Variance reduction for policy gradient with action-dependent factorized baselines

Read on openai.com