[UTMD-102] Accelerating Automated Auction Design using Sufficiency of Local Incentive Compatibility (by Genta Okada)

Genta Okada

Deep learning has emerged as a promising approach for optimal auction design, particularly in complex multi-bidder, multi-item settings. RegretNet (Du¨tting et al. (2019)) is a notable architecture in this domain, capable of learning revenue-maximizing auctions. However, ensuring Dominant-Strategy Incentive Compatibility (DSIC) in RegretNet, specifically through the estimation of ‘regret,’ is computationally intensive, posing a barrier to its practical application. This paper proposes a method to accelerate RegretNet training by leveraging the sufficiency of Local Incentive Compatibility (LIC) for DSIC under common auction settings. Instead of minimizing global regret, which requires an expensive search for optimal deviations, we minimize ‘local regret’, restricting the search to a small neighborhood of the true valuation. This approach significantly reduces the computational burden of regret estimation, primarily by decreasing the number of gradient ascent steps required. Experimental results demonstrate that our approach reduces training time by approximately 80% while maintaining revenue and regret levels comparable to the original RegretNet. This work contributes a more efficient training methodology for RegretNet, thereby enhancing its accessibility for designing optimal auctions.