Speaker: Michael J. Curry (University of Maryland)
Title:Deep Learning for Auction Design: Certifying and Enforcing Strategyproofness
Abstract: The design of revenue-maximizing strategyproof auctions has proved difficult when there are multiple bidders and multiple items. Motivated by the lack of theoretical progress, a recent approach to automated mechanism design, differentiable economics, represents auctions by rich function approximators and optimizes their performance by gradient descent. One such technique, RegretNet, gives state-of-the-art performance and works in settings beyond the reach of previous techniques. A major limitation, though, is that the strategyproofness of the resulting mechanisms can only be measured approximately using empirical techniques. We present two recent papers which attempt to mitigate this problem. The first adapts integer programming techniques from the adversarial robustness literature to exactly compute the maximum degree by which strategyproofness is violated. The second presents a new auction architecture: a variation of affine maximizer auctions modified to offer lotteries and learnable end-to-end. This approach, while not able to represent every feasible auction, gives competitive performance in terms of revenue, while guaranteeing strategyproofness by construction.
