Enhancing Generative Auto-bidding with Offline Reward Evaluation and Policy Search

The article introduces AIGB-Pearl, an advanced method for Auto-bidding that combines generative planning with policy optimization to enhance performance beyond existing offline reinforcement learning approaches. AIGB-Pearl employs a trajectory evaluator for score assessment and implements a KL-Lipschitz-constrained score-maximization scheme to facilitate safe exploration of data. This method shows significant improvements in advertising performance, making it a valuable tool for practitioners seeking to optimize bidding strategies with AI.