New AI Bidding System Boosts Online Ad Performance by 3.6%

A novel hierarchical reinforcement learning framework, named HOBA, has been developed to enhance online advertising bidding systems. This framework addresses common limitations such as the inability to adapt to dynamic auction markets and the need for extensive manual tuning of bid parameters. HOBA operates on three distinct time scales: a high-level large language model infers hyperparameters, a mid-level SARSA agent dynamically selects from expert models, and a low-level pool of experts executes bids. This design minimizes exploration risk by confining online learning to discrete expert selection rather than continuous bid optimization. Evaluations on the AuctionNet benchmark and a large-scale A/B test confirmed HOBA's effectiveness, showing consistent improvements over existing baselines. In a real-world deployment, the system achieved a 3.6% increase in target cost, validating its practical business value.