TAC-2000 was the first autonomous bidding agent competition. While it was a very successful event, some minor improvements would increase its interest from a multiagent learning perspective.
With or without these modifications, we hope to be able to participate in future TACs, with the goal of adding additional adaptive elements to ATTac-2000.
Another direction of future research is to apply the lessons learned from TAC to real simultaneous interacting auctions. It is straightforward to write bidding agents to participate in on-line auctions for a single good if the value to the client is fixed ahead of time: the agent can bid slightly over the ask price until the auction closes or the price exceeds the value. However, when the values of multiple goods interact, such as is the case in TAC, agent deployment is not nearly so straightforward.
One such real application is the Federal Communications Commission's auctioning off of radio spectrum [WeberWeber1997,CramtonCramton1997]. Especially for companies that are trying to achieve national coverage, the values of the different licenses interact in complex ways. Perhaps autonomous bidding agents will be able to affect bidding strategies in such future auctions. Indeed, in related research we have begun down this path by creating straightforward bidding agents in a realistic FCC Auction Simulator [Csirik, Littman, Singh, StoneCsirik et al.2001].
In a more obvious application, an extended version of ATTac-2000 could potentially become useful to real travel agents, or to end users who wish to create their own travel packages.
We would like to thank the TAC team at the University of Michigan, including Michael Wellman, Peter Wurman, Kevin O'Malley, Daniel Reeves, and William Walsh, for constructing the TAC server and responding promptly and cordially to our many requests while conducting the research reported here. We would also thank the anonymous reviewers for their helpful comments and suggestions.