Characterizing the Gap Between Actor-Critic and Policy Gradient

TL;DR

This paper precisely characterizes the difference between actor-critic and policy gradient methods in reinforcement learning, introduces algorithms to correct this gap, and demonstrates improved performance in experiments.

Contribution

It provides a detailed theoretical analysis of the AC-PG gap, introduces practical correction algorithms, and empirically shows enhanced efficiency and performance.

Findings

Corrected AC methods improve sample efficiency

Enhanced final performance of reinforcement learning algorithms

Theoretical framework unifies AC and PG as a Stackelberg game

Abstract

Actor-critic (AC) methods are ubiquitous in reinforcement learning. Although it is understood that AC methods are closely related to policy gradient (PG), their precise connection has not been fully characterized previously. In this paper, we explain the gap between AC and PG methods by identifying the exact adjustment to the AC objective/gradient that recovers the true policy gradient of the cumulative reward objective (PG). Furthermore, by viewing the AC method as a two-player Stackelberg game between the actor and critic, we show that the Stackelberg policy gradient can be recovered as a special case of our more general analysis. Based on these results, we develop practical algorithms, Residual Actor-Critic and Stackelberg Actor-Critic, for estimating the correction between AC and PG and use these to modify the standard AC algorithm. Experiments on popular tabular and continuous environments show the proposed corrections can improve both the sample efficiency and final performance of existing AC methods.