A Simple Agent for Supply Chain ManagementBrian Farrell and Danny LoffredoDecember 7th, 20061 IntroductionThe Trading Agent Competition Supply Chain Management (TAC SCM) gameis a competitive testbed for software agents designed to perform supply chainmanagement. While not a perfect simulation of reality, the game is sufficientlycomplex to encourage the development of complex supply chain managementstrategies.A TAC SCM game consists of 220 simulated days. On each day, customerssend requests for quotes (RFQs) for new computers to the agents. The agentsrespond to these RFQs with offers that can be accepted or rejected by thecustomers. The agents build computers using parts bought from suppliers. Theagents send RFQs to the suppliers, specifying a desired price and due date. Thesuppliers can then make an offer on the RFQs, which the agents can accept orreject. More detailed information about the game can be found in the officialspecification document [1]. Stated simply, there are two main objectives in thegame: to minimize component procurement costs, and to maximize computersales prices.In this paper, we describe Simplicity, an agent designed to compete in TACSCM. The agent was designed using a simple approach in order to see howit would fare against more complex implementations from previous TAC SCMtournaments. The paper is organized as follows. First, we explain the designand implementation of the agent. Next, we analyze the success of our strategies.Finally, we suggest improvements that could be made in future work.2 ImplementationOur chief goal in designing Simplicity was to create a simple agent. While thelimited timeframe of the project necessitated a simple design, there are alsoother benefits to designing a simple agent. However, there are many benefitsof a simple agent design. All other things being equal, a simple agent wouldbe preferred over one that is more complex. It is easier to understand, im-plement, and debug. We hoped to show whether a simple agent could still doreasonably well when compared to more complex approaches. Even if the agentdid not perform as well as some of the other agents entered in the TAC SCM1tournament, it could still be possible to describe it as successful if it performednearly as well as the other agents. There are very likely to be tradeoffs betweenagent performance and complexity. The goal of Simplicity was to see if much ofthe complexity could be reduced while still maintaining a reasonable amount ofperformance.Another advantage Simplicity has over other agents is that it does not useany data from previous games. Several of the top performing agents in the 2005TAC SCM tournament used data from previous games in some fashion [3][4].While this data will always be easily obtained in this competition, it may beharder or impossible to obtain in other domains, so there is some benefit to anagent that does not need past data. However, an agent that does not requirepast data is not necessarily more robust to drastic changes in the market. Eventhough the agent does not use data from the past, it was still tested in marketconditions similar to the ones it would face in the competition. It is hard to saywhether it would perform well if some facets of the competition were drasticallychanged.An important idea that informed the design was the possibility of sellingcomputers at a loss. There is a bias towards low-demand games [5]. During pe-riods of low demand, prices can drop below the level at which sales are profitable.An intelligent agent must know when it is better off accumulating inventory andwhen it is better off selling computers. In order to make this decision reliably,the agent needs an accurate measure of its costs.The tasks of the agent are split into two modules: the Demand Managerand the Supply Manager. The Demand Manager handles customer RFQs andfactory production. It decides which RFQs to bid on, and what prices to pay.It then schedules factory production and deliveries, and projects how manycomponents it will need in the future to maintain production. The SupplyManager deals with suppliers. It uses the usage projections from the DemandManager to buy inventory.2.1 Supply ManagerThe Supply Manager’s main goal when placing normal orders is to be veryflexible. The price offered by a supplier is a function of its committed capacity,so prices are very much affected by the actions of other agents. An agent thatalways buys components with a lead time of 5 days will do well when playingagainst agents that all use long-term buying strategies, but will perform poorlywhen playing against agents with short-term buying strategies. Therefore, it isimportant to be as flexible as possible when considering component lead times.The Supply Manager sends five RFQs to suppliers each day. First, it sendsan RFQ to procure any components necessary to prevent shortages in the nexttwenty days. Next, it sends an RFQ for components with either an extremelyshort or long lead time. Finally, it uses whatever remaining RFQs it has to sendzero-quantity RFQs to suppliers which it uses to model the suppliers’ productionto estimate component costs in the future.22.1.1 Normal RFQsEach day, the Supply Manager uses the list of projected component use suppliedby the Demand Manager to determine when its inventory will drop below aspecified level. This level was set at 800 for non-CPU components and 400for CPU components. If the projected inventory level after twenty days is stillabove the threshold, no normal orders are placed. If not, the Supply Managerlooks at all of the days that it has price predictions for that are before the daythat inventory becomes too low. It plans to buy slightly more than it wouldneed to get inventory levels back to the threshold. This has the effect of keepingcomponent lead times closer to the twenty-day cutoff. Once the Supply Managerfinds the day with the lowest predicted price, it sends an RFQ with a reserveprice slightly greater than the predicted price. Since large procurements byother agents could drastically change the supplier’s offer price, and the agentuses RFQs that are up to three days old in price prediction, a reserve price isnecessary. It keeps the agent from paying a price much greater than it expected.Because of this reserve price, the agent can simply accept every offer that it getsas a result of the RFQ.2.1.2 Short- and Long-Term RFQsWhile the strategy for procuring components via normal orders is fairly flexible,it ignores
View Full Document
Unlocking...