The 2002 Trading Agent Competition: An Overview of Agent Strategies

Suppose you want to buy a used Canon AE1 SLR camera at an online auction. It would be quite a daunting task to manually monitor prices and make bidding decisions at all web sites currently offering the camera—especially if accessories such as a flash and a tripod are sometimes bundled with the camera and sometimes auctioned separately. However, for the next generation of trading agents, autonomous bidding in simultaneous auctions will be a routine task. Simultaneous auctions, which characterize internet sites such as eBay.com, are a challenge to bidders, particularly when complementary and substitutable goods are on offer. Complementary goods are items such as a flash and a tripod that would complement a camera, but a bidder wants the flash and tripod only if he/she is certain to acquire the camera. Substitutable goods are goods such as the Canon AE-1 and the Canon A-1—a bidder desires one or the other but not both. In combinatorial auctions, bidders bid on combinations of items, such as “camera and flash for $295”; in these auctions, the (NP-complete) problem of determining how to allocate the goods to maximize revenue falls in the hands of the auctioneer. In simultaneous auctions, however, the complexity burden lies with the bidders. The International Trading Agent Competition (TAC) annually challenges its entrants to best design online agents capable of bidding in simultaneous auctions for substitutable and complementary goods. The original TAC was designed and operated by a group at the University of Michigan AI Laboratory (Wellman et al. 2001). In 2002, for the third rendition of the competition, the TAC software platform was redesigned by the Intelligent Systems Laboratory at the Swedish Institute of Computer Science (SICS). Since the first International Trading Agent Competition back in 2000, TAC agent design has come a long way. In TAC-00, agent designs were primarily centered around designing algorithms to solve an NP-complete optimization problem. However, by the second year, it became common knowledge that this problem was tractable for the TAC travel game parameters. During the second year, agent designs focused on estimating clearing prices, and some agents designed algorithms that made use of distributional price estimates. Agent design in TAC-02, however, cannot be described so succinctly. Agent architects used numerous general-purpose AI techniques, including machine learning, planning, partially observable Markov decision processes, Monte Carlo simulations, and multiagent systems. Ultimately, however, the most successful agents were primarily heuristically based and domain specific.