Optimal-er Auctions through Attention

RegretNet is a recent breakthrough in the automated design of revenue-maximizing auctions. It combines the expressivity of deep learning with the regret-based approach to relax the Incentive Compatibility constraint (that participants benefit from bidding truthfully). We propose two independent modifications of RegretNet, namely a neural architecture based on the attention mechanism, denoted as RegretFormer, and an interpretable loss function that is significantly less sensitive to hyperparameters. We investigate both proposed modifications in an extensive experimental study that includes settings with constant and varied number of items and participants, novel validation procedures, and out-of-setting generalization. We find that RegretFormer consistently outperforms existing architectures in revenue and, unlike existing architectures, is applicable when the input size is variable. Regarding our loss modification, we confirm its effectiveness in controlling the revenue-regret trade-off by varying a single interpretable hyperparameter.