Unbiased Asymmetric Reinforcement Learning under Partial Observability

TL;DR

This paper introduces an unbiased asymmetric actor-critic method for partially observable reinforcement learning, providing theoretical guarantees and demonstrating improved convergence and policy quality over biased methods.

Contribution

It proposes a theoretically sound unbiased asymmetric actor-critic algorithm that effectively exploits state information under partial observability.

Findings

Unbiased asymmetric actor-critic converges faster than biased baselines.

The method achieves better policies in partially observable domains.

Theoretical analysis exposes issues in common asymmetric methods.

Abstract

In partially observable reinforcement learning, offline training gives access to latent information which is not available during online training and/or execution, such as the system state. Asymmetric actor-critic methods exploit such information by training a history-based policy via a state-based critic. However, many asymmetric methods lack theoretical foundation, and are only evaluated on limited domains. We examine the theory of asymmetric actor-critic methods which use state-based critics, and expose fundamental issues which undermine the validity of a common variant, and limit its ability to address partial observability. We propose an unbiased asymmetric actor-critic variant which is able to exploit state information while remaining theoretically sound, maintaining the validity of the policy gradient theorem, and introducing no bias and relatively low variance into the training process. An empirical evaluation performed on domains which exhibit significant partial observability confirms our analysis, demonstrating that unbiased asymmetric actor-critic converges to better policies and/or faster than symmetric and biased asymmetric baselines.