Cohere talk covers on-policy value learning at 10,000 FPS

Covers on-policy value learning using REPO, achieving 10,000 frames per second. Explains bias-variance tradeoff, resampling, and when PPO beats value methods.

Featured · Claas Voelcker