! 2004 IEEE InternationalConference on IndustrialTechnology (ICIT)
Simulating intelligent routing in Flexible Manufacturing Systems using NetLogo Yves Sallez, Thieny Berger, Christian Tahon Equipe Syst mes de Production Laboratoire d Automatique et de M canique Industrielles et Humaines W C N R S 8530, Le mont Nouy, F-59313Valenciennes cedex Email :
[email protected] Abstract- This paper describes the advantages of using the NetLogo software environment to simulate an intelligent routing system in Flexible Manufacturing Systems (FMS). T h e first part describes the pheromone-based stigmergy concept. The second part presents the NetLogo simulation environment, and the third part is dedicated to a simulation of a flexible assembly cell (in the Valenciennes s AipPrimeca pole).
Key words: routing, Flexible Manufacturing Systems, simulation, stigmergy, NetLogo
2.1 Stigmergy concept In the real world, three basic operations are associated with chemical pheromones : information fusion, information removal and local information distribution. In the first category, deposits from individual entities are aggregated to allow the easy fusion of information. In the second, pheromone evaporation dver time removes obsolete or inconsistent information. In the last, local information is provided according to the diffusion of chemicals only in the immediate neighborhood. ’
I Introduction Today s mass production strategies are unable to cope with the current needs of the manufacturing industry, which must deal with continually changing and increasingly complex product requirements as well as mounting pressure to decrease costs. To meet this challenge, Flexible Manufacturing Systems (FMS) must become more robust, scalable, reconfigurable, dynamic, adaptable and even more flexible. Many international research projects have focused on designing heterarchical (non-hierarchical) architectures, and these new control systems play a prominent role in the field of FMS research [ 13. One interesting approach to such decentralized control is inspired by a biological phenomenon called stigmergy, which signifies insects use of chemicals called pheromones to organize group activity. Briefly, foraging ants lay down chemical trails, and any ant following a useful path adds its odor to the traiI, reinforcing it for future use. This phenomenon allows indirect communication between creatures through the sensing and modifying of the loca1 environment, and this communication determines the creature s behavior. The history of stigmergy in the context of social insects has been described in delail by Theraulaz and Bonabeau [Z]. Section 2 of this article first describes the different pheromone characteristics and the concepts upon which our stigmergic approach is based. Then, some interesting insect-based methods are described for different manufacturing applications. In section 3, the NetLogo platform is presented. Section 4 describes a NetLogo simulation of a flexible assembly cell, located in the Valenciennes’s Aip-Primeca pole. The results of the simulation are given, and the some of the advantages and potential disadvantages of the NetLogo platform are outlined. 0-7803-8662-0/04/$20.00 02004 IEEE
2 Stigmergic approach
In all of these operations, the pheromone field displays several characteristics : - independence : The sender of a pheromone message does not know the identity of the potential receiver and does not wait for any acknowledgement. This characteristic makes pheromone use very useful for communication within large populations of simple entities. - local management : Because diffusion falls off rapidly with distance, interactions through pheromones remain local, thus avoiding the need for centralized interaction management. - dynamism : Continuous reinforcement and evaporation respectively integrate new information and deIete obsolete information. In applications like dynamic product routing, these characteristics provide two main advantages : robustness and adaptability to the environment.
2 2 Applications in F M S The first experiments on the industrial use of stigmergy were conducted by Deneubourg [3] in the early 198O’s, using simulated “ant-like robots”. Many researchers (Ferber [4] , Arhn [5] , Dorigo and Colombetti [ 6 ] ) have applied this concept to their studies of robot collectives and the resolution of optimization problems (Traveling Salesman Problems, Network Routing for telecommunications and the Tntemet). Based on the ant foraging analogy, Dorigo developed the Ant Colony Optimization (ACO) metaheuristic, a population-based approach to the solution of combinatorial optimization problems [7]. The basic ACO idea is that a large number of simple artificial entities are able to build good solations to hard combinatorial optimization problems via low-level
1072
communications. The ACO-based approach can be applied to almost any scheduling problem, such as job shop scheduling and vehicle routing, for example.
nd
n a h w s the entity to reach
Pdp
pd
more rapidly than np
vnk
Researchers have also applied the stigmergy concept to specific situations in manufacturing control systems : Parunak [XI emphasizes the importance of the environment in agent systems. Information flows through the environment complement classic message-based communications between agents. Parunak focuses on the dynamics that emerge from the interactions in multi-agent systems. In Parunak s study, the environment is computational, and agents moving over a graph are used to study suppIy networks among manufacturing companies. Interactions among agents produce emergent dynamics that are analyzed using methods inspired by statistical mechanics.
[o 11,
Pdn
nkneighbor
pd!=
ih
vnk
1 i vnk
i
The best neighbor If (
the
i,
Pdi
nc is such that
Pdc
= m(f'dn),
entity receives insufficient = p d ), it chooses n, randomly.
information
Updating P coefticients : During moving phases, each entity stores data in an embedded memory. This data describes the entity s path through the nodes KZk , including crossing time. Every time an entity reaches a destination node, a fictitious entity retraces its path virtually, updating coefficients at each crossing node. A node ?lk contains :
Br ckner [9] applies the stigmergy concept to manufacturing control. He presents an extensive set of guidelines to enable the design of synthetic ecosystems. This application is supported by an agent-system approach. Different types of agents are used to model resources, part flow and control units within the context of car body routing in a Mercedes Benz paint shop. In Peeters el a1 [lo] and Hadeli el a/ [ l 11, the authors propose a pheromone-based control algorithm with a bottomup design. Both Peeters et a1 and Br ckner based their work on the PROSA reference architecture [12]. (Those interested should consult the Mascada-WPCReport [13] for a description of both the agents and the pheromone life-cycle in the context of car body routing in a paint shop, this one at Daimler-Chrysler.)
-
a Ph matrix whose rows denote all possible destinations n d and whose columns, all existing
neighbors vnk ,
- a matrix including the mean
dk and the standard deviation dk of the time needed to go from nk to n d for all possible destinations.
Tdk is the time span needed to go frOmPlk to n d , When the fictitious entity is located at node nk ,a comparison is made between Tdk and the mean dk of previous TA. This comparison is valid only if dk is stable enough in
terms of the
dk
of the previous Tdk ,
In the end, three cases can be characterized : - when d k is stable and Tdk dk , the use of n e is
2 3 Description of the stigmergic approach
reinforced by increasing P d c , - when d k k stable and
Our approach is based on behavioral biological systems like ant colonies [3], 1141.
Each product entity has a list of services to be obtained successively from the resources (service stations) located on the nodes. These nodes are interconnected by paths that are used by moving entities to move from one node to another until they reach their destination node, where the desired service can be obtained. After choosing their destinations, the autonomous entities must choose the appropriate path. Choice of neighbor node : The entity chooses the shortest route from its current location to the destination node, according to the information provided by the neighbor nodes, in which coefficients associated with each destination node are stored. LetPdn represent the time needed to move from the current node nk to destination n d via node n u , which must bclong to thenk neighbor node set. In the following notation :
, decreasing P d c being forgotten, results in the use of - when dkis unstable, adjustments are made to stabalize it. Tdk
dk
3 NetLogo presentation A simulation of our approach was performed using NetLogo Software [15]. NetLogo is a platform for modelisation and simulation that is based on agents fmctionning in parallel. NetLogo extends MIT s Starlogo platform by extending the traditional Logo language. Developed by the Center for Connected Learning and Computer-Based Modeling at Northwestern University, NetLogo runs on any platform that supports Java. I t s mainly used to simulate natural and social phenomena. NetLogo is particularly well suited for modeling complex systems and analyzing the connection between the behavior of basic entities and the macro-level patterns that emerge from their interactions [ 161.
1073
3.1 NetLogo modeling A NetLogo model principally involves two basic elements: 1) a two-dimensional grid that models the environment, and 2) a set of agents ("turtles") with their own attributes and a number of procedures that they can perfom. Each element of' this cellular world is called ttpatch". Some &tributes are associated to each patch and can be change dynamically. For example, a pheromone evaporation process can be implemented easily by decreasing the variable modeling the amount of pheromone deposited on the patches. All independent "agents" operate concurrently in the "patch" environment. Each entity can read and modify some of the attributes attached to the patches in its vicinity. The behavioral rules defined for the turtles allow the Mle-environment interaction to be satisfactorily described. This last is very important for the simulation of the stigmergic process. The NetLogo language extends Logo to support large numbers of agents interacting concurrently. The language incorporates control flow, input-output primitives and primitives for turtles and patches. 3.2 NetLogo interface The NetLogo interface is user-friendly. The user can easily modify various parameters in the model using such tools as sliders and switches, for example, or can give instructions to the different entiiies (turtles or patches) via ?nonitorH windows. The user can also track the evolution of the variables using graphics tools, such as line graphs, plots and monitors. Commandbutton
1
Switch
Fig. 1 is a typical interface, from the model ANTS in the NetLogo Library [17]. In this example, a colony of ants forages for food. When an ant finds food (in one of the three blue piles), it carries the food back to the nest (in violet, at center), dropping pheromones as it moves. Then other ants follow the trail toward the food. The "evaporation-mte" and the "diffusion-rate" sliders control the dynamic of the pheromones trails. The plot describc the consumption of the food sources.
4 A routing simulation for FMS 4.1 AIP FMS cell description
The simulation was Derformed in the context of a flexit assembly cell at the Valenciennes AIP-Primeca pole. This cell is composed of seven workstations i placed around a flexible conveying system (see Fig. 2). This conveying system ensures a flexible flow of pallets to each workstation. At the beginning, plates are loaded on shuttles by workstation W7. Shuttles reach the various assembly stations according to a predefined assembly process plan. For each station, a robot is able to assemble three kinds of components, accessible in different kinds of storage areas (structured pallets, random position storage or flexible feeder). Every component may be obtained in two different places. This redundancy introduces flexibility into the system, which is particularly useful in case of robot malfunction.
I
Fig. 1: Example of NetLogo interface 1074
1
,
6 axis Robot
Manual LoadingNnloading
Manual correction
Inspection
Robot
Fig. 2: Schematicview of the flexible cell The conveying system is based on the Montrac concept of the Montech society IlX]. Montrac is a monorail transport system that uses self-propelled shuttles to transport materials on tracks (Fig. 3a). Transfer gates are used to connect the different tracks. Shuttles can change direction at each transfer gate (Fig. 3b). Each shuttle is individually controlled and equipped with a collisionavoidance optical sensor.
4.2 Simulation results The preliminary results clearly illustrate the overall adaptability of the system when faced with small perturbations. Fig. 4 displays several images obtained with a NetLogo simulator. Only the right part of the cell is shown and entities are indicated by smaIl arrows. In this example, Nland NI1 are respectively the departure and destination nodes. The simulation can be divided in two periods :
- During the first period (the “teaching” phase), entities travel indifferently along the different paths of the network. (Fig. 4a). At the beginning, the P coefficients are all equal on the different nodes. According to the principle of stigmergy, the optimal path Po (N1 -N2-NlO-N11) emerges through reinforcement (Fig. 4b).
Fig. 3a: Montrac shuttle
- The second period (the “exploitation” phase) begins when the P coefficients are stable enough. This occurs when all the paths have been explored according the different destination nodes Faced with perturbations affecting the fluidity of the path N2-Nl0, the entities travelling on this path have poor performances, and the appeal of this path decreases. In a classic display of the natural routing reconfiguration of the stigmergic approach, the other non-optimal path ? P o (NlN2-N3-N5-N7-N9-NIO-N11) then becomes more appealing (Fig. 4c).
Fig. 3b: Transfer gate
Our simulation results also highlight a frequent problem in ACO algorithms: stagnation. This occurs when the routing network reaches its convergence. An optimal path Po is chosen by all ants, which recursively reinforces the preference for Po.
1075
I
In their survey of ACO systems, Sim and Sun [19] introduce some software mechanisms to solve this problem: EvaDoration : This mechanism is widely used in many studies [8], [9]. Evaporation prevents pheromone concentration on certain paths and allows the exploration of new routing solutions. This approach is inspired by real ants colonies. Aginq :With the aging mechanism, an entity deposits less and less pheromone during its travel from node to node. This solution (often used in conjunction with evaporation) is based on the fact that "old" entities are less successful in finding optimal paths.
Fig. 4a: Teaching phase Limiting Dheromone auantity : This mechanism limits the amount of pheromone on every path. This upper limit prevents the generation of a path that is "too dominant". 4 3 NetLogo evaluation
NetLogo provides a simple and efficient scripting language to control the entities and their interactions with the environment. It is very easy to learn and is an ideal language for the rapid development o f basic simulations and the fast-prototyping of some concepts. Other advantages are worthy of note : Models can be run as applets in a web browser. - The NetLogo website provides a good tutorial and a large quantity of documentation (over 140 example models included). - Notable recent enhancements to NetLogo include improved data exchange with other applications.
Fig. 4b: Optimal path
Nevertheless, there are some limitations : - NetLogo doesn t provide any object-oriented data encapsulation tools, which limits entity structuration and complex systems development. - Advanced direct communication primitives between agents are unavailable, which reduces NetLogo s usefulness in applications where protocols are commonly used between entities (such as contract nets for intelligent manufacturing systems). Fig. 4c: Reconfiguration This reinforcement has two consequences :
5 Conclusion
- congestion of the optimal path,
- an
important decrease in the likelihood of another path being chosen
In our simulation, frequent perturbations on the path N2-NI 0 provoke it decrease in the corresponding coefficient P on the node N2, with the result that the path PNo is favored. Little by little, the optimal path Po is "forgotten".
The stigmergic approach is an original answer to routing problems in FMS. Applied in many research fields (robotics, network routing ), stigmergy offers robustness and adaptability to evolving environments. These properties stem directly from the virtual pheromone and have been highlighted by. our simulation results in the NetLogo environment.
1076
'
The NetLogo platform offers many advantages for simulating a stigmergic approach. NetLogo interfaces are user-friendly, and this platform is ideal for fastprototyping. A simple and efficient scripting language allows entities to be controlled and their interactions with the environment to be described. Still, certain drawbacks (Iack of data encapsulation and absence of direct communication tools) could hamper the development of more complex simulations. In this context, perspectives for research include studying the emergence of behaviors in large flexible manufacturing cells where entities (parts, resources, operators ) can communicate directly (by message transmission) or indirectly (by stigmergy). The concept of bio-inspired systems, such as the bionic manufacturing inwoduced by Vaario and Ueda [20], would seem to be an interesting avenue to explore. In this study, the authors use local attraction fields to direct transporters carrying jobs to particular resources. Dynamic scheduling emerges from the interactions between the different entities. For this type of application, the previously mentioned limitations of NetLogo could become a problem. Luckily, several more advanced, multi-agents platforms, such as Swarm or Repast [ 161, are available if necessary.
References P. Pujo and J. P. Kieffer, M thodes du pilotage des de production, Trait Productique), H e m s Lavoisier, 2002.
syst mes
TC2
( s rie
S . Br ckner, "Return from the Ant synthetic ecosystems for manufacturing control", Ph.D. Thesis Wumboldt-University of Berlin, June 2000.
P.Peeters, H. Van Bmssel, P. Valckenaers, J. Wyns, L. Bongaerts, T. Heikkil and M. Kollingbaum, "Pheromone based emergent shop floor control system for flexible flow shopstt, in Proc. International orkshop on Emergent Synthesis (I ES'99), Kobe, Japan, Dec. 6-7, 1999. Hadeli, P. Valckenaers, M. Kollingbaum and h. Van Brussel, "Multi-agent ccordination and control using stigmergy", Computers in industry, Elsevier Science, vol. 53, January 2004, pp. 75-96.
5. Wyns, "Reference architecture for Holonic Manufacturing Systems - the key to support evolution and reconfiguration", Ph.D. Thesis, K.U. Leuven, 1 999. Mascada-WPCReport : ACA (Autonomous Cooperating Agents) Framework for Manufacturing Control Systems, 1999. G. Di Car0 and M. Dorigo, "AntNet : Distributed Stigmergic Control for Communications Networks". Journal of Intelligence Research, vol. 9, 1998, pp. 3 17-365.
NetLogo website :httu://ccl.northwestern.edu/netlo~o/
G. Theraulaz and E. Bonabeau, "A brief history of stigmergy", Journal of ArtiJiciul ye, vol. 5 (2), 1999, pp. 97-1 16. J.L. Deneubourg, J. M. Pasteels and J.C. Verhaeghe, "Probabilistic Behaviour in Ants : a Strategy of Errors 'I, Journal of Theoretical Biology, vol. 105, 1983, pp. 259-271.
J. Ferber,
H.V.D. Parunak, S. Br ckner and J. Sauter (2001). "EFUM's Approach to Fine-Grained Agents", in Proc. of the NASA JP orkshop on Radical Agent Concepts ( MC'ZU01), Greenbelt, MD, sept. 19-21, 2001.
es syst me$ multi-ugents, InterEditions,
1995. R.C. Arkin, Beha jour-based robotics, The MIT Press, 1998. M. Dorigo and M. Colombetti, Robot shaping An experiment in Beha iour Engineering, The MIT Press, 1998.
M. Dorigo, G . Di Car0 and L. M. Cambardella, "Ant algorithms for discrete optimization", Artipcial ife, vol. 5 (2), 1999, pp. 137-172.
1077
30s M.Vida1, P. Buhler , H. Goradia ,"The Past and
Future of Multiagent Systems", in Proc. AAUAS orkshop on Teaching Multi-agent Systems, NewYork, USA, July 19-23,2004. U. Wilensky, "NetLogo Ants model", 1998 httD://ccl.northwestem.edu/nctlorro/models/Ants, Center for Connected Learning and Computer-Based Modeling, Northwestern University, Evanston, IL. Montech : Montrac conveying system (2003) : http:iiwww.montech.cWmontrac/content/
K.W. Sim and W.H. Sun, "Ant Colony Optimization for Routing and Load-Balancing: Survey and New Directions", IEEE Trans. On Systems, Man and Cybernetics, vol. 3, N 5, Sept. 2003, pp. 560-572. J. Vaario and K. Ueda, "An emergent modeling method for dynamic scheduling, Journal of Intelligence Manufacturing, vol. 9, 1998, pp. 129140.
