Auction learning as a two-player game

Designing an incentive compatible auction that maximizes expected revenue is a centralproblem in Auction Design. While theoretical approaches to the problem have hit some limits, arecent research direction initiated by Duetting et al. (2019) consists in building neural networkarchitectures to find optimal auctions. We propose two conceptual deviations from their approachwhich result in enhanced performance. First, we use recent results in theoretical auctiondesign (Rubinstein and Weinberg, 2018) to introduce atime-independentLagrangian. This notonly circumvents the need for an expensive hyper-parameter search (as in prior work), but alsoprovides a principled metric to compare the performance of two auctions (absent from priorwork). Second,the optimization procedure in previous work uses an inner maximization loop tocompute optimal misreports. We amortize this process through the introduction of an additionalneural network. We demonstrate the effectiveness of our approach by learning competitive orstrictly improved auctions compared to prior work. Both results together further imply a novelformulation of Auction Design as a two-player game with stationary utility functions.