Relative Entropy Pathwise Policy Optimization

摘要

Score-function based methods for policy learning, such as REINFORCE and PPO, have delivered strong results in game-playing and robotics, yet their high variance often undermines training stability. Improving a policy through state-action value functions, for example by differentiating Q with regard to the policy, alleviates the variance issues. However, this requires an accurate action-conditioned value function, which is notoriously hard to learn without relying on replay buffers for reusing past off-policy data. We present Relative Entropy Pathwise Policy Optimization, an algorithm that trains Q-value models purely from on-policy trajectories, unlocking the use of Q function derivatives to compute policy updates in the context of on-policy learning. We show how to combine stochastic policies for exploration with constrained updates for stable training, and evaluate important architectural components that stabilize value function learning. This results in an efficient on-policy algorithm that combines the stability of Q-based policy gradients with the simplicity and minimal memory footprint of standard on-policy learning. Compared to state-of-the-art on two standard GPU-parallelized benchmarks, REPPO provides strong empirical performance at superior sample efficiency, wall-clock time, memory footprint, and hyperparameter robustness.