ANTS in the Supply Chain - CiteSeerX

The original version of this paper was presented at the Workshop on Agent based Decision Support for Managing the Internet-Enabled Supply Chain, Agents 99, Seattle, WA, 1 May 1999

ANTS in the Supply Chain John A. Sauter

H. Van Dyke Parunak

(Corresponding Author) ERIM PO Box 134001 Ann Arbor, MI 48113 USA +1 734 623-2513 (v) +1 734 623-2501 (f)

ERIM PO Box 134001 Ann Arbor, MI 48113 USA +1 734 623-2509 (v) +1 734 623-2501 (f)

[email protected]

[email protected]

Abstract Computer agent based systems provide a natural mechanism to reflect real world human agent based systems. Supply chains are networks of corporations involving multiple human agents in connected but disparate processes. We investigate some ways that computer agent-based systems can assist and supplant human-based interaction and decision making in a supply chain without the need for unwieldy centralized or top-down management schemes. The Agent Network for Task Scheduling (ANTS) architecture uses techniques inspired by both human institutions and insect colonies. In ANTS large populations of simple agents exhibit robust behavior in scheduling supply chains. We describe a new mechanism called least commitment scheduling that defers decisions on process sequences until the last possible moment. A Densitybased Emergent Scheduling Kernel (DESK) uses probabilistic committed capacity profiles of resources over time to provide a surface over which the agents can wander looking for opportunities to optimize. 1. Introduction Supply Chain Management is a common theme in today’s literature. After addressing their own scheduling and material problems OEMS are looking for ways to fix the same problems in their suppliers. In the automotive industry most of the value-added comes from the second and third tier suppliers. If the inefficiencies are not addressed in the entire system, solving the problem at the OEM level simply moves the inventory and material problems to another level in the chain. Supply chains exhibit several problems: • Excessive schedule variation experienced by sub-tier suppliers • Similar capacity bottlenecks exist at multiple suppliers • Inventory or WIP levels often deviate from expectations Many supply chain problems are, in fact, systemic problems. These system problems are unintended consequences of the very way that OEMs and their suppliers do business, and are often difficult to fix. The shortcomings in today’s supply chain management solutions are becoming painfully evident. Early MRP software products were too limited and too brittle in the face of the dynamical environment of the factory floor. That led to MRP II and then Enterprise Resource Planning (ERP) as software vendors desperately tried to keep adding functionality to meet the problems of managing a complex operation. Now the software itself has become so complex that successful implementations require a multi-million dollar multi-year effort. Scaling these complex software systems up to handle the greater complexity of an entire supply chain (that often includes small manufacturers with little or no software staff) is unthinkable. Industry

05/07/99

Copyright © 1999, ERIM (MI), All Rights Reserved.

Page 1

Sauter et al., “ANTS in the Supply Chain”

needs supply chain management software that can handle complex dynamical systems while being much simpler to construct and manage. Small-grained, agent-based systems offer a promising alternative to the monolithic software modules being developed today. A small-grained agent is a simple agent that responds to its environment using simple rules or interacts directly with other agents through predetermined protocols. The rules and interactions that govern insect colonies inspire their design [8]. Section 2 briefly describes the problem we are addressing and the requirements we identify. Section 3 defines a set of agents that satisfy the requirements and describes how least commitment scheduling works. Section 4 offers some conclusions. 2. Supply Chain Management Requirements A manufacturing enterprise is measured by the cost of the goods it produces, their quality, and the timing of their availability relative to the customer’s need. The task of supply chain management is to deploy resources across a supply chain to produce high-quality goods as inexpensively as possible and when the customer wants them. It governs decisions such as which suppliers should be used for which products, in what order products should be manufactured, when new jobs should be started, when new orders should be placed, and what level of inventory should be carried. The problem is the subject of extensive research in the industrial engineering and operations research community. The specific requirements that were identified on the basis of detailed discussions with manufacturing staff at various commercial manufacturing facilities are outlined below: • Least Commitment— The customer's statement of demand (product specification, quantity, price, and delivery time) develops interactively, rather than being specified in detail at the outset. • Empowerment— Human stakeholders (including operators, manufacturing engineers, and managers) receive the information (both as-is and what-if) they need to do their jobs, with interfaces to let them control the system rather than being controlled by it. • Frequent Change— The system adapts its behavior dynamically in response to environmental changes • MRP Functionality— The aggregate behavior of the agent community subsumes functionality currently provided by MRP II systems. • Metamorphosis— The system maintains continuity between different entities that represent different stages in a common life cycle (for example, an order for a part, the part itself, and its production history). • Modality Emergence— An entity’s factory control modality emerges dynamically from its operation in the context of the rest of the system, rather than being hard-coded or the result of an explicit inference. • Uniformity— An operation at the corporate boundary interacts with external Suppliers or Customers in the same way that it does with internal ones. 3. System Design In designing a system to meet these requirements, we must develop a specific set of agents and design behaviors that support the requirements. Agents can represent various functions or they can represent things. The first step is to decide what should be agented.

05/07/99


Page 2


Input Parts

Output Parts

Input Parts

Resource Flow Axis

3.1 The Agent Architecture Previous research on agent-based factory control and scheduling (including our own) differs widely on what is represented as an agent: levels in a hierarchical decomposition of the factory [2, 7, 12], Resources [1, 4, 10, 11], or Parts [3, 6]. In ANTS the agents are things: they represent the elements in the supply chain and within the factory. We identify the following agent classes as described in [9]: Unit Process Broker (UPB)— Represents a single step in the process plan. It negotiates for the raw material and the resources it requires to perform its step in the overall plan. An instantiation of a UPB (e.g., the specific instance of heat treating part 118a performed in Oven 18 at 14:32 3 May 1999) is called an Operation. Resource— The “tools” that are needed to perform an Operation, including machines, operators, material handling devices, energy, tooling, fixtures, gauges, part programs, and documentation. Resources are characterized by such things as maintenance schedules, availability, and cost of use. They instantiate a transient agent, called an Engagement, that represents a commitment to a single Operation to perform a unit process during some time interval. Part Broker (PB)— Represents the inputs (Materials) and outputs (Products) for a UPB. There may be more than one input (e.g., assembly) or output (e.g., sawing up bar stock). PBs arrange material handling operations such as transport and storage (also represented by Resource agents) between UPBs. Material handling is just like a unit process that changes a part’s age and location. However, these changes do not alter the part functionally, and the part number does not change across a material handling operation (as it does across other unit processes). Customer— The linguistic beneficiary; the one who benefits from the execution of the work. The Customer initiates the Request for Bids (RFBs) within a factory for an order. Within the factory the Customer appears as another UPB. Externally, the Customer represents the interface with the factory’s customer. Supplier— Another variety of beneficiary, this time the one from whom the input material is purchased. The Supplier appears as a UPB to the other agents in a factory and represents the interface to the factory’s suppliers. The basic dynamic of manufacturing is that Parts move through a network of UPBs and PBs. Each UPB acquires one or more input parts (materials) from PBs responsible for parts of the necessary types, and engages Material Material Processing Processing Handling Handling certain Resources to produce one or more Resources Resources Resources Resources output parts (products) for PBs of the Engagements for Resources appropriate types. Figure 1 outlines some of Unit Unit Part Part these relationships, emphasizing that the Process Process Broker Broker dynamics of manufacturing is the flow of both Unit Part Part Unit Process Broker Broker Process parts and engagements for Resources through Unit Part Unit Part UPBs. The schedule emerges from the Process Broker Process Broker interactions of the UPBs, PBs, and Resources Part Flow Axis as they seek to fulfill their commitments. Figure 1: ANTS Agents Supply Chain Integration— ANTS’ uniformity supports a novel approach to supply chain integration. Traditionally, supply chains are

05/07/99


Page 3


integrated by visualizing each firm as a monolithic entity with its own internal mechanisms, and constructing a special set of mechanisms to permit separate firms to interact with each other. ANTS instead decomposes each firm itself into a miniature supply chain made up of a series of producers and consumers. As a result, the interfaces between ANTS agents within a firm are the same as those between one firm and another, and integration of ANTS-based firms into a larger supply chain is immediate and transparent. 3.2 Market-Based Negotiations ANTS institutes artificial markets using a modified Contract Net protocol among the various agents, extending the mechanisms explored in [1, 5, 13, 14, 15]. In addition to enabling self-organization among agents, this economic model provides a natural way to measure and trade off the relative importance of conflicting demands at each decision point, just as a market-based price system measures the relative values of goods that would be difficult to compare in a barter system. To facilitate interactions across a supply chain, ANTS’internal currency is denominated in dollars. Further details of the negotiation protocols are provided in [9] and briefly described here. Negotiations are initiated by the Customer UPB when it issues a Request for Bid (RFB) to the PB representing the finished good desired. The request specifies the latest delivery time desired along with other dimensions of the bid detailed in Table 1. The RFB is issued using subject-based addressing. The subject is the type of part specified in the RFB. This allows any PB agent to listen to RFBs for that type of part. There could be several different kinds of PBs for that part type. An inventory PB might be able to respond to the RFB using parts within its inventory. Table 1: Some Dimensions of Customer-Supplier Interaction Dimension Definition Fuzziness Product What does the Customer asks for a red car, but would be happy with any primary Identity customer want? color; wants 3σ quality but would buy 6σ at marginally higher cost Quantity How many does he Customer wants 15, but might buy 20 to get a price break want? Price How much is he On cost-plus pricing, such as many DoD contracts, the exact price is willing to pay? not known until the product is delivered Delivery How long will he wait Tardy delivery is not uncommon. Also, a customer who initially asks Time for product? for 100 by Monday might accept 10 by Monday and 90 by Friday. Information How long will he wait A customer with a custom order who learns that a bid will require a Time for these answers week to prepare might be happy to pick a preconfigured set of before ordering? options that can be priced out of a catalog on the spot. A make-to-order PB would immediately issue an RFB to the UPBs that are listening to RFBs of that part type. UPBs can represent different alternatives to producing that part including outsourcing to various suppliers. Thus a supplier can compete seamlessly with the internal production capacity at any point in the process plan. When a UPB receives an RFB, it will issue RFBs for each of the types of processes it requires. Resources listen to RFB for all the process types they can provide. So a request for a machine tool capable of performing a certain type of process will only be read by those machine tools certified to perform that process. At the same time the UPB issues RFBs for each of the input part types that it requires to all the PBs listening to RFBs for those part types. Resources

05/07/99


Page 4


respond with their hourly rate to provide the desired processing as a function of time. The UPB then searches across all the available alternative resources to find the lowest cost time slot to use. PBs respond with their cost to provide the desired part as a function of delivery time. The UPB picks the best supplier(s) for input materials and the best resource(s) and calculates a total cost as a function of delivery time. This cost function is returned in its response to the RFB. 3.3 Least Commitment Scheduling Using DESK Supply chains are complex systems whose performance is difficult to predict. Any attempt to schedule operations across a supply chain must be able to handle the uncertainty within the system and be able to respond quickly to change. Scheduling approaches that attempt to fix the schedule ahead of time are brittle with respect to change. A least commitment approach delays decisions about fixed schedule sequences until the last possible moment. This allows the system to maintain maximum flexibility as long as the capacity allows it. In addition, the system must still be able to respond to changes even after commitments have been made. In ANTS the primary commitment is represented as an Engagement on a Resource. An Operation has one or more Engagements on one or more Resources. A Resource has a set of one or more engagements with different Operations. The set of engagements on a Resource is referred to as its dance card. ANTS uses an approach called Density-based Emergent Scheduling Kernel (DESK)1 for representing engagements. In DESK an engagement is composed of three parts: • Kernel (k) – the time required for the Operation to be performed on the Resource • Commitment Window (CW) – the interval that represents the Earliest Start Time (EST) for the Operation and the Latest Finish Time (LFT). The start of the interval is denoted CWs and the end is denoted CWe. The Resource maintains the constraint, CWe − CWs ≥ k . • Working Window (WW) – the interval that represents the current EST and LFT for the Operation. Initially this is the same as the Commitment Window, but it may become narrower over time as the Resource seeks to optimize its committed capacity. The start of the working window is denoted WWs and the end is denoted as WWe. The Resource maintains the constraint, WWe − WWs ≥ k . If we assume that the kernel can slide anywhere within its working window with equal probability we can construct a committed capacity profile for any engagement that represents the amount of the resource’s capacity committed to executing that engagement at any point in time. Figure 2 shows various capacity profiles of engagements whose kernel width varies as a percentage of the working window width.

1

05/07/99

The DESK algorithms described in this section are patent pending. Copyright © 1999, ERIM (MI), All Rights Reserved.

Page 5


1.00

1.00 0.80

0.80

0.60

0.60

0.40

0.40

0.20

0.20

0.00 0.00

0.50

1.00

1.50

2.00

0.00 0.00

0.50

0.00

0.50

1.00

1.50

2.00

1.50

2.00

25%

10% 1.00

1.00

0.80

0.80

0.60

0.60

0.40

0.40

0.20

0.20 0.00

0.00 0.00

0.50

1.00

1.50

2.00

50%

1.00

75%

Committed Capacity

Figure 2 Engagement Committed Capacity Profiles The committed capacity profile for an engagement is a function of time (t), the kernel (k), the start of the working window (ts) and the end of the working window (te): t < ts 0,   t − ts , ts ≤t < ts + k ∧ t < te − k  e − ts − k t  k  ts + k ≤t < te − k , cc(t , k , ts , te ) ≡ te − ts − k  te − k ≤t < ts + k 1,  te − t , te − k ≤t < te ∧ ts + k ≤t  e − ts − k t  t ≥ te 0,  These capacity profiles are additive. As additional engagements are added to the Resource’s dance card, their individual committed capacity profiles can be Eng 1 Eng 2 summed to get the combined Eng 3 committed capacity profile for the Eng 4 Resource. Figure 3 shows the dance card of a Resource with 4 t im e 1.4 engagements added. This function → 1.2 provides a profile that can be used to 1.0 guide the location of future 0.8 0.6 engagements and to inform the 0.4 Resource of potential scheduling 0.2 conflicts that need to be resolved. 0 Figure 3 Dance Card with Four Engagements 05/07/99


Page 6


Committed Capacity

In Figure 3 the capacity profiles of Engagements 1, 2, and 3 overlap in such a way that the resource’s capacity is committed Eng 1 beyond 100% for a time. The Eng 2 committed capacity profile provides Eng 3 an easy way for the resource to Eng 4 locate these regions of over commitment on its dance card. In the time 1.4 example, Engagement 1 uses 100% → 1.2 of the capacity of the resource 1.0 during part of its working window. If 0.8 0.6 the working windows of 0.4 Engagements 2 and 3 were changed 0.2 so as to avoid overlapping with that 0 region of Engagement 1 then the over capacity violation would be Figure 4 Dance Card after Adjusting Engagements eliminated. Figure 4 shows the dance card after that adjustment has been made. These decisions can be done with straightforward algorithms without the need for complex reasoning. The capacity conflict was resolved without having to make any changes in the commitments to the upstream or downstream processes. If the Resource is unable to resolve the conflict by moving the working window it may renegotiate the original commitment window on an Engagement and as a last result, renege on its commitment entirely. Resources use their committed capacity profiles to encourage placement of future engagements in regions of low committed capacity by charging increasing hourly rates for increasing committed capacity. These cost profiles can be used to select which engagement to adjust when there are several alternatives available to resolve an over committed region. By determining which move results in a lower overall cost the resource is able to pick the best engagement to adjust. The ANTS software is currently being tested on a simulated shipbuilding enterprise. Deneb is modeling a portion of the Newport New Shipyard in Quest (a discrete event simulation package sold by Deneb). Each fabrication facility in a shipyard acts like a small independent manufacturing operation. Modeling a shipyard as a supply chain blends nicely with the ANTS architecture. The ANTS software is being implemented in Java on Voyager a commercial agent platform from ObjectSpace. Phase I is complete and testing has begun to understand the behavior of the basic algorithms under various operating conditions. The algorithms will be adapted as experience is gained in their use. 4. Conclusion Agent-based systems provide a promising alternative to the monolithic ERP software systems currently used in supply chain management. The ANTS architecture utilizes a uniform architecture that transparently covers the operations within a factory and across a supply chain. The negotiation among the agents provides a means by which resources can be allocated to tasks. The least commitment scheduling strategies are respond to varying loads without having to resort

05/07/99


Page 7


to rescheduling the entire supply chain. The DESK committed capacity profiles guide the allocation of resources to spread the load evenly and guide the resolution of resource conflicts. Acknowledgment ANTS is partially funded by the NIST ATP program. The prime contractor is Deneb (Naresh Raja). ANTS is based on initial research under the AARIA project funded by the DARPA Agile program. The prime contractor for AARIA was Intelligent Automation, Inc. References [1] A. D. Baker. Metaphor or Reality: A Case Study Where Agents Bid With Actual Costs to Schedule a Factory. In S. H. Clearwater, Editor, Market-Based Control: A Paradigm for Distributed Resource Allocation, pages 184-223. World Scientific Publishing Co. Pte. Ltd., 1996. [2] J. Butler and H. Ohtsubo. ADDYMS: Architecture for Distributed DYnamic Manufacturing Scheduling. In A. Famili, D. S. Nau, and S. H. Kim, Editors, Artificial Intelligence Applications in Manufacturing, pages 199-214. AAAI Press/The MIT Press, Menlo Park, CA, 1992. [3] N. A. Duffie, R. Chitturi, and J. I. Mou. Fault-tolerant Heterarchical Control of Heterogeneous Manufacturing System Entities. Journal of Manufacturing Systems, 7(4):315-28, 1988. [4] J. Heaton. Agent Architecture Distributes Decisions for the Agile Manufacturer: Reengineering at AlliedSignal Automotive Safety Restraint Systems. AMR Report, (June):8-13, 1994. [5] G. Y.-J. Lin and J. Solberg. Integrated Shop Floor Control using Autonomous Agents. IIE Transactions, 24(3):57-71, 1992. [6] J. Maley. Managing the Flow of Intelligent Parts. Robotics and Computer-Integrated Manufacturing, 4(3/4):525-30, 1988. [7] H. V. D. Parunak. Manufacturing Experience with the Contract Net. In M. N. Huhns, Editor, Distributed Artificial Intelligence, pages 285-310. Pitman, London, 1987. [8] H. V. D. Parunak. ’Go to the Ant’: Engineering Principles from Natural Agent Systems. Annals of Operations Research, 75:69-101, 1997. Available at http://www.iti.org/~van/gotoant.ps. [9] H. V. D. Parunak, A. D. Baker, and S. J. Clark. The AARIA Agent Architecture: From Manufacturing Requirements to Agent-Based System Design , Presented at the Workshop on Agent-Based Manufacturing, ICAA’98, Minneapolis, MN, 10 May 1998 [10] H. V. D. Parunak, J. Kindrick, and B. Irish. Material Handling: A Conservative Domain for Neural Connectivity and Propagation. In Proceedings of Sixth National Conference on Artificial Intelligence, pages 307-311, American Association for Artificial Intelligence, 1987.28, [11] M. J. Shaw and A. B. Whinston. Task Bidding and Distributed Planning in Flexible Manufacturing. In Proceedings of IEEE Int. Conf. on AI Applications, pages 184-89, 1985.

05/07/99


Page 8


[12] K. J. Tilley and D. J. Williams. Modelling of Communications and Control in an Auctionbased Manufacturing Control System. In Proceedings of IEEE International Conference on Robotics and Automation, pages 962-967, IEEE, 1992. [13] D. M. Upton, M. M. Barash, and A. M. Matheson. Architectures and auctions in manufacturing. International Journal of Computer Integrated Manufacturing, 4(1):23-33, 1991. [14] W. E. Walsh, M. P. Wellman, P. R. Wurman, and J. R. MacKie-Mason. Some Economics of Market-based Distributed Scheduling. In Proceedings of 18th International Conference on Distributed Computing Systems, 1998. Available at http://wwwpersonal.engin.umich.edu/~wew/Papers/icdcs98.ps.Z. [15] M. P. Wellman, W. E. Walsh, P. R. Wurman, and J. K. MacKie-Mason. Auction Protocols for Decentralized Scheduling. http://www-personal.engin.umich.edu/~wew/Papers/mbscheduling-extended.ps.Z, University of Michigan, Ann Arbor, 1998.

05/07/99


Page 9