Agent Service for Online Auctions

We have designed configurable agents to represent users in online auctions, specificMly the Michigan AuctionBot. The agents can be configured, started, and monitored from a web interface. We implemented three types of agents, distinguished by their different ways of using information in the auctions. A competitive agent does not use any information in the auction market. It chooses its actions based on its individual optimization problem. A price modeling agent uses price history as its only information. A bidder-modeling agent uses other agents’ bidding histories to predict their next bids and infer the next clearing price. Our experiments suggest that an agent’s performance in the auctions depends not only on its bidding strategy, but also on the bidding strategies of others. When all the agents behave strategically they may reach a sub-optimal equilibrium, in which they receive worse payoffs than behaving competitively. keywords: Web agents, internet auctions, online learning, strategic bidding Introduction Intelligent agents for electronic commerce are a popular research topic. There are shopping agents that collect price information for users (Doorenbos, Etzioni, Weld 1997), and information filtering agents that collect interesting publications (Bollacker, Lawrence, & Giles 1998). We are interested in designing agents for online auctions, where buyers and sellers interact with each other. Such agents can work on the behalf of users since users usually do not have time or inclination to monitor the activities in an auction. Sometimes an optimal bidding strategy may be computationally intensive, in which case it is especially useful to have a software agent carry out the bidding. The interesting research issue for us is how an agent takes advantage of the information available and achieves maximal profit in the transactions. This usually refers to how an agent uses past observations to make predictions and choose its optimal bids. We also address design issues such as how an agent works in a web environment and how it gathers information and makes decisions in real time. We have designed an agent server that works on a user’s behalf to submit bids to one of the online auctions--the Michigan AuctionBot. The users specify the names of the auctions they want to participate in, the initial amounts of the goods, and the bidding strategies they prefer. The agent then starts bidding on the AuctionBot for the users. The agents keep bidding until the auction closes, and then report the results back to the users. Our experiments suggest that an agent’s performance in the auction depends not only on its bidding strategy, but also on the bidding strategies of others. A greedy bidding strategy may help the agent to gain in the short run, but may also cause it to lose in the long run.