Intermediary-Supervised Auction Mechanism Design With Deep Neural Networks

Abstract

An optimal auction mechanism is frequently characterized by its ability to allocate items to bidders who offer the highest marginal revenue. This optimality is described from the perspective of the auctioneer. Its design focus is typically on the unilateral constraints imposed by the auctioneer on the bidders, which can pose a challenge in achieving both incentive compatibility and maximizing revenue. This paper innovatively proposes an intermediary module, decomposing the auction process into a multi-objective optimization task. Specifically, we propose a novel auction mechanism design method called the Tri-Auction Game Engine (TAGE). In this framework, bidders strive to maximize their utility through bidding; the auctioneer concentrates on maximizing revenue by determining allocations and payments based on these bids; and the intermediary plays a pivotal role in modeling the tolerance to ensure the effective regulation of the auction process. Furthermore, we employ an adaptive annealing strategy, which models tolerance to dynamically adjust the optimization process of the model. This approach balances revenue maximization and incentive compatibility constraints, and eliminates the reliance on ex-post regret inherent in traditional methods. Finally, we demonstrate through experiments that TAGE outperforms baseline models in all settings, thereby providing valuable insights for the design of future auction mechanisms.