Learning with Opponent Modeling in Repeated Auctions

We design an algorithm to learn bidding strategies in repeated auctions. We consider seller and all bidders simultaneously for strategy learning and explore the convergence of this system. We apply and improve the opponent modeling class algorithm to allow bidders to learn optimal bidding strategies in this multiagent reinforcement learning environment. The algorithm uses almost no private information about the opponent and has no restrictions on the strategy space, so it can be extended to multiple scenarios. Our algorithm improves the utility compared to both static bidding strategies and dynamic learning strategies. We hope the application of opponent modeling in auctions will promote the research of bidding strategies in online auctions and the design of non-incentive compatible auction mechanisms.