AIGB Track Winner (CleanDiffuser) Presentation

The authors propose a diffusion bidding policy named DARE, which integrates ensemble Q-learning and action re-labeling. The approach first systematically analyzes the bidding trajectory dataset, clustering critical indicators in the decision-making process, and then implements ensemble Q-learning based on the simplified representation. DARE subsequently backtracks the optimal action in the trajectory using the ensemble Q-function and relabels the dataset. The approach employs a diffusion policy to aggregate multi-peak distribution policies from ensemble Q-functions, thereby enhancing generalization capabilities beyond the original distribution and forming an end-to-end bidding decision model.