEBASTIAN BOHLMANN is a Ph.D. candidate at Department of Simulation and Modelling - Institute of Systems Engineering at the Leibniz Universit¨at Hannover. He received a Dipl.-Ing. (FH) degree in mechatronics engineering from FHDW university of applied sciences. His research interests are machine learning and heuristic optimization algorithms, complex dynamic systems, control system synthesis and grid computing. His email address is .

Bohlmann, S., Klinger, V., and Szczerbicka, H. (2010a) System Identification with Multi-Agentbased Evolutionary Computation Using a Local Optimization Kernel. In Submitted to ICMLA 2010 (International Conference on Machine Learning and Applications). Bohlmann, S., Klinger, V., and Szczerbicka, H. (2010b) System identification with multi-agent-based evolutionary computation using a local optimization kernel. In The Ninth International Conference on Machine Learning and Applications, pages 840–845.

ARNE KLAUKE is a researcher at the university of applied science FHDW in Hannover. He received a Dipl.-Math. from the Gottfried Wilhelm Leibniz Universit¨at Hannover. His email address is .

Bohlmann, S., Klinger, V., and Szczerbicka, H. (2010c) Co-simulation in large scale environments using the HPNS framework. In Summer Simulation Multiconference, Grand Challenges in Modeling & Simulation. The Society for Modeling and Simulation.

VOLKHARD KLINGER has been a full time professor for embedded systems and computer science at the university of applied sciences FHDW in Hannover and Celle since 2002. After his academic studies at the RWTH Aachen he received his Ph.D. in Electrical Engineering from Technische Universität HamburgHarburg. He teaches courses in computer science, embedded systems, electrical engineering and ASIC/system design. His email address is .

Kanungo, T., Mount, D. M., Netanyahu, N. S., Piatko, C. D., Silverman, R., and Wu, A. Y. (2002) An efficient k-means clustering algorithm: Analysis and implementation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24:881–892.

HELENA SZCZERBICKA is head of the Department of Simulation and Modelling - Institute of Systems Engineering at the Leibniz Universit¨at Hannover. She received her Ph.D. in Engineering and her M.S in Applied Mathematics from the Warsaw University of Technology, Poland. She teaches courses in discreteevent simulation, modeling methodology, queuing theory, stochastic Petri Nets, distributed simulation, computer organization and computer architecture. Her email address is .

Law, A. M. and Kelton, W. D. (2000) Simulation Modeling and Analysis. McGrawHill Nelder, R. and Mead, J. (1965) A simplex method for function minimization. Computer Journal, 7(4):308–313.

754

SERVICE OPTIMIZATION FOR SYSTEM-OF-SYSTEMS BASED ON POOL SCHEDULING AND INVENTORY MANAGEMENT DRIVEN BY SMART SIMULATION SOLUTIONS Agostino Bruzzone, Marina Massei, MISS DIPTEM University of Genoa Via Opera Pia 15, 16145 Genova, Italy Email {agostino, massei}@itim.unige.it - URL www.itim.unige.it Enrico Bocca, Simulation Team Via Molinero 2, 17100 Savona, Italy Email [email protected] - URL www.simulationteam.com this complex framework; in fact in the real industrial and business case provide many opportunities to apply "pool management" to system-of-systems; in fact there are several frameworks where to apply this approach and authors have completed successfully several R&D projects for major companies on this subjects such as: − Service for fleets of helicopters and resources devoted to provide Search and Rescue − Service for fleets of resources (i.e. Buses, Metro, Trams, etc.) for a mass transportation companies − Service for different fleets of tank vessels supporting different chemical industries − Industrial Plants Service Pool Management

Keywords: Simulation, Power Plants, System-ofSystems, Pool Management, Service, Maintenance, Decision Support System, Optimization. ABSTRACT The aim of this research is to support service and maintenance of pools of System-of-Systems, such as power plants or vessel/aircraft fleets by using simulations model dynamically integrated with smart optimizer driven by Artificial Intelligence techniques. The proposed methodology permit to create a framework to evaluate, optimize and test the service and maintenance policies (involving both inventory and scheduling); this framework is based on a simulator combined with an intelligent optimizer The authors proposed a new metrics to evaluate the real performance of pool service management of the whole complex system and support the optimization processes.

In fact along the years the authors have developed methodologies and simulation models to face these challenges and in particular they have developed LAPIS (Lean Advanced Pooling Intelligent optimizer and Simulator) suite, integrating M&S (Modeling & Simulation) and AI, to support decision making over this context; therefore this paper proposes an example applied to power plants service over a pool of different sites and units; the results obtained from the simulator are used for demonstrating the potential and benefits of the new methodology proposed as well as validation support. This papers focus in fact on applying LAPIS to power plant pool management for optmizing their service over a wide spectrum of target fucntions. In addition by this approach is possible to control and optimize different aspects of system-ofSystems maintenance, as power plant pools, in terms of different hypotheses in term of the balance between service quality (i.e. availability), cost estimation and constrain respect; the case study proposed represent an example of system-ofsystems simulation able to manage all the aspect above mentioned and to guarantee good results in a real industrial case.. The research activities are synthesized in the LAPIS model description and the

INTRODUCTION The work performed by the authors in this research is to verify the benefits of the integration of simulation models with Artificial Intelligence (AI) techniques in complex system service and maintenance; in fact in most of the case the complex systems rely and require very expensive and sophisticated maintenance/service support; in fact this problem become even more difficult when the complex systems due to their interaction evolve becoming a system-of-systems; in this case the use of simulation is usually the only reliable approach to face the difficulties related to the management of their service/maintenance; in particular in this context it is critical to consider some KPI (Key Performance Indexes) with special attention to availability, costs, resource utilization, readiness. From this point of view one interesting opportunity for improving these KPIs and for guaranteeing better overall performances is to develop pool management strategies able to generate synergies in

755

is still a very important factor, but the choice to use just that parameter was even due even to the fact that estimating profitability was quite complex until few years ago due to data availability and format; today, due to the high variability in demand and especially in energy prices over time, the target function to be maximized is often the profit achievable by a plant evaluated by a combined estimation of capability and prices integrated over time; the goal is to have the unit operative as much as possible during the most profitable timeslots; special algorithms for estimating this kind of target function are proposed by the authors (Bruzzone , Madeo, Tarone F 2010). The final solution should include the definition of schedule, inventory management by fixing compatible timeframes with the service time cycles of each item; this solution needs to optimize concurrently availability, profitability, costs and technical commercial constraints that are often defined based on the specific case and point of view (i.e. different from user to service provider); it is evident that such application represents a very hard problem that to be solved need to be approached by the innovative techniques combining simulation and optimization as proposed in this paper. In fact within real applications the stochastic factors (i.e. failures, repairing time, spare part delivery times, item refurbishment lead times) combined with the complex processes (i.e. refurbishment processes, commercial procedures, technical constrains) have a strong influence to the overall performance of the maintenance solutions. By the innovative approach proposed in this paper the solution is defined by applying a DSS (Decision Support System) combining stochastic simulation and intelligent optimization; therefore the positive results achieved in several previous application suggest this as appropriate approach to solve this problem (Bruzzone & Simeoni, 2002)

validation of this approach is obtained by the experimentation on a real power plant case study application (the data proposed in the analysis are modified due to confidential reason).

SERVICE & MAINTENANCE FOR A POWER PLANT POOLS A pool of power plants is a set of power plant sites where multiple units of different machines are operating; for instance today in most of the case the traditional power generation relies on combined cycle power plants: each plant have usually more than one combined cycle each one incorporating just a main machine a Gas Turbine, a Steam Turbine and two Generators; each machine is composed by several systems (i.e. main and secondary systems) as well as by several auxiliary systems (i.e. Aqua Demi, HVAC, firefighting etc); in several case the subsystems are even very complex (i.e. Digital Control System, Burning Control System etc.); in addition the maintenance for such components is driven by preventive actions related to their use (equivalent operative hours) and on failures; the first component is strongly related to the power demand and utilization modes of each plant that is strongly affected by exogenous stochastic factors, while the failures obviously are characterized by high complex statistical distribution combining different phenomena (i.e. basic failures and rare catastrophic events). It results evident that to provide efficient service such set of power plants corresponds to define a pool management strategy for a system-of-systems; the paper focuses in fact in the identification, design and engineering of best service and maintenance policies for such system-of-systems; in the proposed case it is considered a group of combined cycle power plants (4 up to a dozen); this management need to operate considering available resources (i.e. personnel), scheduling and available timeframes (i.e. technical, commercial and contractual constraints), inventory (i.e. spare part storage and replacement policies), acquisition/refurbishment policies (i.e. for item subjected to regeneration such as several layers of Gas Turbine Blades). The final aim is to define the preventive service and to support decisions able to optimize costs and power plant availability and, at the same time, to reduce the risks and to guarantee a robust management in case of unexpected breakdowns, It is interesting to note that today availability concept evolved and it is more important respect to the traditional availability a new estimator related to the plant profitability; therefore the availability it

POWER PLANT POOL MANAGEMENT Due to high order of interactions among different entities, stochastic factors, different objects and a lot of target functions each Power Plants Service represent itself a pretty complex framework. In fact, often, many machines (i.e. generators, gas turbines, steam turbines, boilers) in multiple sites need to be maintained concurrently both in term of preventive maintenance as well as in term of failure recovery; for this reason modern service strategies deal with pooling the power plants and generating synergies to compensate the high degree of complexity. In the power plant maintenance case propose there are many important elements (i.e.

756

intensity of the use Power Plants (related to the demand variations) and the mode of use (related to the policy for managing the machines by the users); that parameters are affected by stochastic factors and the results generates a complex behavior of the components; in fact currently the LAPIS simulator is able to be integrated with DCS and keep up-todate on the EOH of each system in the power plant pool. Through this example is easier to understand the complexity of the service and the emerging of complex behaviors in this system-of-systems; to keep the system under control and to guarantee good performances in this case it is critical to understand that the optimization process should be related to different target functions defined based on the specific case; in general the two main components of the these target functions for power plant service are related to service costs, power generation profitability and plant availability. These are obviously competing functions and requires a multivariable combined optimization. From other point of view the degrees of freedom for controlling and improving the power plant service management are related to the following main elements: • Power Plant Preventive Maintenance Scheduling • Power Plant Component Inventory Management • Refurbishment Component Planning & Sequencing The goals of this research was to create a power plants manager, that interact with an intelligent decision support system to estimate correctly the plant performances and to optimize resources, inventory and scheduling. The approach proposed in this paper permits different and interactive modes: automated optimization integrated with the simulation as well as what if analysis; so the different solutions and policies are simulated driven by the intelligent optimizer or by the users and estimates the KPIs; the optimizer proposed for this case is based on GAs (Genetic Algorithms)due to the high number of variable and the strong influence of stochastic factors that could drive to local minimum traditional optimization techniques (Bruzzone, Signorile 1998). In addition the approach is very robust and reliable even for evaluating different performance in term of service and maintenance of Power Plants considering market change (Bruzzone, Giribone 1998). Therefore Pooling Power plant maintenance is based on the idea of a concurrent collaborative planning and management of the service over a set

rotors elements have a cost of a million USD each, hot gas parts require specific controls), therefore for sure Gas Turbines and, in particular, their blades represent the most critical element to be optimized in term of service due to their costs, lead times and sensibility to different operative modes. Each machine have many components that need to be checked, substituted and/or refurbished; for many types of turbine blades, for example, is possible the refurbishment or re-coating: it corresponds to an hi-tech process devoted to rebuild blade surface of the blades; depending by the model of the blades the process can be repeated one, two or even three times before to require the substitution with new ones; in addition the refurbishment of used component is usually costing about 1/10 respect acquisition of new elements (and a layer of gas turbine blades cost about 1 million dollar, while each single turbine include several layers), that means that an optimized management able to rotate among different sites and machines a blade layer maximizing the refurbishment is able to guarantee very big savings; therefore it is necessary to consider that at least some percentage of the elements subjected to the refurbishment processes should need to be substituted each time this is applied by new ones due to the too deep damages on the material (this is usually defined as scraping percentage in power plant blades or other regenerating items). In this context, typical stochastic phenomena are failures, scraping percentages and the quantity of components/items/subsystems to be substituted, duration of inspections as well as duration of minor and major revisions. There are complex constraints among the maintenance over different components in the same unit, site or for the same users (i.e. space for dismounting the machines or wiliness to concentrate the operation in the same time frame). Due to this fact optimizing the preventive maintenance scheduling is not enough to manage effectively this system-of-systems, but it is required to define inventory management policies in coordinated way and to plan refurbishment activities: so it becomes necessary to simulate long term scheduling to check mutual influence of different choices, to check transversal constraints. In fact the Preventive Maintenance of many components is regulated by Equivalent Operative Hours (EOH) able to consider not only the their use, but also special operational mode that are reducing component life cycles (i.e. startups, shot downs, etc.); the EOH value is evolve obviously as a not deterministic variable due to many factors; the authors defined a set of parameters to characterize each machine that combine solar working hours of the plants (related to power production), the

757

results is that pooling always improve the service levels (Taragas 1989); A reliable support for manager of service part inventories are demand pattern identifier based on statistical methods (Cerda et al. 1997; Sugita et al. 2005; Paschalidis et al. 2004; Muckstadt 2005; Beardslee et al. 2006). Several authors investigates the methodologies for optimize inventory management and service for similar cases (Cohen 1990; Silver 1991; Nahmias 1994; Harris 1997). The use of statistical techniques is effective to support service and maintenance modelling and to validation and verification of the conceptual model (Hill 1997; Hill 1999; Aronis et al. 2004). The demand of spare parts for supporting service and maintenance procedures is an critical variable to model and requires specific and accurate analysis (Grange 1998). In the past was developed multi location inventory models combining simulation models and optimizers (Federgruen, 1993; Kochel 1998; Nielander 1999); to optimize inventory and transportation cost polynomial-time algorithms was used (Wang & Cheng 2007). Some Traditional algorithms applied to maintenance optimisation may be questionable (i.e. Hallefjord et al. 2003) The combination of Linear and non linear approaches (Gupta, Zhang 2010) was investigated on specific target function as well as Scheduling approach in combination with inventory optimisation (Fan et.al 2009). Complex techniques, such as genetic algoritms, permit to approach multi-objective optimization (Srinivas & Deb 1994). In power plants maintenance the decision management about replace and order new item referred to limited life cycle components was analyzed by separate and join optimization (Armstronga 1996) Supply and inspection of components must be considered by the mathematical models developed (Chen et al. 2009). The Simulation Team DIPTEM researchers have long experience in these methodologies and techniques applied to this sector (Giribone, Bruzzone & Tenti 1996). Some innovative approach was proposed, based on simulation, by the authors to support management strategies in the identification of best solutions considering multiple constraints and target functions (Bruzzone et al. 1998). In fact it was also developed a methodology for supporting pooling strategies for multiple power plant service (Bruzzone, Mosca, Pozzi, Cotto, Simeoni, 2000) as an innovative approach for defining criteria for serving the sites by clustering

of plants; in fact combining different machines, sites and power plant users it becomes possible to create a pool of entities requiring service and to identify how to manage the pool of resources for satisfy them; the success and the benefits in this care are strictly connected to the possibility to revamp items dismounted from one machine or plants and to use on another one reducing the acquisition of new components; these results are achievable by defining an effective sequence of major and minor inspections and a coordinated schedule; in fact it becomes possible to optimize the reuse of elements without affecting the availability of the power plant acting on the schedule of the service operations, refurbishment actions and on the inventory. In Power plants many kinds of components may be reused more than one time after a refurbishment process (usually considering the percentage of scraped item for each treatment); to reduce new item acquisition by increasing refurbishment component uses it is necessary to finalize an optimized management of the inventory and an effective scheduling; the service planning, for instance, should be adapted anticipating or delaying the inspection/revision in order to have refurbishment items from other plant available in time; this require to correctly major and minor revision schedule taking into account technical and commercial constrains (i.e. component life cycle vs. EOH or period of the year where it is not allowed to conduce maintenance operation due to the service contract); this activity should be based on a comparative analysis on cost reduction and profitability improvements over a stochastic scenario and in presence of risks. This approach is even more effective in large sets of plants, that require service within the same timeframe; in this case the good results emerge by a collaborative management obtained by maximizing the sharing of refurbished and new items as well as the demand. Creating a common power plant pool is possible to optimize inventory management, component safety stocks in order to reduce costs and to improve the profitability and availability over all plants and to optimize shared resources. (Bruzzone, Bocca 2008).

M&S FOR POWER PLANT SERVICE & MAINTENANCE. There are consolidated experience in managing service and maintenance in industrial power plants; modeling in simple case the service quality obtained from different management strategies the

758

well the inspections, revisions and turnover strategies which permit to minimize spare parts acquisitions costs; increasing the number of machine (units) involved in the maintenance services it is critical to be able to track each component for each unit and to be able to process and elaborate this context (Bruzzone, Simeoni 2002)

the machines in subsets able to guarantee compatible timeframes respect life cycles of spare parts, components and items; this approach leads to the optimization of availability, costs respecting technical commercial constraints. Power Market Price Behavior over a Year 200.00 180.00 160.00

NEW PERFORMANCE INDEXES

Value [Euro/MWh]

140.00 120.00

In order to identify the optimum in terms of service the authors proposed a specific metrics that permit to balance, in weighted way, cost and availability. The costs are defined considering refurbishment and acquisition of spares, the availability is defined in relation to the value of the market price for energy considering that in different days, hours and months (see fig.1), unavailability of the items (and then of the plants) correspond to profit lost of profit. Profit lost is different in different time frame.

100.00 80.00 60.00 40.00 20.00 0.00 0

30

60

90

120

150

180

210

240

270

300

330

360

Time [days]

Power Market Price Behavior over a Week 80

70

Value [Euro/MWh]

60

50

40

SP = kaVA + kcVC

30

20

0 1

2

3

4

5

6

np

np

np

i =1

i =1

i =1

OCe = ¦ Acei + ¦ Rcei + ¦ Scei

10

7

Time [days]

º ª ΔTi naij = int « » ¬« LC j (nre j + 1)»¼

Power Market Price Behavior over a Day 160

140

nrij =

ΔTi − na j LC j

Value [Euro/MWh]

120

nc

Acti = ¦ naij ca j

100

80

j =0

60

VC =

40

20

SP ka kc np nc nrej

0 0

4

8

12

16

20

24

Time [hours]

Figure 1 – Power Market Price Behaviours Therefore the improvement of power plants performances in terms of service and maintenance, go trough fast supply chain, flexible stock management, lean process that became key elements for a correct management DSS; so the integration of operation, administration, maintenance and business have a relevant importance on the power plant performances such as efficiency, reliability, availability, environmental impact and lifecyclecost (Bruzzone, Mosca Simeoni, Pozzi Cotto, Fracchia 2000). The difficulties in the definition of a correct service and maintenance procedures in power plants need to be approached by modeling the revamping and refurbishment processes for specific components as

ΔTi OCe Acei Rcei Scei caj crj VC

759

nc

Rcti = ¦ nrij crj j =0

Acti + Rcti Oce service performance availability importance factor cost importance factor number of plant in the pool number of maintenance components maximum number of possible refurbishment for j-th component time frame for maintenance of i-th plant Effective Overall Cost Component Acquisition Cost for i-th plant Component Refurbishment Cost for i-th Plant Extra Cost due to stop for i-th Plant unit acquisition cost for j-th component unit refurbishment cost for j-th component cost performance index as ratio effective vs. minimum theoretical service costs

planning GT Theoretical maximum revenues i =1 ti 0 GT Theoretical maximum revenues of i-th plant i t 0+ ΔT i ' i* i* i* i* i* i* without any planned maintenance G i (t F , t F , t F ) = ³ PN g a (t , t FM , t Fm , t Fisp ) f (t )dt G’i Maximum revenues of i-th plant with optimal t0 i i i i Planning maintenance g a (t , t i FM , t i Fm , t i Fisp ) = g isp (t , t i Fisp ) g M (t , t i FM , g m (t , t i Fm ) Rei effective revenues based on decided planning ­0 t i FM ≤ t ≤ t i FM + Δt i F tFispl technical/contractual interval between ° i inspections g M (t , t i FM ) = ®1 t > t i FM + Δt i FM i technical/contractual interval between minor t °1 Fml t < t FM ¯ revisions ­0 t i Fm ≤ t ≤ t i F + Δt i Fm tFMl technical/contractual interval between ° i i i i major revisions g m (t , t Fm ) = ®1 t > t Fm + Δt Fm technical/contractual tolerance between λ i °1 isp t < t Fm ¯ inspections ­0 t i Fisp ≤ t ≤ t i Fisp + Δt i Fisp λm technical/contractual tolerance between minor ° i i revisions g isp (t , t F ) = ®1 t > t i F + Δt i F i λ technical/contractual tolerance between °1 M t < t Fisp ¯ major revisions t i 0 + ΔTi ΔtFMi theoretical/contractual duration of major R e i = ³ PN i g i e (t ) f (t ) dt review for i-th plant ti 0 i theoretical/contractual duration of minor Δt Fm i * * * * * t * Fisp , t i Fm , t i FM / G ' i (t i Fisp , t i Fm , t i FM ) < G ' i (t i Fisp , t i Fm , t i FM ), review for i-th plant ΔtFispi theoretical/contractual duration of inspection ∀t i Fisp , ∀t i Fm , ∀t i FM , ∀k , for i-th plant | t i Fisp (k ) − t i Fisp (k − 1) |< λisp t Fispl , ti 0 + ΔTi

np

GT = ¦ GTi

³P

GTi =

Ni

i

isp

m

f (t )dt

i

M

i

i

M

m

isp

isp

isp

The aim of the application of the model is to find the best & feasible schedule and inventory management in order to optimize the general performance of the plants. The best result is evaluated by considering expected prices of power along days/weeks/months/years. To do this, and test different scenarios, is possible to use historical data of power consumption integrated with trend hypothesis. Using Genetic Algorithms integrated with a simulation it is possible to estimate the fitness function in term of value and confidence band related to the different performance indexes considering the complex relationships among variables and parameters; the authors realized LAPIS framework in order to support the optimization and analysis of service and maintenance planning & policies as well as for supporting decision making in this framework,

| t i Fm (k ) − t i Fm (k − 1) |< λmt Fml , |t

i

FM

(k ) − t

i

FM

(k − 1) |< λM t FMl

t i 0 + ΔTi

³ ge (t )dt i

e

ti 0

Avi =

Prod p =

ΔTi np

Prod c = ¦ i =1

Ri ΔTi ⋅ G 'i (t i

VA =

tio f(t) i ga

¦G i =1

'

i

(t

i*

Fisp

,t i

e

*

PNi ⋅ Ri

np

* Fm

1 np Ri ¦ GT i =1 ΔTi

Fisp

,t i

* Fm

,t i

* FM

)

e

,t i

* FM

)

np

¦ PN

i

i =1

initial time for i-th plant Power Price at t time nominal operative state

vector with time of m1Major revision for i-th plant vector with time of m2 minor revision t i Fm for i-th plant t i Fisp vector with time of m3 inspections revision for i-th plant i t (k ) element k-th of the vector t i t i FM

MODEL VARIABLES The main purpose of the simulation model defined by the authors is to be a DSS able to properly evaluate the KPIs over complex scenarios; the LAPIS is stochastic discrete event simulator integrated with an Optimizer based on Genetic Algorithms; in fact key performance indexes are affected by several variable such as:

*

optimal set of Major revision time for i-th plant optimal set of Major revision time for i-th plant t Fm i optimal set of Major revision time for i-th plant t Fisp ge(t) effective operative state based on decided

ti

FM

i*

*

760

• • •

• •

Effective Planning for each Plant Effective substitution/mounting Sequencing for compnents subjected to refurbishment Schedule Performances quantify the respect of time constraints such as : • Inspection/Revision exceeding the allowed time due to delays/problems (i.e. extra time for a revision due to spare part shortage) • Technical Times Interval among inspections/revision not respected (i.e. too many EOH before substituting some blade layer) • Dates not acceptable due to contract constraints (i.e. desire to avoid maintenance in some months with higher power prices or viceversa desire to concentrate all the maintenance within summer holiday break) • Too short interval among two sequential different machine inspection/revision on the same site respect desired value (i.e. desire to avoid to operate them concurrently inside the same power building due to interference and lack of space) • Too long interval among two sequential different machine inspection/revision on the same site respect desired value (i.e. desire to operate them concurrently inside the same power building creating synergies with service resources and kits) • Too few machine concurrently unavailable respect desired value (i.e. desire to distribute the maintenance to guarantee average power generation capability of the plant users) • Too long interval among two sequential different machine inspection/revision on the same site respect desired value (i.e. desire to operate them concurrently inside the same power building creating synergies with service resources and kits) Power Plant Number of Stops and Duration Inventory behavior for each Component • Warehouse Quantities • Stock-out Times, Importance and Quantity (i.e. how many spare parts of the component are missed, how critical it is the component for plant operations and how much it costs to acquire through unconventional channels) • Component Service Level • Component Rotation • Expected Final Status of the Component at the end • Quantities and values of spare parts of the component distinguishing among new ones and refurbished at the



different levels are mounted in the machines/plants at the beginning of the contract • Quantities and values of spare parts of the component distinguishing among new ones and refurbished at the different levels are mounted in the machines/plants when service contract expires • Quantities and values of spare parts of the component distinguishing among new ones and refurbished at the different levels are available in the warehouses when service contract expires Costs & Profitability detailed as: • New Spare Part Acquisition Costs • Refurbishment Costs • Warehouse Fees • Expediting Fees • Initial Costs for the defined Configuration • Plant Expected Profitability Risk Reports • Risks of Service and Maintenance Delays • Risk on Component Shortage • Risk of Power Plant Stops (Number and related duration)

The model implements many alternative management policies for power plants service as well as estimation criteria and these are defined inside the simulation model. In fact are several simulation parameters that need to be set in order to properly estimate the performances such as: • • • • • • • • • • • • •

761

Replication Number for each scenario evaluation in order to estimate the stochastic factor influence Pseudo Random Number seeds (or automatic initialization) Simulation Duration Power Plant Pool Configuration Inventory initial configuration Initial Scheduling Operative Management Criteria Inventory Management Policies Policies for restoring of Safety Levels Policies for managing Expediting Policies for Interchanging compatible Components Policies for Cannibalization of Components in planned maintenance occurrences Policies for Cannibalization of Components due to failures

• • • • •

The box called “FUSE” is an interesting module of the system, where the application of Fuzzy Techniques is used to evaluate the interaction among technical (machine lifecycle), operational (interference among inspections), contractual (periods preferable for maintenance) and commercial factors such energy request from the market (Bruzzone et al. 2004). In fact LAPIS is fully integrated with this fuzzy logic performance evaluator, developed in previous researches, devoted to evaluate the quality of the planning of maintenance operations by a hierarchical approach (Bruzzone & Williams 2005). Fuzzy Logic is very useful in this context due to the uncertainty on variables and constraints (Cox 1994). Failures, planned maintenance events, critical time points such as shut downs, start-up, contract closure, item delivery and several other events are driving the time advance in LAPIS simulator; while the power demand behavior and unit EOH (Equivalent Operating Hours) are computed by the integration of the function between two consecutive events. The stochastic variables are computed by using Montecarlo Tecniques, the simulator for each time and in each run extract the value of the variable from distribution function. The probability distribution of the variables was identified analyzed by statistical techniques (Test Chi2 T and by the Subject Matter Expert in order to identify the best fitting of the real data with the known Probability Distributions. For several reason (few historical data, short history, errors in records, confidential nature of the information etc.) in most of the cases without strong historical background the authors used Beta Distributions to model stochastic variables. In fact Beta distribution allows to integrate easily the expert estimations with historical data in order to have consistent data. There are three types of optimization modes supported by LAPIS architecture:

Policies for processing Automatic Collected Data related to Power Demand and Plant use Policies for managing contract duration Definition of Technical Constraints Definition of Contractual Constraints Definition of Resource Constraints

In fact the initial conditions set in the model are used to start the simulation of a specific scenario; during the simulation run the model reproduce the operation in service and maintenance of the plants considering unexpected failures, managing initial schedule and inventory. In order to verify and validate the model on all the events, costs and indexes the simulator generates a log file that contains all estimations of different stochastic components compared to initial planning and management strategies

Figure 2 – LAPIS General Architecture LAPIS simulator generates a lot of reports that represent, also in graphical mode, the temporal evolution of the following variables: • Each Plant Unit EOH • KPIs • Profitability • Costs • Component Consumption • Component Levels on the Warehouses • Refurbishment Quantities • Failures • Inspection, Minor and Majors Revisions

• Planning Optimizer (PO) • Inventory Optimizer (IO) • Combination of PO and IO All the optimization models deal with the dynamic interaction among the GAs optimizer and the stochastic discrete event simulator. The optimizers, as anticipated, use Genetic Algorithms (GAs) in order to find robust and cost effective solutions (Bruzzone A.G., 1995). In this application the GAs are initialized from a set of solutions called “population” including proposals of the users; the genetic operators referred to fitness

LAPIS ARCHITECTURE & COMPONENTS As anticipated, LAPIS solution is the combination of simulation and optimization (see figure 2), as sown in the architecture scheme the simulator is connected to each component (is the core of this proposed approach).

762

needs in terms of maintenance. The analysis by the simulation of the maintenance policies of each component may be and hard and unproductive work, for this reason in combined power plants the maintenance procedures are driven by the most important item that is the Turbo Gas (TG) The template of the sequence of events has the following order: I-I-PR-I-I-GR (I=Inspection, PR=Partial Revision and GR=General Revision). All the characteristics of the Items are stored in the company Data Base (DB). All the scenario processed are stored in DB Scenario in order to collect a historical set of analysis performed (useful for future research). In order to have the better interaction with the users the authors develop three different kind of output reports: a customer report (it is possible to send it directly to the customer), a pre-customer report (these reports need to be checked by the user before sending them), and a user-report (to control if the results are consistent with the related scenario). Due to the high computational workload related to this very complex model, it was decided to implement it in C++ allowing to run optimization process within reasonable time (i.e. few minutes for simple case/partial optimization, few hours for complex scenarios); at the same the authors are used to consider the automatic optimization not as a stand alone solution, but as a procedure to be run by decision makers interactively while they test new hypotheses and ideas; in fact changing some hypotheses on exogenous factors the best solution change and the decision makers need to compare the results and evaluate the reliability of data and evaluation provided by experts. So in order to guarantee an easy access to simulation and optimization results as well as en effective analysis tool the LAPIS report are post processed and handled by a module implemented using MS Office Suite, while FUSE model provides additional capabilities in evaluating the solutions. By using the report carried out by the combined use of LAPIS and FUSE the decision makers are able to quickly evaluate, accept, modify or reject alternative proposed solutions. The authors worked on optimization and simulation of pooling strategies for managing several power plants with different spare parts; therefore due to complexity of the system it was critical to guarantee an effective interface to make the use of the program easy for decision makers and to support all the functions provided by LAPIS. LAPIS VV&A (Verification, Validation and Accreditation) was extensively applied even in term of dynamic analysis of simulation/optimizer over several complex scenarios. Several month of work, desktop review and dynamic testing in cooperation

function, obtained by running automatically the simulator, to guide the search and improving solutions (Bruzzone, Bocca 2008); the genetic algorithms include: • selection • recombination (crossover) • mutation The optimizer elaborate the population and evaluate their fitness by running automatically the simulator and it recombines the solutions by the above mentioned algorithms for several generations; the parameters of the GAs could be set by the user as well as the weights of the fitness function to be used for each optimization.

LAPIS FRAMEWORK Due to its complexity the model need a very high quantity of input data, it is so hard to modify this data to create different scenarios maintaining it consistent; in fact most of data needed for running the simulation (such as existing planning, technical data of items and spares, levels of the storage) are extracted by the company ERP System; therefore in LAPIS an easy interface is defined to quickly check and change hypotheses for creating different scenarios.

Figure 3 Lapis Function Configuration Figure 3 show the LAPIS components, the user can interact with the model by using a Graphical User Interface (GUI) with whom is possible to modify the input parameters and the settings (of the model to change the characteristics of the scenario and also of the simulator) (Bruzzone & Simeoni, 2002). The scheduling contains the data about the interventions planned by the user to make maintenance to each power units of the pool. The scheduling depends both of contracts and customer; it is not a fixed variable but the user must input it by using the GUI. Each complex plants is composed by different item with different characteristics and then different

763

with Subject Matter Experts (SME); in particular power plant service experts (i.e. project managers, supply chain management team member, planners) supported the validation phase. Several Statistical Test Technique was applied to LAPIS models such as Analysis of Variance (ANOVA), Mean Square pure Error (MSpE), Confidence Band, Statistical Comparison and Sensitivity Analysis (as presented in figure 4).



Parameter D: Computing/Neglecting the Residual Value of the item mounted in the machines at the end of the contract in the overall service cost KPI; this parameter is important because permit to consider the residual value of the material in the plants (especially if the customer don’t renovate the contract).

As shows in Figure 5, considering the availability of the power plants the sensitivity analysis shows how, in all the scenario tested, the residual value of the plants don’t affect the availability of the plants. Analyzing the contrast the graph shows that the scraping impact on the availability of the plants in inverted proportional way; unfortunately the scraping is not easy to control, in general sense scraping should be considered as an exogenous factor that evolves with the time, usually positively due to new technological material developments and operational experience; so this variable is able to produce future benefits that are estimated by LAPIS, but it is not subjected to a direct control by users or service providers.

Figure 4 – LAPIS Verification and Validation In the paper the results of sensitivity analysis based on Design of Experiments (DOE) are proposed as VV&A; the analysis allows to quantify the effects and the contrasts of selected parameters on the target functions as well as their interaction. The case proposed for this analysis and optimization is related to a realistic scenario involving a collaborative service provided to nine different combined units (steam turbine plus gas turbine) located over different sites. The aim of this analysis is to identify what, and in which way (direct or inverse proportionally), a controlled variable impact on a target function (i.e. availability of the pool and the overall cost); in similar way all the combined effect of different variable are estimate, in fact due to the high degree of not linearity of the problem under analysis the high level order effects are usually significant and cannot be neglected . In the paper the analysis is focalized on some of the main parameters, according to the SME suggestions: • Parameter A: Number of kits of Gas Turbine blades available (condition tested over the values: 3 kits or 6 kits). • Parameter B: Enable or disabled the Cannibalization of new kits of blades. • Parameter C: Scraping Percentage, that indicate the percentage of damaged blades that can’t be refurbished at each step (based on company historical data and SME: from 1% to 5%)

Figure 5 – Sensitivity Analysis Availability Trough the analysis of Figure 6 related to a different target function (overall service costs) is possible to identify a particular behavior where the combination of 2 variable (A and B) is very big; in fact the influence on the related target function is

764

Figure 7 propose the evolution of the optimization process during an optimization; it is evident the significant saving and benefits provided by the integrated use of simulation and GAs.

much more than the impact of the individual parameters if considered separately (please note that the scale is logarithmic); this is a classical confirmation of the complexity of the phenomena generating a very not-linear behavior. In fact the possibility to cannibalize the new kits permits to optimize the maintenance operation and then to reduce total cost (as contrast shows).

CONCLUSIONS The architecture of the model resulted very reliable and robust even when applied to pretty complex scenarios; currently this approach is proposed by the author to support decision making in terms of service and maintenance in system-of-systems service for major companies . The evaluation of the profitability combined with availability related to the market prices represent an example of how it could be possible to redirect the service management to most efficient approaches by properly evaluating the overall effectiveness and efficiency; in power industry for instance this aspect have a growing impact and the use of such models could guarantee significant improvements and high level of competitiveness both for service providers as well as for users; in fact a major benefit of introducing these advanced models in the service of complex systems it is related to the sharing among providers and users of a common understanding of the scenario with possibility to achieve much better results and lean decision making process both in management and acquisition of service contracts. the introduction of this innovative approach in power plant service allows to define pooling management strategies able to evaluate more properly the real performances and to change the scheduling criteria for inspections/revisions as well as the policies in use (inventory management, safety stock, refurbishment services). Due to its effectiveness LAPIS was effectively used to be integrated in the company decision making process for supporting complex power plant’s service and maintenance.

Figure 6 – Sensitivity Analysis Total Cost The obtained results support to development of robust solutions able to consider inventory costs, stop costs, availability, contractual term respect and constraints in fact the simulator driven by the to genetic algorithms is able to provide useful service management solutions.

REFERENCES

Service Combined Optimization Process on Scheduling and Inventory Management

1.

100% 98% 96% 94%

SP

92% 90% 88%

2.

86% 84% 82%

20

70

120

170

220

270

320

370

420

Generations

Figure 7 Lapis Optimisation Evolution

765

Armstronga M.J., Derek R. Atkinsa (1996) "Joint optimization of maintenance and inventory policies for a simple system" IIE Transactions, Volume 28, Issue 5 May 1996 , pages 415 - 424 Aronis P.K., I. Magou, R. Dekker and G. Tagaras, (2004) “Inventory control of spare parts using a Bayesian approach: A case study”, European Journal of Operational Research 154, pp. 730–739

3.

4.

5. 6.

7.

8.

9.

10.

11.

12. 13.

14. 15. 16.

17. Cerda R., C. B., and Espinosa de los Monteros F., A. J., (1997) “Evaluation of a (R,s,Q,c) multi-item inventory replenishment policy through simulation”, Proceedings of the 1997 Winter Simulation Conference, pp. 825-831. 18. Chen SC, Yao MJ, Chang YJ (2009) "A Genetic Algorithm for Solving the Economic Lot and Inspection Scheduling Problem in an Imperfect Production/Inventory System" 8th International Conference on Information and Management Sciences, Kumming, China, July 19. Cohen, Morris, Pasumarti V. Kamesam, Paul Kleindorfer, Hau Lee and Armen Tekerian, 1990, "Optimizer: IBM's Multi-Echelon Inventory System for Managing Service Logistics," Interfaces, vol. 20, no. 1 (Jan-Feb), p. 65-82. 20. Cox E. (1994) "The Fuzzy System Handbook", AP Professional, Chestnut Hill, MA 21. Fan, BQ; Chen, RJ; Tang, GC (2009) "Bicriteria Scheduling on Single-Machine with Inventory Operations" 3rd International Conference on Combinatorial Optimisation and Applications, Huangshan China, June 10-12 22. Federgruen, A., (1993) “Centralized planning models for multi-echelon inventory systems under uncertainty”, in Graves, S., Rinnooy Kan, A., Zipkin, P. “Handbook in Operations Research and Management Science”, Vol. 4, Logistics of Production and Inventory. NorthHolland: Amsterdam. 23. Giribone P., Bruzzone A.G. & Tenti M. (1996) "Local Area Service System (LASS): Simulation Based Power Plant Service Engineering & Management", Proceedings of XIII Simulators International Conference SMC, New Orleans LA, April 8-11 24. Grange F. (1998) "Challenges in modeling demand for inventory optimization of slowmoving items", 30th Wintersim Washington, D.C., United States, November 25. Gupta S, Zhang N (2010) "Flexible Scheduling of Crude Oil Inventory Management", Ind. Eng. Chem. Res., pp 1325–1332, Washington DC, USA, December 26. Harris, Terry, 1997, "Optimized Inventory Management," Production and Inventory Management Journal, vol. 38, no. 1, p. 22-25 27. Hallefjord Å., K. Jörnsten, S. Storøy and S. W. Wallace (2003) “Inspection and maintenance optimization methods” Michelsen Institute, Fantoft, Norway. 28. Hill R.M., (1997) “Applying Bayesian methodology with a uniform prior to the single period inventory model”, European Journal of Operational Research 98, pp. 555–562.

Beardslee E.A., Theodore B. Trafalis (2006) "Discovering Service Inventory Demand Patterns from Archetypal Demand Training Data" University of Oklahoma, USA Bruzzone A.G. (1995) "Fuzzy Logic and Genetic Algorithms Applied to the Logistical and Organisational Aspects of Container Road Transports", Proc. of ESM95, Praha, June 5-7 Bruzzone A.G., Kerckhoffs (1996) “Simulation in Industry ”, Genoa, Italy, October, Vol. I & II, ISBN 1-56555-099-4 Bruzzone A.G., Giribone P., Revetria R., Solinas F,, Schena F. (1998) "Artificial Neural Networks as a Support for the Forecasts in the Maintenance Planning", Proceedings of Neurap98, Marseilles,11-13 March Bruzzone A.G., Giribone P. (1998) "DecisionSupport Systems and Simulation for Logistics: Moving Forward for a Distributed, Real-Time, Interactive Simulation Environment", Proceedings of the Annual Simulation Symposium IEEE, April Bruzzone A.G., Signorile R. (1998) "Simulation and Genetic Algorithms for Ship Planning and Shipyard Layout", Simulation, Vol.71, no.2, pp.74-83, August Bruzzone A.G., Mosca R., Pozzi Cotto S., Simeoni S. (2000) "Advanced Systems for Supporting Process Plant Service", Proc.of ESS2000, Hamburg, Germany, October Bruzzone A.G., Mosca R., Simeoni S., Pozzi Cotto S., Fracchia E. (2000) "Simulation Systems for Supporting Gas Turbine Service Worldwide", Proceedings of HMS2000, Portofino, October 5-7 Bruzzone A.G., Simeoni S. (2002) "Cougar Concept and New Approach to Service Management by Using Simulation", Proc/of ESM2002, Darmstad Germany June 3-5 Bruzzone A.G. (2002) "Supply Chain Management", Simulation, Volume 78, No.5, May, 2002 pp 283-337 ISSN 0037-5497 Bruzzone A.G., Simeoni S., B.C. (2004) "Power Plant Service Evaluation based on advanced Fuzzy Logic Architecture" Proceedings of SCSC2004, San Jose' Bruzzone A.G., Williams E. (2005) "Summer Computer Simulation Conference", SCS, San Diego, ISBN 1-56555-299-7 (pp 470) Bruzzone G. A., Bocca E. (2008) “Introducing Pooling by using Artificial Intelligence supported by Simulation”, Proc.of SCSC2008 Bruzzone A.G, Madeo F., Tarone F. (2010) "Pool Based Scheduling And Inventory Optimisation For Service In Complex System" Proc. of the 16th International Symposium on Inventories, Budapest, August 23-27

766

29. Hill R.M., (1999) “Bayesian decision-making in inventory modeling”, IMA Journal of Mathematics Applied in Business and Industry 10 (1999), pp. 165–176. 30. Köchel P. (1998) “A survey on multi-location inventory models with lateral trans-shipments” In: S. Papachristos and I. Ganas,, Editors, Inventory Modelling in Production and Supply Chains, 3rd ISIR Summer School, Ioannina, Greece, pp. 183–207 31. Muckstadt, J.A., (2005) “Analysis and Algorithms for Service Parts Supply Chains” Springer, New York 32. Nahmias, Steven, 1994, "Demand Estimation in Lost Sales Inventory Systems," Naval Research Logistics, vol. 41, pp. 739-757. 33. Nieländer, U., 1999 “Simulation Optimisation of Kanban Systems using a non-standard Genetic Algorithm” Proc. Of 4th ISIR Research Summer School on Inventory Modeling, Exeter, UK. 34. Paschalidis, I. C., Liu, Y., Cassandras, C. G., Panayiotou, C. (2004) “Inventory control for supply chains with service level constraints: a synergy between large deviations and perturbation analysis”, Annals of Operations Research 126, pp. 231-258. 35. Silver, Edward A., 1991, "A Graphical Implementation Aid for Inventory Management of Slow-Moving Items," Journal of the Operational Research Society, vol. 42, no. 7, pp. 605-608. 36. Srinivas N., K.Deb,(1994) “Multiobjective optimization using nondominated sorting in genetic algorithms”, Evol. Comput.,vol.2, no.3, pp.221-248. 37. Sugita, K., and Fujimoto, Y. (2005) “An optimal inventory management for supply chain considering demand distortion”, Industrial Informatics, 3rdIEEE International Conference on (INDIN), pp. 425-430. 38. Tagaras G. (1989) "Effects of Pooling on the Optimization and Service Levels of TwoLocation Inventory Systems" IIE Transactions, Volume 21, Issue 3 September 39. Wang XL , Cheng TCE (2007) "Logistics scheduling to minimize inventory and transport costs" International Conference on Industrial Engineering and Systems Management, Beijing

767

MODELING OF OBESITY EPIDEMICS BY INTELLIGENT AGENTS Agostino G. Bruzzone, MISS DIPTEM University of Genoa Email [email protected] - URL www.itim.unige.it Vera Novak BIDMC, Harvard Medical School Email [email protected] - URL http://www.bidmc.org/SAFE Francesca Madeo M&S Net Email [email protected] - URL www.m-s-net.org Cecilia Cereda Simulation Team Email [email protected] - URL www.simulationteam.com

KEYWORDS Health Care, Simulation, Intelligent Agents, Human Behavior Modeling ABSTRACT The paper focuses on a large scale problem related to population health care with special attention to obesity. The authors present a proposal for modeling human behavior and its influence on the evolution of obesity epidemics, and its effects on social networks, infrastructures and facilities. This approach is based on Intelligent Agents tools developed for reproducing country reconstruction and human factors. These models represents the base that allows to add specific and complex aspects related to pathologies and correlated behaviors that allows to reproduce these phenomena.

2011 Situation

1991 Situation INTRODUCTION

Figure 1. Obesity Epidemic evolution in last 20 years

Today one of the main problems is to adapt health care to the existing challenges in the society. This means that both public and private health care system have to face the problem to obtain and properly allocate resources for prevention, treatment and rehabilitation of a large assisted population with limited assets. In fact, these challenges are expected to grow in coming decades, due to a variety of reasons in different world regions: i.e. population growth, changes in life-span expectation, aging of the population, social and economic evolution. To solve this problem it is necessary to improve the effectiveness and efficiency in allocating the available resources. Therefore, a better understanding of these complex phenomena affecting population health aspects is currently very critical to properly define health policies, actions and to plan future infrastructures and services to be able to face such challenges.

It is evident that in health care more than in other sectors the high influence of stochastic factors and very complex correlation make difficult analysis on large scale without using modeling and simulation; so a very interesting aspect is obtain a better understanding of the phenomena generating/affecting current development of obesity epidemics that is dynamically and quickly evolving in many countries; and therefore, providing challenges today and threats for the future. The authors propose to develop simulation models able to reproduce human behavior to address this problem and to provide support for decision makers; a first important benefit expected by these simulators should be the possibility to test and to validate the different existing hypotheses related to the mutual influence among different factors related to physiological and psychological issues, behavioral aspects and regional/ethnical/social/geographical and economical

768

factors. In fact, by conducting experiments using information available in larger samples, it will allow us to validate the consistency of these predictions within a population over time and also across diverse populations.. Once the models become validated, it will be possible to use a forecasting system, that would allow to conduct risk analysis and estimation of future development; the simulator, used in this way, reproduces scenarios and generates forecasts in term of resources and facilities required for treatments, as well as estimation of the impact of costs and demand on the country infrastructures and industry. It would also allow to predicting the impact of behavioral, social and economic interventions to prevent further development of obesity epidemics and to curtail its costs. Finally such simulators could become a very strategic advantage to support decision and to evaluate the impact of actions and countermeasures of public and private institutions and organizations on such a critical sector as the health care. The authors decided to start a joint research on these issues by focusing on obesity due to its strong impact on country economies and to its very complex and dynamic evolution (as proposed in figure 1, source World Health Organization Statistics Reports); in fact obesity of a population evolve based on individual and social behaviors over time (i.e. depression due to some social problems, lack of mobility, socially acceptable overeating etc. ). The use of intelligent agents in this area represents an innovative opportunity for research; currently the authors present the first development on this track based on some available data and some adaptation of their simulation frameworks to this new context. It is expected that such research will be further developed in support of specific R&D programs.

Childhood: Influence of Parents

Changes in Behavior

BMI =

W H2

W H

Weight [kg] Height [m]

BMI in Childhood

Adult BMI

Mortality & QALYs

Related Pathologies

Figure 2. Basic Model of Obesity in Childhood

Children at risk for overweight are defined as the 85th and 95th percentiles of body mass index. Currently, (2010), no state has prevalence of obesity less than 20% and 26 states had a prevalence of 25% or more (Centers for Disease and prevention; www.cdc.gov); 66% of adults are overweight or obese; 16% of children and adolescents are overweight, and 34% are at risk of overweight. Minority and low-socioeconomic-status groups are disproportionately affected at all ages. By 2015, 75% of adults will be overweight or obese, and 41% will be obese. Obesity is associated with increased risk of diabetes, hypertension, cardiovascular diseases, strokes and dementia, mobility dysfunction, cancer and mortality thus increasing significantly health care cost in the society. Elevated body mass index is being increasingly recognized as a risk factor for stroke, cardiovascular disease, and cognitive decline (Falkstedt et. al. 2006; Cournot et. al. 2006). Diabetes epidemic follows obesity spread in the world, affecting countries that previous had a lower prevalence of diabetes. It is expected that diabetes prevalence of people with diabetes will dramatically increase further between 2010 and 2030 i.e. India from 50.8 to 87 millions; China from 43.2 to 62.6;United States of America from 26.8 to 36.0;Pakistan from 7.1. to 13.6; Brazil from 7.6 to 12.7 (Wang et al. 2007). A long-term population study with 27 years of the follow-up has shown prospectively that in the multiethnic population, midlife obesity increases the risk of dementia later in life. Obese people (BMI ≥ 30) had a 74% increased risk of dementia (hazard ratio 1.74, 95% confidence interval 1.34 to 2.26), while overweight people (body mass index 25.0-29.9) had a 35% greater risk of dementia (1.35, 1.14 to 1.60) as compared with those of normal weight (body mass index 18.6-24.9) (Whitmer et al. 2005). The increased risk for Alzheimer’s disease and dementia later in life is independent even if adjusted for elevated blood pressure, smoking, socioeconomic status and genetic factors (Wolf et al.2007;Kivipelto et al.2005).

The obesity epidemic (Wolf et al. 2007) has been increasingly spreading worldwide in the past three decades, involving even the countries that never in the past showed obesity among their population. The United States has observed of the highest rate of obesity increase in the world (Wang et al. 2007), affecting people of all ages including children, both genders, different ethnic and racial backgrounds, and various socioeconomic groups. Adult overweight and obesity are defined using body mass index (BMI): normal weight < 25; overweight: 25-29.9; obesity > 30 Body Mass Index

Weight

Childhood: Preventive Actions

THE OBESITY EPIDEMIC

BMI

Imbalance Nutrition

769

have, always considered in comparison to a normal weight individual. A recent study in Australia has estimated the intangible costs arising from obesity in adulthood, reaching a value between 13 to 18 billion AU$; in addition to these aspects, many studies have proposed estimations of indirect costs of obesity highlighting the fact that their economic impact could exceed the direct health care costs, both in term of absolute and percentage of GDP (Magarey et al 2001). For instance in the U.S.A., it is estimated a loss of production due to the obesity corresponding to 23 billion USD (1989-1990), while in Australia the loss for the country is estimated into 272 million AU$ (Australian Bureau of Statistics). Recently some researchers have carried out a study with respect to the Chinese context, the analysis was derived from two reference years, 2000 and 2025 (Popkin et al 2006); based on these estimates it results, for the year 2000, an economic impact of 49 billion USD (4.06% of GDP), of which 43 billion dollars in indirect costs (3.58% of GDP). The magnitude of the economic impact is expected to reach a size even more critical, both for increased health spending and the consequences on the labor market. The projections to the year 2025 describe an expected total cost overpassing 112 billion USD representing 9.23% of GDP, of which 106 billion USD (8.73% of GDP) is attributable to indirect costs; these researches estimated that, even in China, the largest component of costs was represented by the loss of productivity due to absences due to illness, which causes 75% of indirect costs. The issues related to sickness, early retirement and disability, have been investigated in Sweden, Finland and Denmark (World Health Organization 1997). The results produced have highlighted the link between increased BMI and sickness absence in the long term. Furthermore, considering the child obesity as a risk factor for obesity in adulthood, a fraction of the costs, direct and indirect, previously mentioned, the adult is generated from the high number of adolescents in which obesity has persisted over time. In this direction, some researchers have investigated the effects expected along the course of life resulting from a weight reduction program implemented in American schools. It was introduced the parameter QALYs to quantify the results obtained: the QALY (Quality Adjusted Life Years acronym) is a unit of measure used in cost utility that combines the life span with the same quality. One QALY equal to 1 corresponds to the expected life of one year in normal health, the value 0 corresponds to death (Pliskin et al. 1980). The measurement scale is continuous and a few years of life may also be given negative values (if you have serious conditions and acute suffering of immobility). QALY is used as an index weighting in the assessment of increases in life expectancy associated with health interventions.

Obesity epidemic is associated with significant burden for the people, families and society and contributes significantly to increased health care cost, morbidity and mortality. The exact cost of obesity to the society and people is not known.

Figure 3. BACCUS: Behavioral Advanced Characters and Complex System Unified Simulator

However, a cost of diabetes could be used as an example; i.e. in 2010 -418 billions of international dollars, 8787 ID per person in the world and 55.7 billions of national income loss in China alone. It is expected that diabetes will increase death rates globally by 17% in 2030 and by th 25-27% in Middle East, India and China (IDF Atlas 4 ed. International Federation on Diabetes). Therefore, health care cost will increase exponentially, as the obesity epidemics spreads into younger population and other countries. Obesity is a multifactorial process that results from interactions among the individual health status, functional and social habits, social networks, education and other factors that cannot be predicted from a single variable (i.e. body mass) but requires nonlinear modeling of multiple variables and their interactions. Although that reduced physical activity, increased food intake and social networking are commonly cited factors, the mechanism underlying obesity spread worldwide mechanisms remain poorly understood. The mechanisms of obesity spread and their impact on the world’s health, personal life and economy have not been well studied and therefore, the effective strategies to prevent obesity epidemics are lacking. Impact of Obesity in Term of Costs The social cost of obesity is related to several aspect (Cereda 2011); as first element it was demonstrated that in fact the obese individuals impact on national health care system costs exceed the average expenditure per capita of an individual's normal weight; therefore much of the social costs associated with obesity are due to different social interactions that obese individuals may

770

Thus, for example, whether the introduction of a new surgical technique allows the patient to survive on average 6 years older, but the conditions are such that after the operation be considered equal to 0.2 QALY (eg., Because of serious motor deficiencies and frequent pain), the intervention effect on life quality will be weighted for only 1.2 years. Cost-utility analysis performed in the above studies was a cost of 4305 USD/patient per QALY gained, while the sum of the avoided health care costs and productivity losses avoided, you get an expected benefit, net of implementation costs of 7313 USD/patient for each program implemented (Wang et al.2003).

MODELLING THE PHENOMENA The idea to reproduce the phenomena related to obesity is to use intelligent agent reproducing the population and let them evolve based on their behavior and on the applied actions and scenario evolution (Bruzzone et al.2011); the behavioral models are defined inside the agents, defined IA-CGF (Intelligent Agent Computer Generated Forces); for instance an high level example is proposed in figure 2; a combined stochastic simulation engine manage the interactions among the agents; this framework defined as NCF (Non Conventional Frameworks) could be tailored over very different contexts and areas. Previous researches have been carried out by Simulation team to develop models able to reproduce human behavior over town or regions (Bruzzone et al.2008); in fact these models were used originally for epidemic evolution (Avalle et al.1999), for analyzing urban disorders (Bruzzone et al.2006) and for country reconstruction (Bruzzone & Massei 2010). In this case it was decided to create an ad hoc NCF with full inheritance of IA-CGF Libraries. For the specific research it was possible to start the development of new conceptual model and to design a first shell of NCF defined BACCUS (Behavioral Advanced Characters and Complex System Unified Simulator) that introduce several additional parameters related to physiology, health status and behavior; these parameters includes: − − − − − − − − − −

Figure 4. Investigation Process of the Obesity over a Region based on Modeling the Population Behavior

The BACCUS simulator is implemented using IA-CGF engine of the Simulation Team as anticipated and it is proposed in Figure 3. The development concept is based on a several step process; first phase is related to defining the population in term of PIG (people initialization groups); the PIGs represent the different groups present on a scenario; PIGs are defined in term of statistical distribution for their main factors (Social Level, Educational Level, Political Attitudes, Religion, Ethnic, Tribe, Gender, Age); based on these parameters the population is generated randomly respecting original statistics on the geographic region, but even considering the different groups characteristics with their specific structure (i.e. some ethnic group have different social status statistics respect other one living in the same region); this process generates the people agents; in addition inside the region it is possible to define Zones as

BMI Sport Profile Alcohol Profile Stroke Infarct Diabetes Cancer Hypertension Atrial Fibrillation Hyperlipidemia

771

objects that have affinities with the main factors and so the people agents are distributed geographically in term of living and working locations in consistency with their affinities; it is possible to overlap zones with different affinities in order to represent inconsistent mixes of different groups of the population. In addition to the individual aspects, in the model the aggregation parameters are defined to regulate the generation of social networks based on stochastic distributions; by Montecarlo technique the people agents are associated in term of families and working connection generating the social network; the behavior of each agent is defined by models that regulates how it operates under regular (i.e. working days, holidays) or special conditions (i.e. natural disasters, sickness period); an overall presentation of the procedure is proposed in Figure 4.

Figure 5. IA-CGF in BACCUS Simulator allows to evaluate the population behavior and its influence on the obesity evolution in the town

Table I: Extract of Results based on non parametric rank correlation on available samples

In addition it was used a sample related to 170 subjects in the Syncope and Falls in the Elderly Laboratory (Lab SAFE), active in the BIDMC; these 170 patients in the sample was volunteers, aged between 50 and 85 years, recruited from the SAFE Lab for four samples, conducted in the years 2006 to 2010. All four samples considered issues related to diabetes; in fact the data were related to studies correlating diabetes with other pathologies, so the sample includes other diseases such as patients with myocardial infarction or stroke. All samples have a quite balanced ratio between balance between male patients and female patients. There are some limitations in using samples of existing populations that were collected or originated for different studies, and therefore do not provided sufficient information i.e. about frequency of measurements, include specific populations, or have missing values for certain parameters. Therefore, this paper represents a first step forward for modeling and tuning simulators for investigating obesity, so the focus is mostly on creating the conceptual models, introducing consistent data and tuning the parameters in order to obtain reasonable results by the simulation runs. Some examples, of the parameters used to check mutual influences are reported in table I e II; statistical analysis and ranking methodologies was applied in order to check data significance and their correlation. The low level of correlation obtained was expected due to the reason above mentioned; therefore the analysis was useful to identify procedures and aspects to be investigated in future data collections and researches. In order to proceed it was decided to implement some behavioral model and correlation algorithms among the factors based on author's hypotheses consistent with the data available.

Table II: Some Correlation Extract among factors and BMI based on the available samples

Obviously dealing with human modeling the verification and validation of the simulator is critical as well as the data collection and analysis; for the preliminary test the hypotheses on the conceptual model relating behaviors with obesity was based on current researches (Christakis et al. 2007) as well on the analysis of data available in Beth Israel Deaconess Medical Center (BIDMC) in Boston and Harvard Medical School affiliate͘ In fact some general data used for simulation were derived from previous studies obtained by the databases of associations (i.e. World Health Organization, the American Heart Association and the American Diabetes Association).

772

In fact the MSpE allows to quantify the experimental error due to influence of the stochastic components; as presented in figure 6 the variance of all the target function reach steady state situation over a reasonable number of replications and over a time horizon of about 1 year. so this confirm that simulator provides consistent results on a stable situation with capability to define the confidence band for estimating the obesity target functions.

Mean Square pure Error 0.00000120

0.00003

Regular OverWeight Obese BMI [kg2/m4]

0.000025

361

353

345

337

329

321

313

305

297

289

281

273

265

257

249

241

233

225

217

209

201

193

185

177

169

161

153

145

137

129

97

121

89

113

81

105

73

0.00000000

65

0.000005

57

0.00000020

49

0.00001

41

0.00000040

9

0.000015

33

0.00000060

1

0.00002

25

0.00000080

17

Y (MSpE)

0.00000100

0

CONCLUSIONS The use of agents and simulation to investigate large scale health care problems represent an important opportunity; obviously in these case it is critical to guarantee a multidisciplinary approach to the problem; in fact from this point of view the authors represents a good example of different skills and background with common interest. The obesity epidemic represents a very important and interesting application framework that could be very useful to consolidate research in this area of M&S related to Medicine and Health Care. The research highlighted the critical aspects related to collecting, mining and filtering the data to define the conceptual models related to such complex problems as well as to support parameter fine tuning and simulator VV&A. The model and the present results in this first phase are promising and the potential of using Intelligent Agents in this context is very great considering the impact of the obesity epidemic. This presentation represents the first step on this research track, and currently the authors are working on some experimental analysis in term of impact on different industries (i.e. beverage and airlines) as well as in the further development of the models and their validation using datasets of a larger , longitudinal cohorts of diverse populations in different countries.

Time [days]

Figure 6. BACCUS VV&A based on Dynamic Analysis of Mean Square pure Error on Population Obesity Classes

The scenario used to the tests was a small town of about 15'000 inhabitants; target functions included the population sharing among the different obesity classes and the average BMI: − Regular − Overweight − Obese − Average BMI In figure 5 it is proposed the BACCUS during execution of the proposed scenario, analyzing each single individual. The VV&A (Verification, Validation & Accreditation) of BACCUS is based analysis of MSpE (Mean Square pure analysis) as measure of the variance of the target functions among replicated runs over the same boundary conditions; by this approach it becomes possible to identify the number of replications and the simulation duration able to guarantee a desired level of precision; MSpE values in correspondence of these experimental parameters determines the amplitude of the related confidence band: n0 ª m º m 1 « Sri (t ) − n0 ¦ Srj (t )» ¦ i =1 ¬ j =1 ¼ MSpE m (t , n0 ) = n0 n0

2

CBAm (t , n0 , α ) = ±tα , n0 MSpE m (t , n0 )

REFERENCE

t m

[1]

simulation time m-th target function estimated by the simulator no number of replications with same boundary conditions and different random seeds Srkmk(t) m-th target function value at t time of the k-th replicated simulation MspE m(t,no, α) Mean Square pure Error at t time and with no replications for the m-th target function α percentile CBAm(t,no, α) Confidence Band Amplitude at t time, with no Replications for the m-th target function

[2]

[3]

773

Bruzzone A.G. Tremori A., Massei M., Adding Smart to the Mix, Modeling Simulation & Training: The International Defense Training Journal, 3, 25-27, 2011 Cereda C., Models and Analysis of Complex Systems for the Evaluation of Future Scenarios and of Related Infrastructural and Operational Needs, Genoa University Thesis, DIPTEM Press, 2011 Bruzzone A.G., Massei M., Intelligent Agents for Modelling Country Reconstruction Operation", Proceedings of AfricaMS 2010, Gaborone, Botswana, September 6-8, 2010

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15] Wang L.Y., Yang Q., Lowry R. and Wechsler H.,Economic Analysis of a School-Based Obesity Prevention Program, Obesity Research 11, 1313– 1324, 2003 [16] Magarey AM, Daniels LA & Boulton JC, Prevalence of overweight and obesity in Australian children and adolescents: reassessment of 1985 and 1995 data against new standard international definitions. Medical Journal of Australia 174: 561564, 2001 [17] Avalle L, A.G. Bruzzone, F. Copello, A. Guerci, P.Bartoletti Epidemic Diffusion Simulation Relative to Movements of a Population that Acts on the Territory: Bio-Dynamic Comments and Evaluations, Proc. of WMC99, San Francisco, January, 1999 [18] Pliskin, J., Shepard, D, Weinstein, M, Utility Functions for Life Years and Health Status". Operations Research (Operations Research Society of America) 28 (1): 206–224, 1980 [19] Spearman C.E. "The Abilities of Man: Their Nature and Measurement", Macmillian, London, 1927 [20] Foster G.D., What is a reasonable weight loss? Patients' Expectations and Evaluation of Obesity Treatment Outcomes, Journal of Consulting and Clinical Psychology, 65 (1): 79-85, 1997 [21] WHO, World Health Statistics: Reports, 2011 [22] WHO, Prevening and Managing the Global Epidemic of Obesity: Report of the World Health Organization Consultation of Obesity, Geneva, June 1997 [23] National Institutes of Health, Clinical Guidelines on the identification, Evaluation and Treatment of overweight and Obesity in Adults: The Evidence Report, 1998

Bruzzone A.G., Scavotti A., Massei M., Tremori A., Metamodelling for Analyzing Scenarios of Urban Crisis and Area Stabilization by Applying Intelligent Agents, Proceedings of EMSS2008, September 17-19, 2008, Campora San Giovanni (CS),Italy, 2008 Wolf PA, Beiser A, Elias MF, Au R, Vasan RS, Seshadri S: Relation of obesity to cognitive function: importance of central obesity and synergistic influence of concomitant hypertension. The Framingham Heart Study. Curr Alzheimer Res 4:111-116, 2007 Wang Y, Beydoun MA: The Obesity Epidemic in the United State- Gender, Age Socioeconomic, Racial/Ethnic and Geographcis Characteristics : A systematic Review and Meta-Regression Analysis. Epidemiologic Reviews 29:6-28, 2007 Christakis N.A., Fowler J.H., The Spread of Obesity in a Large Social Network Over 32 Years, The New England Journal of Medicine, 357-4, 370-379, 2007 Falkstedt D, Hemmingsson T, Rasmussen F, Lundberg I: Body Mass Index in late adolescence and its association with coronary heart disease and stroke in middle age among Swedish men. International Journal of Obesity 31:777-7783, 2006 Cournot M, Marquie JC, Ansiau D, Martinaud C, Fonds H, Ferrieres J, Ruidavets JB: Relation between body mass index and cognitive function in healthy middle-aged men and women. Neurology 67:1208-1214, 2006 Popkin B.M., Kim S., Rusev E.R., Du S., Zizza C., Measuring the full economic costs of diet, physical activity and obesity-related chronic diseases, Obesity Review, 7, 271-293, 2006 Bruzzone A.G., Bocca E., Rocca A., Algorithms devoted to Reproduce Human Modifiers in Polyfunctional Agents, Proc. of SCSC2006, Calgary, Canada, July 30-August, 2006 Whitmer RA, Gunderson EP, Barrett-Connor E, Quesenberry CP, Jr., Yaffe K: Obesity in middle age and future risk of dementia: a 27 year longitudinal population based study. BMJ 330:1360, 2005 Kivipelto M, Ngandu T, Fratiglioni L, Viitanen M, Kareholt I, Winblad B, Helkala EL, Tuomilehto J, Soininen H, Nissinen A: Obesity and vascular risk factors at midlife and the risk of dementia and Alzeheimer disease. Arch Neurol 62:1556-1560, 2005 Australian Bureau of Statistics, National Health Survey 2004-05: Summary of results. ABS cat.no. 4364.0. Canberra, 2005

774

MARITIME SECURITY: EMERGING TECHNOLOGIES FOR ASYMMETRIC THREATS Agostino Bruzzone, Marina Massei, Alberto Tremori, MISS-DIPTEM, University of Genoa, Via Opera Pia 15, 16145 Genova, Italy Email {agostino, massei, tremori}@itim.unige.it -URL www.itim.unige.it Francesca Madeo, Federico Tarone Simulation Team, via Molinero 1, 17100 Savona, Italy Email {madeo, tarone}@simulationteam.com - URL www.simulationteam.com Francesco Longo MSC-LES, Mechanical Dept, University of Calabria, Via P. Bucci 44C, 87036 Rende, Italy Email [email protected] - URL www.msc-les.org SECURITY IN MARITIME: AN EVOLVING SCENARIO

ABSTRACT This paper analyses the evolution of a complex scenario for security: maritime environment and in particular coastal areas and harbors the critical nodes in the whole system. In new technologies provide an effective support in this asymmetric framework for situation awareness and threat assessment. M&S, CGF, Data Fusion are techniques that allow the users to obtain efficient awareness on the general on-going situation in real time and to support decision over complex scenarios.

Today Maritime Security is a very critical aspect on Marine Framework introducing the concept of Asymmetric Marine Environment with new special attention to Threats such as: – Piracy – Conventional Terrorism – CBRN (Chemical, Biological, Radiological and Nuclear) Threats Some important aspects are expected to increase over Next Years their impact in General as well in Marine Framework increasing on Asymmetric Threats such as:

Keywords: Maritime and Harbor Security, Human Behavior Modeling, Computer Generated Forces, Data Fusion

• INTRODUCTION •

This paper provide an overview on a combined approach using M&S (Modeling & Simulation) and Data Fusion techniques to analyze complex scenarios involving asymmetric marine environments; the idea to use intelligent agents (IAs) as driver for Computer Generated Forces is a very critical aspect for modeling scenarios where many entities interact (i.e. commercial and nautical traffic around a port); in order to succeed in this sector it is critical to identify the requirements for such combined solution; the authors focus this paper on the following aspects • To provide a quick Overview on the Modern Complex Scenarios and to identify the related Challenges • To present the Potential of R&D within this Framework • To Outline innovative enabling Technologies, Methodologies and Solutions for succeeding • To present actions, investments on the R&D Tracks to support these activities • To outline R&D potential Outcomes • To present Examples and Approaches in this context

• • • • • • • •

775

Movement of European Region Social Economic Center of Gravity to South increasing maritime traffic with North Africa Stabilization and Normalization Processes and Country Reconstruction Initiatives Overseas Overseas Developing Areas Growth, Production/Demand & Sustainability Issues Technologies Easier access to New Dimensions for preparing and creating critical threats (i.e. Cyberspace) Multiple opportunities to Access to Resources to develop WMD (i.e. smallpox, RDD) IT & Web empowering the potential of individuals and small groups (i.e. C2 capabilities) Increasing new reachable targets such as Oil Platform, Environmental Threats, Social Service Political Issues Political Instability on Critical Regions (i.e. Africa) Evolution of Principle of Nations and Populations (i.e. Commercial States) Evolution of new critical issues requiring rational on joint Defense and Homeland Security Budgets (i.e. natural resource issues: water)

network; in addition to physical technologies and infrastructures it is even more important the benefits provided by innovative soft computing techniques and methodologies.

The Real World: Multi Dimension and Multi Layer Resolution Asymmetric warfare is a very complex framework and modeling and simulation need to properly address all the related issues; in fact this context is: • A Real World on 5 Dimensions: • Surface • Underwater • Air • Space • Cyber • A Multi Layers & Resolutions Frame • Fleets and Parties • Ships and Commercial Traffic • Crew & People Acceding Ports/Vessels • Services & Infrastructures As explanatory Example from the new challenges in this context it could be useful to consider the evolution over the years; in fact today Modeling is critical to evaluate Strategies in Threat Identification, Decision Making & Evolution Prediction based on their behavior much more respect on their features: • Once upon Time people were used to identify threats based on Platform Detection, Identification and Classification • In Some case the same Platform is in use on multiple sides by different actors, someone friendly and someone "extremely foe". • In some case the Platform is becoming a Menace just based on own it is operating It is sufficient to consider the case of piracy to realize that the problem is not to detect the kind, class or even name of a ship based on their silhouette or using EMS, but to identify suspect behaviors that suggest presence of pirates inside a fisherman boat.

Impact of Innovative Technologies such as IA, CGF and HBM in Marine Frameworks In fact, innovative IA (Intelligent Agents), CGF (Computer Generated Forces) & HBM (Human Behavior Modeling) represent a Strategic Issues in different application areas to be applied to asymmetric marine warfare; in particular it is possible to consider the following application area and related benefits provided by these innovative solutions: – Simulation Based Acquisition and Test & Analsysis • Capability to Proceed in Data Farming on Different Hypotheses on Vessel and System Design on Virtual Prototypes – Training and Exercise • Reduction of human personnel for Training & Exercising • New Scenarios involving Dynamic Simulated Complex System vs. the old pre-defined scripts – Operational Planning • Reducing Time for Planning Development due to the reduction of human experts employed in the different roles • Possibility to Experiments different Alternatives by replicated runs carried out in Automatic way – Mission rehearsal and conduct operations • Capability to keep the simulation on-line and to conduct statistical experimental analysis Therefore, in order to apply R&D to current and future Asymmetric Marine Framework, the authors identified the following innovation tracks to be investigated: • Cognitive Technologies – Data Fusion (i.e. Situation Assessment) – Human Behavior Models – Intelligent Agents & CGF – Decision Support Systems (i.e. Web 3.0) • Modeling & Simulation – Concept and Doctrine Development (i.e. Interoperable Simulation) – Simulation Based Acquisition (i.e. Virtual Prototyping) – Training (Mobile Training, Serious Games) – Serious Games used to create complex scenarios with multiple players and to investigate different strategies • Equipment & Devices – Integrated Solutions (i.e. Mobile Tactical Control Systems) – Platforms (i.e. UAV, AUV, NMM) – Sensors (i.e. Through Wall Sensors) – Weapons (i.e. Non Lethal Weapons)

Port Protection and Asymmetric Naval Warfare For facing threats in asymmetric naval warfare it is necessary to develop new Models and Solutions able to Interoperate with the critical components of such Scenarios such as: • Non Conventional Operations • Human Behaviors on (i.e. Crew, Stakeholders, Domestic Opinion) • Services & Infrastructures • Commercial Traffic & Yachting • Port Infrastructures and Resources • Joint Operations (i.e. Ship Inspections, Littoral Control, C5I2 ) NEW ENABLING TECHNOLOGIES The existing and new technologies have a great potential in this area; for instance communication infrastructures and mobile solutions allows today to distribute information as well as data collection, data processing and decision making over a large complex

776

AVAILABLE EXPERIENCES The authors have experience in using several of these techniques in different applications, for instance, currently, the authors are developing a solution (PANOPEA), to test complex scenarios related to piracy and to investigate different C2 solutions as well as strategies and technologies within this framework; obviously the authors have even long experience in traditional applications (i.e. data fusion over conventional air naval scenarios). In the paper some experience and simulation model are presented as example of the potential of these techniques within marine asymmetric scenarios.

In the following figures the different fallout areas and contamination risks are summarized respect different kind of CBRN devices

An Example of Simulated Attack to a Port As Example of Port Attack Simulation it is proposed an unclassified simulation scenario developed in cooperation among Bulgarian Academy of Sciences, Lockheed Martin Canada, MISS DIPTEM Univ. Genova, CRTI, NATO PBIST Experts, Port Authorities, CUBRC within a NATO working group; the scenario is based on the following objects and hypotheses: • Small boat (fishing or pleasure normally seen in the harbour, not regulated by ISPS code) filled with a mixture of explosive and CBRN; • The boat is heading for a oil Terminal or tanker within a port • No predicted pattern from the pleasure boat is available • No clear strategic warning available until close to the attack • Necessity to concentrate on attack assessment and response • A Priori and HUMINT information are essential in the fusion process and need specific models

Fig.1

Areas of contamination Case 1: 100Ci 241 Am Two oil-well logging sources

For this scenario it is necessary to model and simulate available technologies such as: • Tracking small targets : Track continuity • Sensors for CBRN detection • Unusual maneuver detection for threat assessment • Data fusion for Recognizing in a very short time the attacking boat • Solutions for permanent and continuous observation and surveillance The use of simulation is an important benefits to address critical questions that are necessary to clarify for port security assessment, for new equipment design and for security procedure definition; an example of this question is following: if no automated radiation detectors are available in an Harbor Area, when is the true nature of the threat discovered?

Fig.2

777

Areas of contamination Case 2: 20kCi 90SrRussian RTG

well as all the other elements present in a specific framework; in fact the complexity to Coordinate Humans in not-conventional operations for improving their coordination and capabilities to face complex challenges is a well known element in Navy. So considering asymmetric threats it is even more important to models these Human Factors both for the directing the threats, for reproducing the boundary elements as well as for being actors for our resources; this point is even more evident by a simple example: looking to each single Vessel it is evident that for Simulating its capabilities in reacting to threats it is very critical to model the crew and its human behavior modifiers (i.e. stress, fatigue, harmony).

Fig.3

IA-CGF & Human factors in maritime security The authors have developed models for reproducing human factors and to represent intelligent agents able to direct objects within interoperable simulators; in fact the Simulation Team create a new generation of CGF, titled IA-CGF (Intelligent Agent Computer Generated Forces) for this purpose and some application within marine environment is already available and in the experimentation phase over complex scenarios. These new IA-CGF are organized based on Modules that are interoperable in HLA Federation (High Level architecture) and they include: • IA-CGF Units (i.e. commercial ships, contractors on the ship, special teams, fisherman boats, coast guard units) • IA-CGF HBL Human Behavior Libraries (i.e. fatigue, stress, aggressiveness, trustiness) • IA-CGF NCF Non-Conventional Frameworks devoted to reproduce specific scenarios (i.e. piracy) The IA-CGF are available to support different aspects in the marine asymmetric threat simulation, such as:

Areas of contamination Case 3: 10kT Nuclear Weapon

It is evident the importance to define best solution to face these challenges by considering the area impact of such devices as well as the implication of a port attack in term of economic costs and strategic issues. The use of M&S, CGF, and Intelligent Agent CGF is critical to test new algorithms to detect suspect behaviors or to develop the requirements for new security solutions; today the use of mobile networks provides very interesting opportunity to share info quickly and easily and to develop deployable netcentric solutions, therefore the use of M&S is critical to properly design the specific configuration and operative modes.

ASYMMETRIC MODELING

THREATS:

IA Drive the General Traffic & Critical Entities In fact the use of Intelligent Agents provide the capability to create large simulation frameworks where airplanes, yachts, ships, ground entities act in consistency with their nature and within the Scenario and react dynamically to the Simulation Evolution.

ASYMETRIC

The complexity of marine scenarios is due often to the involvement of many entities that generate a very challenging framework for detecting real asymmetric threats from false alarms or uncommon behaviors; in order to face this challenge it is necessary to create models able to reproduce complex behaviors such as that ones that characterize general cargo operations, commercial traffic, pleasure boats as well as the threat tactics. By this approach it is possible to create models that support the marine asymmetric threats assessment.

IA Direct the Port and Coast Protection within its 5 Dimensional Space IA are able to direct actions of the different resources for port and coast protection, so it becomes possible to run extensive experimental campaigns by simulation for defining optimal protection solution and to assess the threats over a complex scenario; these results are achievable by testing and evaluating the effectiveness and efficiency of the all Naval Resources, including platforms, weapons, individual sensors, ground infrastructures, C2 and different information sources for protecting assets against new threats including not conventional use of civil resources.

Human Factors and Marine Simulation Most of the critical issues in generating and simulating large maritime scenarios is dealing with the necessity to model the humans factors that affect the activities of the vessels, boats, airplanes, coast infrastructures as

778

organizations, entities and operators in Emergency & Environmental Management. MESA is a modular system based on combined simulators running on PC able to export directly the results on WWW servers.

Fig.4

FLODAF As Asymmetric Data Fusion example FLODAF framework was developed by the authors as tool devoted to support engineering and performance estimation of Data Fusion architectures and algorithms; this suite includes a Scenario Generator and a Simulator for analyzing the Data Fusion performances over complex Air-Naval scenarios including surface and underwater vessels, aircrafts.

Asymmetric Data Fusion

In the following part of the paper, different Interoperable M&S solutions, developed by the authors for Marine Environment are presented as example of their potential in facing the above mentioned challenges.

ST-VP ST-VP was by Simulation Team originally as a framework to support Training in marine evnrionemtns; in fact the Interoperability of ST-VP simulators is based on HLA and guarantees in addition to traditional stand-alone training, even Concurrent Cooperative Training in complex Operations and Policies; ST-VP have a long experience in being applied within commercial ports. In fact The ST-VP includes all the different port equipment and even other marine devices and platforms; ST-VP in addition to Operator Training supports even Safety and Security Training, Procedure Definition, Equipment Design and Virtual Prototyping; among ST-VP innovative capabilities the following aspects are interesting • ST-VP is a fully containerized real-time distributed HLA Simulator reproducing Marine Environments and Ports. ST-VP is integrated within a 40’ High Cube Container ready to be used on site immediately after arrival. • ST-VP Simulator allows to operate all the different Equipment in a Virtual World by an immersive Cave (270 ° Horizontal and 150° Vertical), reproducing Sounds, Vibrations, Motion in all weather conditions • ST-VP includes a Full-Scope Simulation for Training Operations & Procedures, an Integrated Class Room, the Instructor Debriefing Room, and secondary Interoperable Simulators of all the Port Cranes and a Biomedical Module for Safety, Ergonomic and Posture Enhancement. • ST-VP World is customizable for each Port, Procedure and Equipment

PANOPEA PANOPEA is a IA-CGF NCF that use IA for reproducing a complex framework related to piracy involving several thousands of vessels, plus all related activities (i.e. intelligence, ports, special forces, contractors, helicopters, UAV, etc.) In fact PANOPEA simulation allows to model Piracy activities; a specific study on-going by using this simulator is related to the evaluation of different strategies in NEC C2 M2 (Netcentric Command and Control Maturity Models) and in quantify benefits related to guarantee C2 agility. PANOPEA reproduces military vessels, helicopters, ground base units, cargos, as well as small medium boats, fishermen and yachts traffic as well as Pirates; all these entities are driven by Intelligent Agents and apply strategies for succeeding in their specific tasks. PLACRA The Placra simulator was developed by the authors in order to reproduce the crew activities and behavior on Oil Platforms as well as on vessels; in this case the simulator takes care of reproducing operative procedures, on-board micrologistics as well as the human behavior modifiers and their impact on crew efficiency; the model consider the workload, individual and team characteristics, their history and previous experiences as well as the platform infrastructures and equipment; the simulation evolve over a scenario where regular or critical events have to be handled. MESA MESA is an integrated environment, developed by the authors, to perform simulation and risk analysis in ports and maritime sector considering the evolution of emergencies; MESA combines the simulation with GIS to support safety and security assessment plans and operations; in fact MESA is devoted to support port

An example of ST-VP federation applied to a marine security scenario is proposed in the following scheme: In fact ST-VP is able to interoperate with other simulators (virtual and constructive) as well as with real equiptent; among the others it is possible create connections with: ST_PT & ST_RS Simulators

779

(driving simulatgors), Seaports (Simulator of Terminal and Ports Security Procedures and Operations), TRAMAS (Simulation of the Logistic Network & Impact on the Town of port activities), KATRINA LIKE (regional scenario simulation reproducing a large scale crisis)

• • • • • • •

Fig.5

ST_VP FEDERATION

CONCLUSIONS



It evident that Maritime Security is part of a wide Scenario and need to be addressed by an integrated approach: due to the complexity of the framework the use of simulation is very effective and this paper proposes some of the critical methodologies and techniques to be used in this direction In addition it important to outline that Marine Asymmetric Warfare is fast evolving introducing new issues and new threats affecting more and more subjects, so it is becoming very urgent to create capabilities in defining, evaluating and optimizing solutions to face these challenges; so it becomes evident that Simulation and Cognitive Technologies are the key issues for succeeding in this goal; in fact today it is very critical to proceed in research and investigation on these domains respect the new evolving threats and to develop of New models and simulators for supporting the development of Systems, Devices and Equipment. In fact the importance of these aspects suggests that it is critical to create new capabilities in security for Maritime Scenario and to network with all international research centers operating in this context In order to succeed it is critical to develop critical assessments as well as to establish connections with Agencies, Companies and Institutions operating in this area.







• •



REFERENCES •



Alberts, D.S. (2008) “NATO NEC C2 Maturity Model Overview,” Draft for Peer Review, SAS065 Study Group, 2008.

780

Alberts, D.S. (2007) “Agility, Focus, and Convergence: The Future of Command and Control.” The Int. C2 Journal, Vol. 1, No. 1, 1-30 Alberts David S., Hayes Richard E., (2006) "Understanding Command and Control", Washington: CCRP, fig. 11, p. 35 Alberts D.S., Hayes R.E. (2003) "Power to the Edge", CCPR Publications, Washington D.C. Alberts D. S., (2002), "Information Age Transformation", revision June 2002, CCRP Publication Series Alberts David S., Gartska J.J., Stein F.P. (2000) "Net Centric Warfare", CCRP, Washington Anderson N. (2006) "Are iPods shrinking the British vocabulary?", Ars Technica On-Line Magazine, December 15 Bocca E., Pierfederici, B.E. (2007) “Intelligent agents for moving and operating Computer Generated Forces” Proceedings of SCSC, San Diego July Bruzzone A.G., Massei M., Madeo F., Tarone F. (2011) "Simulating Marine Asymmetric Scenarios for testing different C2 Maturity Levels", Proceedings of ICCRTS, Quebec, Canada, June Bruzzone A.G. Tremori A., Massei M. (2011) "Adding Smart to the Mix", Modeling Simulation & Training: The International Defense Training Journal, 3, 25-27, 2011 Bruzzone A.G., Tarone F. (2011) "Innovative Metrics And VV&A for Interoperable Simulation in NEC, Urban Disorders with Innovative C2", MISS DIPTEM Technical Report, Genoa Bruzzone A.G., Massei M. (2010) "Intelligent Agents for Modelling Country Reconstruction Operation", Proceedings of Africa MS2010, Gaborone, Botswana, September 6-8 Bruzzone A.G., Massei M.(2010) "Advantage of mobile training for complex systems", Proceedings of MAS2010, Fes, Morocco, October 13-15 Bruzzone A.G. (2010) "CGF & Data Fusion for Simulating Harbor Protection & Asymmetric Marine Scenarios", Proceedings of SIM&SEA2010, La Spezia, June 8 Bruzzone A.G., Cantice G., Morabito G., Mursia A., Sebastiani M., Tremori A. (2009) "CGF for NATO NEC C2 Maturity Model (N2C2M2) Evaluation", Proceedings of I/ITSEC2009, Orlando, November 30-December 4 Bruzzone A.G. (2008) "Human Behavior Modeling: Achievement & Challenges", Invited Speech at SIREN Workshop, Bergeggi, Italy, June 6th



• •



• •





• • • •



• • •



Bruzzone A.G., (2007) "Challenges and Opportunities for Human Behaviour Modelling in Applied Simulation", Keynote Speech at Applied Simulation and Modelling, Palma de Mallorca Bruzzone A.G., Figini F. (2004) "Modelling Human Behaviour in Chemical Facilities and Oil Platforms", Proceedings of SCSC2004, San Jose' Bruzzone A.G., Viazzo S., Massei M., BC (2004) "Modelling Human Behaviour in Industrial Facilities & Business Processes", Proc. of ASTC, Arlington, VA, April Bruzzone A.G., Massei M., Simeoni S., Carini D., B.M. (2004) "Parameter Tuning in Modelling Human Behaviours by Using Optimization Techniques", Proceedings of ESM, Magdeburg, Germany, June Bruzzone A.G., Figini F. (2004) "Modelling Human Behaviour in Chemical Facilities and Oil Platforms", Proceedings of SCSC, San Jose Bruzzone A.G., Mosca R., Simeoni S., Massei M., B.M., B.C., (2004) "A Methodology for Estimating Impact of Different Human Factor Models on Industrial Processes", Proceedings of EUROSIM, Paris, France, September Bruzzone A.G., Procacci V., B.M., B.C. (2001) "FLODAF: Fuzzy Logic Applied to a MultiSensor Data Fusion Model", Proceedings of FLODAF2001, Montreal, August 7-10 Bruzzone A.G., Rapallo S., Vio F., (1999) "MESA: Maritime Environment for Simulation Analysis", Tech.Report of ICAMES, ENSO, Bogazici University, Istanbul, May 15-21 Bruzzone A.G., Page E., Uhrmacher A. (1999) "Web-based Modelling & Simulation", SCS International, San Francisco, ISBN 1-56555-156-7 Bruzzone A.G., Giribone P. (1998) "DecisionSupport Systems and Simulation for Logistics: Moving Forward for a Distributed, Real-Time, Interactive Simulation Environment", Proc. of the Annual Simulation Symposium IEEE, Boston Bruzzone A.G., Giribone P. (1998) "Quality Service Improvement by Using Human Behaviour Simulation", Proceedings of ESM,Manchester, UK, June Bruzzone A.G. (1996) "Object Oriented Modelling to Study Individual Human Behaviour in the Work Environment: a Medical Testing Laboratory ", Proc. of WMC, San Diego, January Cantice G. (2008) "Serious Games... Serious Experimentations?", Proc. of SeriGamex, November CTA (2002) "Agents for Net-Centric Warfare and Time Critical Targets", CTA Technical Report Fogel D. (2005) "Volutionary Computation-toward a new philosophy of machine intelligence", IEEE Press series on Computational Intelligence

• • •





• • •

• •







781

Goldstein J. (2007) "Trial in Absentia Is Ordeal for Veteran Who Was Cleared by U.S. in a Killing", NY Sun, July 16 Krulak C.C. (1999) "The Strategic Corporal: Leadership in the Three Block War" Marines Magazine Ladner R., Petry F. (2005) "Net-Centric Web Approaches to Intelligence and National Security", Springer, NYC Ladner R., Warner E., Petry F., Katikaneni U., Shaw K., Gupta K., Moore P. (2009) "Web Services: Evolving Techniques in Net-Centric Operations", Proceedings of MTS/IEEE OCEANS, Liddy L. (2005) "The Strategic Corporal: Some Requirements in Training and Education", Australian Army Journal, Volume II, Number 2, 139-148 Merkuriev Y., Bruzzone A.G., Novitsky L (1998) "Modelling and Simulation within a Maritime Environment", SCS Europe, Ghent, Belgium, ISBN 1-56555-132-X Molagh J. (2009) "How Afghanistan's Little Tragedies Are Adding Up", Time, May 26 Moniz D. (2002) "Afghanistan’s Lessons Shaping New Military", USA Today, October 7 Mosca R., Viazzo S., Massei M., Simeoni S., Carini D., B.C. (2004) "Human Resource Modelling for Business Process ReEngineering",Proceedings of I3M2004, Bergeggi, Italy, October Patton M.S. (2003) "ES2: Every Soldier is a Sensor", The Washington Post, November 5 Ray D.P. (2005) "Every Soldier Is a Sensor (ES2) Simulation: Virtual Simulation Using Game Technology", Military Intelligence Professional Bulletin Reverberi A. (2006) "Human Behavior Representation & Genetic Algorithms Characteristics Fitting", Invited Presentation on Piovra Workshop, Savona, February 7 Shahbazian E., Rogova G., Weert M.J. (2009) "Harbour Protection Through Data Fusion Technologies", Series: NATO Science for Peace and Security Series C: Environmental Security, Springer Warne L., Ali I., Bopping D., Hart D., Pascoe C. (2004) "The Network Centric Warrior: The Human Dimension of Network Centric Warfare", Tech.Report DSTO, Edinburgh, Australia

ON THE SHORT PERIOD PRODUCTION PLANNING IN INDUSTRIAL PLANTS: A REAL CASE STUDY Agostino Bruzzone(a), Francesco Longo(b) ,Letizia Nicoletti (c), Rafael Diaz(d) (a)

MIS-DIPTEM, University of Genoa, Italy (b) MSC-LES, University of Calabria (c) CAL-TEK S.r.l, Italy (d) VMASC, Old Dominium University (a) [email protected]; (b)[email protected]; (c) [email protected]; (d) [email protected]

et al, 2001). The problems arising in production scheduling are notoriously very difficult and technically complex because they involve a large number of tasks and resources subject to different constraints and objectives; the complexity grows even further due to uncertainties in the manufacturing environment ( Smith, 1992). Note, in addition, that optimal allocation of the jobs to production resources over time is a combinatorial problem (Garey et al., 1976). Scheduling problems can be formulated using analytical methods like mathematical programming or network theory. In this way , for small size problems, optimal solutions can be detected but, in most cases, the assumptions required for the analytical formulation are too restrictive so the resulting mathematical model may be not able to represent with accuracy the real problem (Son et al,1999). In other words theoretical notions tend to oversimplify crucial factors of the actual production process proving that an analytic formulation and resolution is inadequate. Many research works on scheduling problems have been carried out with analytical approaches but most of them consider only one or few constraints (e.g. setups, failures, blocking, etc.) at the same time and as often as not one scheduling objective (criteria) while multiple scheduling objectives subject to several constraints have to be considered in real manufacturing systems. Also the enumerative methods and in general exact methods (usually applied when analytic procedures are not available) are prohibitive to use because of their unrealistic computing requirements (Riane et al, 2001). It is evident that the advances of theory have had a limited impact in practice but it does not mean that advances in scheduling theory have been a waste of time because they have provided interesting insights into the scheduling problem (Pinedo, 2008). An alternative approach to face this problem lies in the use of Modeling & Simulation, simulating reality by building a simulation model (Johtela et al, 1997). Modeling & Simulation allows to overcome the gap between theory and real-world scheduling problems thanks to the capability to represent real word systems and its constraints (Frantzen et al, 2011).Different

ABSTRACT Due to the increased level of competition, nowadays production systems have to keep high performances ensuring customer satisfaction, costreduction and high product quality. The features of the actual competitive scenario drive to pursue even higher levels of efficiency in companies management. In this perspective production planning, with special regard to short period production planning, plays a key role. As a matter of fact, while the long period planning aims at the evaluation of production quantities for each product, the short period planning aims at the definition of an optimal schedule to achieve even higher system performances. As well known a scheduling problem encompasses a great complexity, this kind of problem can be seen as a double allocation problem where the allocation of the jobs to production resources and the allocation of the jobs in a specific time production horizon have to be defined. The complexity grows even further considering that many interacting and variables must be taken into account simultaneously and the stochastic system behaviour cannot be neglected. This paper faces scheduling problems in a real manufacturing system proposing an approach based on genetic algorithms, dispatching rules and Modelling & Simulation. Keywords: Shop Order Scheduling, Discrete Event Simulation, Genetic algorithms, Dispatching Rules. 1. INTRODUCTION The short period production planning tackles the problem of assigning the arriving jobs to workers, machines, equipment and other resources over time. As stated in (Kiran, 1998), scheduling problems are concerned with the determination of which resources should be used and the determination of the completion and starting time for each operation of each order so that no constraint are violated and some scalar functions, measuring the effectiveness of a particular schedule, are maximized (or minimized). Getting a lower inventory level, a high plant efficiency (it means high machine and labor utilization), and respecting due dates, are some examples of scheduling criteria.(Riane

782

simulation modeling approaches taken in the literature about job-shop have been reviewed by (Ramasesh, 1990)providing a state-of-the-art survey of the simulation-based research on dynamic job shop scheduling. In the literature, there are two major approaches to deal with simulation-based scheduling problems, namely: x A simulation-based approach using dispatching rules; x A simulation-based approach using metaheuristic search algorithms. The first approach allows to put in comparison dispatching rules establishing which one performs better (Andersson et al, 2008) . Carri (1986) describes this approach as the experimentation of scheduling rules and the assessment of the effect of different rules on shop’s ability to meet delivery dates and utilize machines. Experimentation with simulation models makes it possible to compare alternative scheduling rules, test broad conjunctures about scheduling procedures and develop greater insight into the job shop operation ( Vinod and Sridharan, 2011 ). Many research works about this approach can be mentioned. Parthanadeea and Buddhakulsomsirib (2010) develop a computer simulation for canned fruit industry and conduct computational experiment on the simulation model to determine a set of appropriate dispatching rules. In Liu (1998) a two-stage simulation process has been presented: in the first stage, a number of dispatching rules are used as input parameters to generate candidate production schedules from a simulation model; in the second stage the performances of these production schedules are evaluated by another simulation model. Goyal et al. (1995) have carried out a simulation study in order to analyze the scheduling rules for a flexible manufacturing system. Different combinations of scheduling rules have been applied evaluating their effect on system performances. Huq and Huq (1995) have developed a simulation model, using a hypothetical hybrid job shop, to study the performance of different scheduling rules combinations with variations in arrival rates and processing times. Flow time, tardiness and throughput have been used as performance measures. They have found out that the rule combination performance varies with the performance criteria, and the combinations are sensitive to arrival rates and processing times. Holthaus (1997) developed new scheduling rules by the combination of well known rules, and conducted a simulation-based analysis of those rules in the dynamic job shop environment. He concluded that the new scheduling rules are quite efficient. Many other simulation studies have been carried out to evaluate the performances of dispatching rules: Holthaus and Rajendranb (1997),(Hicks and Pupong , 2006).

However, in general, this approach does not allow to find the optimal schedule. The second approach mentioned above is based on the combined use of meta-heurist optimizer with simulation and allows to detect the optimal schedule (Andersson et al, 2008). Among the meta-heuristic algorithms, genetic algorithms (GA) have been recognized as a general search strategy and an optimization method which is often useful for finding combined problems; for these reasons GA have been used with increasing frequency to address scheduling problems ( Jeong et al, 2006). The application of genetic algorithms to scheduling problem has been proposed by Bierwirth (1995), Syswerda (1991), Dorndorf and Pesch (1993), Yamada and Nakano (1992), Sakawa and Mori (1999) , Ghedjati (1999), Haibin and Wei (2001), Yun (2002), Vinod and Sridharan (2011) and many others. The joint use of genetic algorithm and simulation are further proposed in Hou and Li (1991), Rabelo et al.,(1993), Ferrolho and Crisóstomo, (2007). A comparison of these two approaches has been presented by Kim et al.,(2007) for job shop schedules of standard hydraulic cylinders and genetic algorithm were found to be better than dispatching rules (LPT, SPT, most work remaining MWKR, and least work remaining LWKR). Similar results were found out in (Sankar et al., 2003) where the results obtained with GA are compared with the results obtained using six different dispatching rules including SPT, LPT, EDD, largest batch quantity (LBT), smallest batch quantity (SBQ) and highest penalty (HP). In this study it has been found out again that the solutions generated by GA outperform the solutions obtained by using Priority Dispatching Rules. PDRs and meta-heuristic optimizer can also be jointly used with good results as shown by (Andersson et al, 2008) In this research work we present a study in which simulation is jointly used with genetic algorithms and dispatching rules to face stochastic scheduling problems in a real manufacturing system. The main goal of the present work is to provide a useful tool that can be integrated in the management system of the company and that can be profitably and efficiently used for short period production planning. The paper is structured as follows: Section 1 presents an accurate description of the system under study, Section 2 presents the steps that have been followed to built the simulation model, Section 3 deals with the verification and validation of the simulator, in section 4 the main results have been presented and finally the last section describes the main conclusions . MANIFACTURING PROCESS DESCRIPTION The project has been developed in collaboration with a small company, which produces high pressure hoses, under specific request of the company top management. 2.

783

During the initial meetings and analyzing the initial collected data it had been evident the efficiency eduction due to the short period production planning. In particular the effective production was smaller than the target production and there were continuous delays in Shop Orders (here in after S.O.s) completion that caused the decrease of the customers’ satisfaction level. So the purpose of this study is to create a decision making tool (specifically a simulator) that could be easily integrated and profitably used in the company management system to support the short period production planning. It is useful to give a brief description of the manufacturing process in order to provide a greater understanding of the steps carried out in the present research work. Each product (see figure 1) is made up by a high pressure hose, two adapters and two hydraulic fittings.

3.

MODELING & SIMULATION FOR THE MANIFACTURING PROCESS This research work faces a dynamic-stochastic scheduling problem. It is dynamic because new S.Os arrive during the scheduling horizon and the system allows the passing between jobs. Normal and priority S.Os can enter in the system. Usually normal S.Os are scheduled on a 2-weeks time window and each new S.O. enters in the last position of the queue. On the contrary, a priority S.O., depending on its priority level, can enter the 2-weeks queue in any position at any timeEach S.O. has a finite number m of operations, all the S.Os entered into the system have to be necessarily completed. The stochastic nature of the problem is due to the presence of stochastic numerical quantities. In effect set-up’s time can be considered as stochastic variables each one with a specific statistical distribution. Further, during the scheduling period, some failures can occur reducing the availability of machines. In the present work failures have been modeled by using a negative exponential distribution for both the Mean Time To Failure (MTTF) and the Mean Time To Repair (MTTR), where MTTF expresses the average time between two consecutive machine failures and MTTR expresses the average time required for repairing the machine. Once the main features of the problem and the production process have been described, the main steps of the research work can be presented. The simulation model development can be summarized as follows: x initial analysis, data collection and distribution fitting; x simulation model development; x Verification, Validation and Accreditation (VV&A); x Genetic Algorithms implementation to support Shop Order scheduling, x simulator integration in the company management system as real time decision tool for short period production planning . All the phases for the simulation model development are detailed in the following sections.

Fig. 1- Hydraulic hoses The production process is made up by 8 operations: x Preparation : all the materials, needed for each Shop Orders, are taken from the warehouse x Fittings stamp : the information required by customers are stamped on the hydraulic fittings x Cut : hydraulic hoses are cut in order to obtain the right hose length x Hose skinning: the external (internal) hose diameter is reduced (increased) in order guarantee an optimal junction between hose, adapters and fittings. x Assembly: hoses, fittings and adapters are assembled. x Junction: all the components are definitively joined x Test : hydraulic hoses are opportunely tested to check the resistance to high pressures x Final controls and packaging These operations are performed in the same order in which they are described but the cutting phase and the fittings stamp operation can take place in parallel since they involve two different components not yet assembled. Further for the cutting phase, two different machines are available: manual and automatic; these machines have different setup times and working times so different levels of productivity.

3.1.

Initial data analysis, data collection and distribution fitting The most important information were collected by means of interview and by using the company informative system. Data collection is concerned with information regarding products, working methods, short period production planning and management, actual S.O.s scheduling rules, inventory management and company informative system. In particular the collected data regard: customers, production mix, bill of materials, work shifts, process times, stocks and refurbishment times, due dates, frequency of customers requiring orders, frequency of customer orders, number of S.O. for each customers, quantity of pieces for each S.O.

784

The most important information were collected by means of interview and by using the company informative system. In particular a key role in data collection has been played by the company informative system from which a database has been extracted. The database reports information regarding final products as: operation identifying number, worker name, Shop Order identifying number , number of pieces, operation competition date, operation competition time, drawing identifying number, hose description, adapters and fittings description. In the same database are also reported information regarding final products opportunely ranked for due date and S.O. identifying number. All the stochastic variables have been analyzed in order to find out statistical distributions capable of fitting the empirical data with satisfactory accuracy. Figure 2 shows the histogram obtained putting in relation the time process observed for the junction operation with the frequency of occurrence. The same kind of histogram has been built for each operation which makes up the production process. Is then possible to find out the most suitable statistical distribution.

In effect the simulator flexibility cannot be easily achieved by using library objects: each library object should represent a specific component/part of a real system but sometime such objects do not represent the real system with satisfactory accuracy. To overcome this problem a programming code must be used for the simulator development. In the present work classes and objects have been implemented by using Simple++ , a simulation language provided by eM-Plant. In this way classes can be accessed and modified at any time and also, if needed, used in other simulation models. So the use of a programming code in developing a simulation model ensures a great accuracy and offers the possibility to change it in the future according new emerging needs; as a consequence, high level of flexibility can be achieved. Concerning the computational efficiency of the simulator and the time required for executing simulation runs, we should take into consideration how a discrete event simulation software works. In a discrete event system the state of the system changes at discrete event time points due to the flow of entities inside the system, for example at the end of an operation, at the arrival of a new shop order, etc. In other words entities with their actions change the state of the system. Usually entities are defined as classes instantiated inside the simulation model. So each entity can also have attributes in which specific information are stored. Note that the number of entities defined in a simulation model is strongly related to the computational load of a simulator: the higher is the number of entities flowing in the simulation model the higher is the computational load of the simulator. Consider that in most production processes thousands of components and products usually flow inside the system, it means thousands of entities flowing inside the simulation model and consequently a high computational load. To overcome this difficulty the approach used for developing the simulation model proposed in this paper is based on the idea to substitute the flow of entities with a flow of information opportunely stored in tables. The events generation is committed to specific objects (provided by the eM-Plant library) called event generators. Ad-hoc programmed routines manage the change of the state of the system due to the generation of an event; the information stored in the tables are updated by the programming code. By following this approach, two main advantages can be obtained: (i) a great gain in term of computational load of the simulator; (ii) reduction of the time required for executing simulation runs. Figure 3 shows an example of information stored in table for each entity (shop order) flowing into the simulator. The simulator main frame is called model. It contains 10 secondary frames (see figure 4). In particular 8 frames are built to the recreate the operations described in section 2 (Preparation, Fittings stamp, Cut, Hose skinning , Assembly, Junction, Test, Final controls and packaging) whilst the remaining 2 frames are respectively:

Fig 2: Histogram and Statistical Distribution of the Process Time for the Junction Operation 3.2. Simulation model development Without doubt the most important step of a simulation study is the modeling phase. In this research work a new approach, quite different from traditional approaches, has been adopted; in the following there is a detailed description of the architecture used during the modeling phase. The main requirement that has been taken into account was to develop a flexible and time efficient simulator. A flexible simulator is a simulator able to easily integrate new additional features over the time while a time efficient simulator is not time consuming in the execution of simulation run. So during the modeling phase we do not use the classic object oriented approach characterized by library objects and entities (that opportunely set and define the simulation model) but we propose a structural design completely based on programming code and tables to store the information.

785

The LT of the i-th S.O. is the difference between the S.O. Completion Time and the S.O. Due Date (DD), as expressed by equation 2. Finally the FR, as expressed by equation 3, is the percentage of S.Os meeting the due date. FTi

CTi  RTi

(1)

LTi

CTi  DDi

(2)

k

FRi

Fig. 3: An example of information stored in table for each entity (shop order) flowing into the simulator

¦ S.O.

i

¦ S.O.

i

i ! n

(3)

i 1

x the Production Manager (PM); x the Graphic User Interface (GUI). The PM generates the S.Os and the relative production planning, takes care of S.Os scheduling, resource allocation and inventory management. The graphic user interface allows the user to select the dispatching rule to be used for S.Os scheduling or to select S.Os scheduling based on the results of genetic algorithms.

3.3.

Simulation model Verification , Validation and Accreditation In the course of a simulation study the accuracy and the quality are not guaranteed “a priori”, for this reason the verification, validation and accreditation processes have to take place to assess the goodness of the developed simulation tool (Balci 1998). Usually a conceptual model is an abstract representation of a real system; in a simulation study the conceptual model is required to build a computerized simulation model. The verification allows to verify if the translation of the conceptual model into the computerized simulation model is accurate and correct. Furthermore the simulator has to be able to reproduce the behaviour of the real system with accuracy since it will take over from the real system for the purpose of experimentation. The validation phase is devoted to assess the accuracy of the simulation model. Accreditation is “the official certification that a model or simulation is acceptable for use for a specific purpose.” (DoD Directive 5000.59). For further details on simulation model Verification, Validation & Accreditation, refer to the American Department of Defence Directive 5000.59. There are two basic approaches for testing simulation software: static testing and dynamic testing (Fairley, 1976). In static testing the computer program is analyzed to determine if it is correct by using such techniques as structured walk-throughs, correctness proofs, and examining the structure properties of the program. (Sargent, 2000). Dynamic techniques require model execution and are intended for evaluating the model based on its execution behavior.(Balci 1997) The simulator verification has been carried out by using the Assertion Checking dynamic technique. Detailed information about this technique can be found in Adrion et al. (1982). We inserted global region and local assertion in order to check the entire model. In this way some errors, most about raw materials inventory management, were detected and corrected. The simulator validation has been carried out by using the Mean Square Pure Error analysis (MSPE). The MSPE The MSPE is a typical technique devoted to find the optimal simulation run duration that guarantees the

Fig. 4 – Simulator Main Frame Furthermore the GUI provides the user with many commands as, for instance, simulation run length, start, stop and reset buttons and a Boolean control for the random number generator to reproduce the same experiment conditions in correspondence of different operative scenarios. The dispatching rules that have been implemented in order to study the Shop Orders scheduling are the Short Production Time (SPT), the Longest Production Time (LPT), Due Date (DD). Some performances indexes have been implemented in the simulation model to evaluate the S.Os scheduling: x the average and the variance of the Flow Time (FT), x the average and the variance of the Lateness (LT), x the Fill Rate (FR). The FT of the i-th S.O., as reported in equation 1, is the difference between the S.O. Completion Time (CT) and the S.O. Release Time (RT).

786

goodness of the statistical results in output from the simulation model. Considering the stochastic distributions implemented in the simulation model we can assert that the outputs of the simulation model are subjected to an experimental error with normal distribution, N(0, V). The best estimator of Vis the mean squares error. The simulation run has to be long enough to have small values of the MSPE of the performance measures being considered. In other words, the experimental error must not “cover” the simulation results. Considering the Flow Time, we can write: n ( FTh (t )  FT (t )) (4) MSpE(t ) ¦ n 1 h 1

problems. A simulation tool allows to monitor the system performances under different S.Os scheduling but an optimization algorithm is required to improve the S.Os scheduling . Interfacing the optimization algorithm with the simulation model it is possible to find out the most suitable solution (evaluated optimizing the scalar function chosen to measure scheduling goodness). The interface between the simulation model and genetic algorithms was created through specific subroutines written using the simulation language Simple++. In this way the optimization algorithms and the simulation model jointly work for the scheduling problem resolution: the former finds out acceptable solutions while the latter validates and chooses the best solutions.

x

FTh(t), value of the Flow Time at instant of time t during the replication h; x h=1,…,n number of replications. Analogous equation can be written for the LT and the FR. The simulation run length chosen is 200 days. Such time, evaluated with four replications, assures a negligible mean squares error for the Flow Time. The same analysis for the Lateness and the Fill Rate gives lower simulation run lengths. The accreditation analysis has been carried out in the present work by monitoring the performance indexes (FT and LT). The best results in terms of mean daily flow time can be obtained using the Longest Production times scheduling rule (LPT). Taking into account that the model proposed in theory is too simplified, this result can be completely accepted even if it is in contrast with theory. Concerning the impact of the different scheduling rules on the mean daily lateness, the difference between the scheduling rules is not so remarkable.

4. SIMULATION RESULTS AND ANALYSIS The research work faces the Shop Orders scheduling problem into a real manufacturing system devoted to hydraulic hoses production. The proposed approach is based on the use of Modelling & Simulation jointly used with dispatching rules and genetic algorithms. The system performances, under different dispatching rules have been tested, as well as the guidelines obtained by using genetic algorithms. The scheduling rules (implemented in the simulator) being tested in the following analysis are: x the Shortest Production Time (SPT); x the Due Date (DD); x the Longest Production Time (LPT). The average values of the FT, LT and FR in correspondence of each scheduling rule are shown in Table 1. As it can be seen in table 1 the SPT rule guarantees the best performances in terms of Flow Time, while the DD rule allows to get the best performance in terms of lateness and Fill Rate. Table 2 reports the standard deviation values for each performance measure in correspondence of each scheduling rule.

3.4.

Genetic Algorithms implementation to support Shop Order scheduling The modeling architecture has been opportunely programmed to be interfaced with genetic algorithms. So once tested the validity of the simulation model, further implementations were carried out to introduce Genetic Algorithms (GA) as support tool for short period production planning. The use of genetic algorithms goes through three fundamental steps: x initial S.Os scheduling (proposed by the user); x setting of genetic operators and algorithms initialization x optimization. The GA was implemented as a functional part of a particular tool called optimizer. This object aims at: optimising S.Os scheduling by means of GA, testing the proposed scheduling, monitoring the manufacturing system performances by using the Flow Time, the Lateness and the Fill Rate indexes. In the following part the problem (which has to be solved) and the optimizer have been described. Simulation tool is not the only way to solve stochastic shop orders scheduling

SPT DD LPT (FT) 4,580 5,590 [days] 3,600 (LT) 1,090 2,370 Lateness [days] 1,500 (FR) 78,640 79,250 73,780 Fill Rate [%] Table 1: Average values of the Performance Measures Flow Time

SPT DD LPT (FT) 0,031 0,039 0,035 [days] (LT) 0,028 0,031 0,036 Lateness [days] (FR) 0,21 0,17 0,19 Fill Rate [%] Table 2: Standard deviation of the Performance Measures for each Scheduling Rule Flow Time

787

The same approach has been applied for the Lateness optimization. The main results have been reported in table 4 and in figure 6. As in the previous case after 25 replications the best, the average and the worst solutions converge to 0.92 days. This value is still better than the result obtained with the DD dispatching rule, the improvement is about 16 % . Finally the table 5 and the figure 7 reports the optimization results for the FR. In this last case after 25 generations the algorithm converges with an improvement of about 1,4%.

The S.Os scheduling has also been investigated by using genetic algorithms trying to minimize the FT, minimize the LT and maximize the FR. Table 3 reports the simulated FT in correspondence of each generation; for each generation are reported the best, the average and the worst FT values . After 25 replications the best, the average and the worst solutions converge to the value of 3.20 days. Note that such value is lower than best result obtained with the SPT rule (the improvement is about 9.17%). The optimization on the FT with genetic algorithms is also shown in the figure 5.

LT Generation

Generation

FT Best

FT FT Average Worst 1 8,76 9,54 9,82 2 7,00 7,76 8,43 3 6,12 6,85 7,98 4 5,91 6,60 7,83 5 5,23 6,09 7,65 6 4,76 5,73 7,60 7 4,95 5,69 6,85 8 4,63 5,53 6,82 9 4,48 5,01 6,12 10 4,32 4,99 5,96 11 4,30 4,80 5,58 12 4,10 4,73 5,51 13 4,02 4,49 5,20 14 3,73 4,17 4,55 15 3,64 3,96 4,17 16 3,64 3,92 4,25 17 3,48 3,87 4,00 18 3,48 3,29 3,38 19 3,35 3,29 3,29 20 3,27 3,29 3,29 21 3,27 3,20 3,20 22 3,27 3,20 3,20 23 3,27 3,20 3,20 24 3,27 3,20 3,20 25 3,20 3,20 3,20 Table 3: Best, Average and Worst values of Flow Time obtained by GA

Best

LT Average

LT Worst

1

3,85

4,25

5,61

2

3,78

4,03

5,49

3

3,62

3,95

4,99

4

3,46

3,82

4,67

5

3,19

3,67

4,28

6

2,92

3,46

4,05

7

2,68

3,24

3,91

8

2,20

2,97

3,72

9

2,05

2,86

3,35

10

1,90

2,51

3,27

11

1,79

2,12

3,16

12

1,65

2,07

3,06

13

1,49

2,00

2,94

14

1,26

1,83

2,29

15

1,19

1,64

2,03

16

1,13

1,48

1,82

17

1,09

1,29

1,76

18

1,02

1,14

1,58

19

0,99

1,03

1,32

20

0,95

1,00

1,20

21

0,92

0,99

1,14

22

0,92

0,95

0,99

23

0,92

0,92

0,92

24

0,92

0,92

0,92

25

0,92

0,92

0,92

Table 4: Best, Average and Worst values of Lateness obtained by GA

Fig. 6: LT Optimization with Genetic Algorithms Fig 5: FT Optimization with Genetic Algorithms

788

FR Generation

Best

5. FR Average

FR Worst

1

82,52

80,20

77,99

2

83,70

81,62

79,64

3

84,45

81,37

78,40

4

85,43

81,92

78,52

5

86,21

83,20

80,30

6

87,04

84,32

81,71

7

88,08

85,43

82,89

8

89,39

86,44

83,60

9

90,38

87,66

85,04

10

91,31

87,81

84,42

11

92,16

89,63

87,21

12

92,91

90,31

87,82

13

93,48

91,15

88,93

14

94,21

92,01

89,91

15

94,75

92,00

89,35

16

95,09

93,03

91,07

17

95,43

94,35

93,38

18

95,75

95,01

94,37

19

96,09

95,39

94,79

20

96,24

95,65

95,16

21

96,34

95,87

95,51

22

96,38

96,14

96,00

23

96,42

96,36

96,36

24

96,42

96,42

96,42

25

96,42

96,42

96,42

CONCLUSIONS

The authors implemented a discrete simulation model by using an advanced modeling approach. The simulation model was developed with the purpose to study the behaviour of different dispatching rules and the potential of genetic algorithms for the S.Os scheduling within a manufacturing system devoted to produce hydraulic hoses. The simulator architecture is totally different from the traditional modeling approach proposed by the commercial discrete event simulation packages. Such architecture is completely based on a different modeling approach : x all the objects have been modeled by means of code x all the information have been stored in table; These features allow to get high flexibility in terms of future changes and new tools implementation (for example genetic algorithms or neural network). The behavior of three different scheduling rules was analyzed in terms of Flow Time, Lateness and Fill Rate. In addition, Genetic algorithms were also used to perform three different optimizations: FT, LT and FR . Comparing the results obtained in these two steps of the research work it was found out that the genetic algorithms are capable of finding better shop orders scheduling improving the results obtained by using the classical scheduling rule. Further, thanks to the high computational efficiency, the simulator has the potentials to be used real-time for short period production planning. In conclusion, the approach proposed in this case study during the modeling phase has been useful for creating a decision and problem solving tool that can be profitably used by the integration in the company informative system and used real-time to support stochastic S.O.s scheduling. REFERENCES Adrion, W.R., Branstad, M.A., & Cherniavsky, J.C., Validation, verification, and testing of computer software, Computing Surveys, 14 (2), pp. 159– 192, 1982. Andersson M, Ng AHC, Grimm H. , Proceedings of 2008 winter simulation conference; 2008. p.200411. Balci O. Verification, Validation And Accreditation Of Simulation Models. Proceedings of the 1997 Winter Simulation Conference ed. S. Andradóttir, K. J. Healy, D. H. Withers, and B. L. Nelson Balci, O., Verification, Validation and Testing. In Handbook of Simulation, edited by J. Banks, pp. 335-393, New York: Wiley Interscience, 1998 Bierwirth C., A generalized permutation approach to job shop scheduling with genetic algorithms. OR Spektrum 17(213):87–92, 1995 Carri A., Simulation of Manifacturing Systems (Wiley, New York, 1986) Castilla i, Longo F. (2010). Modelling and Simulation Methodologies, Techniques and Applications: a

Table 5: Best, Average and Worst values of Fill Rate obtained by GA

Fig. 7: Fill Rate Optimization with GA

789

State of the Art Overview. INTERNATIONAL JOURNAL OF SIMULATION & PROCESS MODELLING, vol. 6(1); p. 1-6, ISSN: 1740-2123 Cimino A., Longo F., Mirabelli G., Papoff E. (2008). Shop orders scheduling: dispatching rules and genetic algorithms based approaches. In: Proceedings of the European Modeling & Simulation Symposium. Campora S. Giovanni (CS), Italy, 17-19 September, vol. I, p. 817-823, ISBN/ISSN: 978-88-903724-0-7. Curcio D, Longo F. (2009). Inventory and Internal Logistics Management as Critical Factors Affecting the Supply Chain Performances. International Journal of Simulation & Process Modelling, vol. 5(4); p. 278-288, ISSN: 1740-2123 Dorndorf U., Pesch E., Combing genetic and local search for solving the job shop scheduling problem. APMOD93 Proc, pp 142–149,1993 Fairley, R. E., Dynamic testing of simulation software, Proc. of the 1976 Summer Computer Simulation Conf., Washington, D.C., 40–46. Ferrolho A., Crisóstomo M., Optimization of Genetic Operators for Scheduling Problems, Journal of Advanced Computational Intelligence and Intelligent Informatics, Vol.11, No.9 pp. 10921098, 2007 Frantzen M., Amos H. C. Ng, P. Moore, A simulationbased scheduling system for real-time optimization and decision making, Robotics and ComputerIntegrated Manufacturing 27(2011)696–705 Garey M.R, Johnson D.S., Sethi R., The complexity of flowshop and jobshop scheduling, Mathematics of Operations Research 1 (1976) 117}129. Ghedjati F., Genetic algorithms for the job-shop scheduling problem with unrelated parallel constraints: heuristic mixing method machines and precedence. Comput Ind Eng 37(1-2):39–42, 1999 Goyal S. K., Mehta K., Kodkodali R., Deshmukh S.G., Simulation for analysis of scheduling rules for a flexible manufacturing system, Integrated Manufacturing Systems[ING], 6(5), pp. 21–26, 1995. Haibin Y, Wei L., Neural network and genetic algorithm-based hybrid approach to expanded jobshop scheduling. Comput Ind Eng 39:337–356, 2001 Hicks C., Pupong P., Dispatching rules for production scheduling in the capital goods industry, Production Economics 104 (2006) 154–163 Holthaus O., Design of efficient job shop scheduling rules, Computers and Industrial Engineering [CIE], 33(1,2), pp. 249– 252, 1997. Holthaus O., Rajendranb C. Efficient dispatching rules for scheduling in a job shop, Production Economics 48 (997) 87-105 Hou, E.S.H., Li, H.-Y., Task scheduling for flexible manufacturing systems based on genetic algorithms, IEEE International Conference on Systems, Man, and Cybernetics, 1991. 'Decision

Aiding for Complex Systems, Conference Proceedings.,vol-1,pp- 397 – 402, 1991 Huq F., Huq Z., The sensitivity of rule combinations for scheduling in a hybrid job shop, International Journal of Operations and Production Management [IJO], 15(3), pp. 59–75, 1995. Jeong S. J., Lim S. J., Kim K. S., Hybrid approach to production scheduling using genetic algorithm and simulation, Adv Manuf Technol (2006) 28: 129 136 Johtela T., Smed J., Johnsson M., Lehtinen R.and O. Nevailainen O., Supporting production planning by production process simulation, Computer Integrated Manifacturing Systems 10 (1997) 193}203 Kim I, Watada J, Shigaki T., A comparison of dispatching rules and genetic algorithms for job shop schedules of standard hydraulic cylinders. Soft Computing 2007;12:121–8. Kiran AS., Simulation and scheduling. Banks J, editor. Handbook of Simulation: Principles, Methodology, Advances, Applications, and Practice. New York: John Wiley and Sons, Inc; 1998. p. 677–717. Liu K. C., Dispatching Rules For Stochastic Finite Capacity Scheduling, Computers ind. Engng Vol. 35, Nos 1-2, pp. 113-116, 1998 Longo F. (2010). Design And Integration Of The Containers Inspection Activities In The Container Terminal Operations. International Journal of Production Economics, vol. 125(2); p. 272-283, ISSN: 0925-5273, doi: 10.1016/j.ijpe.2010.01.026 Longo F., Mirabelli G. (2008). An Advanced Supply Chain Management Tool Based on Modeling & Simulation. Computers & Industrial Engineering, vol. 54/3; p. 570-588, ISSN: 0360-8352, doi: 10.1016/j.cie.2007.09.008 Longo F., G. Mirabelli, E. Papoff (2006). Modeling, Analysis & Simulation of Tubes Manufacturing Process and Industrial Operations Controls. In: Proceedings of the Summer Computer Simulation Conference. July 30th – August 03rd 2006, Calgary Canada, SAN DIEGO, CA: SCS, vol. 38, p. 54-59 Parthanadeea P., Buddhakulsomsirib J., Simulation modeling and analysis for production scheduling using real-time dispatching rules: A case study in canned fruit industry, Computers and Electronics in Agriculture 70 (2010) 245–255 Pinedo M., Scheduling: theory, algorithms, andsystems,.3rdEd..NewYork, NY: PrenticeHall;2008. Rabelo L. et al., Intelligent Scheduling For Flexible Manufacturing Systems, 1993 IEEE International Conference on Robotics and Automation. Proceedings., vol.3, pp-810 – 815 Ramasesh R., Dynamic job shop scheduling: A survey of simulation research, Omega, Volume 18, Issue 1, 1990, Pages 43-57 Riane F., Artiba A., Iassinovski S. An integrated production planning and scheduling system for

790

hybrid flowshop organizations, Production Economics 74 (2001) 33}48 Sakawa M, Mori T., An efficient genetic algorithm for job-shop scheduling problems with fuzzy processing time and fuzzy due date. Comput Ind Eng 36(2):325–341, 1999 Sankar, S. S., Ponnanbalam, S. G., Rajendram C., A multiobjective genetic algorithm for scheduling a flexible manufacturing system. International Journal in Advanced Manufacturing Technologies 22:229– 236, 2003 Sargent R. G. Verification, Validation, And Accreditation Of Simulation Models. Proceedings of the 2000 Winter Simulation Conference J. A. Joines, R. R. Barton, K. Kang, and P. A. Fishwick, eds. Smith S.F., Knowledge-based production management: Approaches, results and prospects, Production Planning & Control 3 (4) (1992) 350}380. Son Y. J., Rodríguez-Rivera H., and Wysk R. A., A multi-pass simulation-based, real-time Scheduling and shop floor control system, Transactions of the Society for Computer Simulation International modeling and simulation in manufacturing Volume 16 Issue 4, Dec.1.(1999) Syswerda G., Schedule optimization using genetic algorithms. In: Davis L (ed) Handbook of genetic algorithms. Van Nostrand Reinhold, New York, pp 332–449, 1991 Vinod V., Sridharan R., Simulation modeling and analysis of due-date assignment methods and scheduling decision rules in a dynamic job shop production system, Production Economics 129 (2011) 127–146 Yamada T, Nakano R., A genetic algorithm applicable to large scale job shop problems, vol 2. Parallel problem solving from nature. North Holland Publishers, Amsterdam, pp 281–290, 1992 Yun Y. S., Genetic algorithm with fuzzy logic controller for preemptive and non-preemptive jobshop scheduling problems. Comput Ind Eng 43:623–644, 2002

791

Carraro 44, 465 Cartlidge 299 Castellanos 642, 654 Castro 384 Cereda 768 Cesarotti 423 Cetinkaya 709 Ceylan 278 Chaczko 703 Chai 508, 715 Choi 308 Chouhal 11 Coates 470 Codetta-Raiteri 545 Crespo Pereira 400, 626 Crouch 470 Dabrowski 659 D'Ambrosio 696 Davendra 593 De Nicola 201 De Prada 211 del Rio Vilas 400, 626 Dello Stritto 325 Di Gregorio 696 Diaz 782 Dorfer 433 Dowman 363 Dreija 62 Dreiseitl 176 Drouin 363 Dupont 241 EL Boukili 144 Eldemir 535 Elizondo Cortes 187 Faschang 433 Federici 351 Ferguson 465 Fernandez de Miguel 635 Fiebig 181 Filippone 696 Fischer 448 Flores 150 Fowler 264, 363 Franz 195 Friðgeirsson 44, 50, 87 García 648, 654 Gardeshi 17

Authors’ Index Acebes 211 Adegoke 675 Affenzeller 448, 454 Aguilar 410 Aguirre 384 Ahmed 289 Aizstrauts 62 Altmann 272 Aquilar Chinea 62 Arhip 380 Ávila 654 Azadeh 562 Azimi 17 Babarada 371, 380 Backfrieder 100, 111 Bagnaro 44 Banks 740 Bao 105 Barbey 25 Battista 325 Belmabrouk 375 Bendella 375 Benhari 220 Benkhedda 375 Bizub 78 Blanco 619 Blanco Fernandez 635 Blank 433 Bocca 755 Bogdan 722 Bohlmann 747 Bossomaier 314 Boudraa B. 460 Boudraa M. 460 Bracale 44 Brandstätter-Müller 488 Brandt 78 Bruzzone 755, 768, 775, 782 Bubevski 555 Burmeister 87 Cafaro 384 Callero 410 Capocchi 351 Carneros 72 792

Gargiulo 44, 50, 87, 465 Ghaemmohamadi 562 Ginters 62 Giordano 325 Giuiusa 423 Göbel 137 Gray 264 Griffin 264 Grundspenkis 62 Gunal 278 Guntinas-Lichius 87 Hamad 229 Hamadou 675 Han 308 Harrè 314 Haueisen 50 Hawe 470 Helgason B. 465 Helgason T.44, 465 Helm 272 Hernandez-Romero 690 Hoare 333 Hoess 478 Hölzlwimmer 488 Huerta Barrientos 187 Hunt 659 Iannone 325 Ighoroje 669 Ingvarsson 44, 465 Introna 423 Jacak 345, 454 Jasek 593 Ji 340 Jimenez-Macias 580, 587, 613 Kang 308 Karakaya 535 Kebabla 5 Kern T. 433 Kern H. 44, 465 Kesserwan 516 Khadem Geraili 220 Kim 319 Klancar 118 Klauke 747 Klempous 703 Klinger 747 Klingner 87

Kolstad 247 Kommenda 454 Kopp 264 Krieg 478 Kronberger 448, 454 Krzesinski 137 Kulczycki 488 Lamas Rodriguez 400 Latorre-Biel 580, 587, 613 Lauberte 62 Lazarova-Molnar 516, 526 Lecca 36 Lee D. K. 308 Lee J.O. 541 Lee K.S. 308 Lee T.E. 319 Lee Y.J. 308 Legato 93 León Samaniego 605 Leventhal 363 Li B.H. 508, 715 Li Q. 502 Lin 508 Lipovszki 394 Lirk 488 Liu 340 Longo 775, 782 López 72 Lowenstein 363 Lulu 127 Luo 105, 502 Madeo 768, 775 Mahdaoui 11 Maïga 669 Maio 496 Mandl 44 Manzo 201 Martínez Camara 605 Massei 755, 775 Mayr H 195 Mayr W. 44, 465 Mazaeda 211 Mazza 93 Medina-Marin 690 Mendez 384 Merazi Meksen 460 Merino 211 793

Merkuryev 62 Merzouki 241 Mizouni 516 Molnar 394 Mondal 363 Morvan 241 Mouss N. 5 Mouss L. H. 5, 11 Mouss M.D. 11 Mujica 357, 734 Music 118, 681 Neumann J. 478 Neumann G. 568 Nicoletti 782 Niculiu M. 254 Niculiu T. 254 Nieto de Almeida 626 Nikodem J. 703 Nikodem M. 703 Nissen 235 Novak 768 Novelli 599 Novitsky 62 Núñez 72 Onggo 333 Oplatkova 593 Or 439 Orsini 264 Osl 176 Özbaú 439 Özlem 439 Ozmen 55 Page 137 Pan 340 Parsapour 488 Perea 150 Pérez de la Parte 605, 619, 635 Pérez V. 654 Perez-Lechuga 690 Perna 696 Petrovic 722 Petz 433 Pfeifer 272 Piera 357, 734 Popa 1 Proell 345 Ramanan 283

Ramon 50 Rarità 201 Ravariu 371, 380 Rego Monteil 400, 626 Ren 105, 502 Reynisson 465 Rios Prado 400, 626 Rodriguez A. 211 Rodriguez D. R. 619 Rongo 696 Rust 709 Sabuncuoglu 283 Sáenz-Díez Muro 605 Saetta 166 Saft 235 Sakne 62 Schiraldi 325 Schoenberg 78 Schuler 272 Seck 709 Seck-Tuoh-Mora 690 Senkerik 593 Seo 541 Shah 289 Shen 340 Shibata 728 Silva 496 Simon-Marmolejo 690 Sindicic 722 Sodja 574 Sokolowski 740 Somolinos 72 Soyez 241 Spataro 696 Spingola 696 Sriram 299 Stekel 454 Strasser 272 Suk 541 Szczerbicka 747 Tao 105 Tarone 775 Tarvid 158 Tiacci 166 Togo 675 Toma 351 Traoré 669, 675 794

Tremori 775 Uangpairoj 728 Ulbrich 478 Ulgen 30 Unnþórsson 465 Usher 247 Viamonte 496 Volk 87 Wagner 448, 454 Wellens 642, 648, 654 Wenzler 709 Williams 30 Wilson 470 Winkler 433, 448, 454 Woda 703 XU 418 Yang 715 Yilmaz 55 Zelinka 593 Zhang L. 105, 502 Zhang Y. 502 Zhongjie 127 Zottolo 30 Zupancic 574 Zwettler 100, 111

795