Learning Dynamics and Generalization in Reinforcement Learning

Solving a reinforcement learning problem poses two competing challenges: fitting a potentially highly-discontinuous value function, and generalizing well to new observations. In this paper, we analyze the learning dynamics of temporal difference algorithms to gain novel insight into the tension between these two objectives. We show theoretically that temporal difference learning encourages agents to fit `high-frequency’ components of the value function early in training, and at the same time induces the second-order effect of discouraging generalization between states. We corroborate these findings in deep RL agents trained on a range of environments, finding that it is the nature of the TD targets themselves that discourages generalization. Finally, we investigate how post-training policy distillation may avoid this pitfall, and show that this approach improves generalization performance to novel environments in the ProcGen suite.

Authors' notes