Optimization of Advertising Bidding Model With Generative AI and Deep Reinforcement Learning Collaboration: Design and Validation of Closed‐Loop Simulation Framework

ABSTRACT Real‐time bidding (RTB) requires decisions in milliseconds under nonstationary market dynamics; traditional optimization and online RL suffer from sample efficiency, stability and generalization limitations. This study proposes a closed‐loop simulation framework coupling generative AI with deep reinforcement learning. A cascaded encoder combining a conditional variational autoencoder (CVAE) with a denoising diffusion probability model (DDPM) synthesizes high‐fidelity market states. A proximal policy optimization (PPO) agent trains in this environment via a three‐stage ‘generate, evaluate, optimize’ mechanism, with a dynamic feedback regulator forming a co‐evolutionary loop. Experiments on 47.82 million anonymized bidding records improve CTR by 23.7%, CVR by 31.2%, ROI by 23.5% over DQN and increase ROI by 6.1% over GAVE and 8.6% over DiffBid. A 2 × 2 factorial ablation confirms significant generator‐agent interaction ( = 0.37 ROI units, 95% CI [0.24, 0.51], = 0.008, = 0.024). Cross‐dataset transfer on iPinYou preserves rankings with 7.0% ROI advantage over GAVE; counterfactual OPE on real logs (SNIPS, doubly robust, auction replay) agrees with simulator ranking. The system achieves 87.4% sample utilization (distinct from GPU utilization 92.30%) and 14.8 ms mean inference latency (p50/p95/p99 = 12.4/18.7/23.1 ms; 100% SLA at 100,000 req/s). A live A/B test on a production platform is planned as future work.