In this project we have created a configurable agent that can participate in certain auctions hosted on the Michigan AuctionBot using three different bidding strategies. We perform regression on the bidding histories of other agents and use this to predict a clearing price for the auction. We then make an adjustment to the reservation prices in hopes of getting more surplus when the transaction is made. The bidding agents are started with parameters specifying the AuctionBot user, the type of bidding strategy, and the initial endowments.
This project was designed for a specific type of auction (synchronous, Mth-price auctions) but in the future we would like to generalize our agents to participate in other types of auctions as well. For example, stock exchanges are asynchronous double auctions. Another future enhancement will be to add additional bidding strategies. We will then be collecting detailed performance statistics to determine which strategies perform better under which types of auctions. We would also like to compare these strategies to human strategies. We will also provide semi-automated bidding services which allow a human to have control over the bidding process.