Analysis and Design of Supply-Driven Strategies in TAC SCM Wolfgang Ketter, Elena Kryzhnyaya, Steven Damer, Colin McMillen, Amrudin Agovic, John Collins, and Maria Gini Dept. of Computer Science and Engineering, University of Minnesota Abstract We describe two sales strategies used by the MinneTAC agent for the 2003 Supply Chain Management Trading Agent Competition. Both strategies estimate, as the game progresses, the probability of receiving a customer order for different offer prices. Offers are made to maximize the ex- pected profit margin on each order. The main difference be- tween the strategies is in how they compute the probability of receiving an order and the offer prices. The first strat- egy works well in high-demand games, the second was de- signed to improve performance in low-demand games. We analyze empirically the effect of the discount given by sup- pliers on orders received the first day of the game, and we show that in high-demand games there is a correlation be- tween the offers an agent receives from suppliers the first day of the game and the agent’s performance in the game. 1. Introduction Competitive scenarios are increasingly being used as testbeds for the development of multiagent systems. A new game, called TAC SCM, was introduced for the 2003 Trad- ing Agent Competition [8]. This game involves a Supply Chain Management (SCM) scenario in which agents at- tempt to maximize profits by manufacturing personal com- puters and selling them to customers. The TAC SCM competition is interesting for many rea- sons. Agents must base their decisions on limited informa- tion about the state of the market and the strategies of other agents. Agents must simultaneously compete in two sepa- rate but interrelated markets: the market from which they buy their supplies and the market to which they sell their finished products. Agents have a large number of decisions to make in a limited time, so the computational efficiency of the decision-making process is important. We describe two sales strategies used by our agent, Min- neTAC, and analyze their performance in different games. We show how the start-effect caused by the large discount given by suppliers on orders made the first day, coupled with the random order in which agent requests are considered, af- fects the outcome of the game. 2. Overview of TAC SCM Six autonomous agents compete to maximize profits in a computer-assembly scenario. The simulation takes place over 220 virtual days, each lasting fifteen seconds of real time. Each agent has a bank account with an initial balance of zero. The agent with the highest bank balance at the end of the game wins. Agents earn money by selling comput- ers they assemble out of parts purchased from suppliers. To obtain parts, an agent must send a Request For Quotes (RFQ) to an appropriate supplier. Each RFQ specifies a component type, a quantity, and a due date. The next day, the agent will receive a response to each request. Suppliers respond by evaluating each RFQ to determine how many components they can deliver on the requested due date and how long it would take to produce all the components re- quested, considering the outstanding orders they have com- mitted to and the RFQs they have already responded to this turn. If the supplier can produce the desired quantity on time, it responds with an offer that contains the price of the supplies. If not, the supplier responds with two offers: (1) an earliest complete offer with a revised due date and a price, and (2) a partial offer with a revised quantity and a price. The agent can accept either of these alternative of- fers, or reject both. Suppliers may deliver components late, due to randomness in their production capacities. If the sup- plier has excess capacity, the price will be discounted; dis- counted prices may be as low as 50% of the base price. Every day each agent receives a set of RFQs from po- tential customers. Each customer RFQ specifies the type of computer requested, along with quantity, due date, reserve price, and late penalty. Each agent may choose to bid on some or all of the day’s RFQs. Customers accept the low- est bid that is at or below their reserve price, and notify the agent the following day. The agent must ship customer or- ders on time, or pay a penalty for each day an order is late. If a product is not shipped within five days of the due date the order is cancelled, the agent receives no payment, and no further penalties accrue.