A constraint based agent for TAC-SCM

The annual international Trading Agent Competition-Supply Chain Management (TAC-SCM) game is based around the manufacture and supply of PCs. There are multiple agents in the game, scheduling production, competing for orders from customers and components from suppliers. A key decision to be made each day in the game is what offers should be made to customers. Each day, the agents receive a set of request for quotes (RFQ) from customers, agents respond with offers, and then the customers select the lowest bid. We have developed an agent to compete in the competition that combines constraint-based optimisation, reasoning with probabilities, and learning of market conditions in an attempt to determine what customer requests to bid on and what prices to bid. Our agent maintains prices that correspond to different probabilities of success in winning contracts, using an online learning approach. By keeping track of the ratio of offers accepted to those made, the prices can be updated iteratively to move closer to their target probability. This range of price/probability pairs is then used as input to a constraint model. For each request, the model chooses whether or not to bid, and selects a price from the range. These decisions are restricted by capacity and supply constraints. A capacity constraint ensures that we will be able to schedule any new orders we receive with existing orders such that the factory capacity for each day in the current horizon is not exceeded. The agents production ability is also subject to component availability, By ordering components in advance, we know the current amount of components available, and we also know how much of each component will be arriving at each clay. This allows us to add a constraint for availability of supplies. An objective function is specified that maximises our expected profit, where the profit on a request is calculated by subtracting from the selling price the cost of components together with late delivery penalties. The agent is implemented in Java, using OPL Studio for the constraint-based optimisation. The agent is competing in the competition, which is a real-time simulation of 220 trading days, each day lasting 15 real seconds. Initial results show that the combination of online learning, uncertainty reasoning and constraint-based optimisation is effective and robust, producing a competitive trading agent.