Lecture
This lecture explains the importance of subtracting the mean reward in policy gradient methods for deep reinforcement learning. It covers topics such as the log-likelihood trick, online gradient rules for one-step and multi-step horizons, learning value functions, and the use of baselines. The instructor also discusses the REINFORCE algorithm with a baseline, the variance reduction achieved by subtracting the mean, and the outlook on deep reinforcement learning with alpha-zero networks.