Q-Regularized Generative Auto-Bidding: From Suboptimal Trajectories to Optimal Policies
The episode discusses "Q-Regularized Generative Auto-Bidding," a paper by researchers from Wuhan University and Alibaba. It addresses the challenge of using suboptimal historical data for auto-bidding. The authors propose QGA, a method that combines generative modeling with a value-based critic to guide the model toward optimal actions, achieving significant performance gains in real-world tests on platforms like Taobao.