Learning Domain Invariant Representations in Goal-conditioned Block MDPs

TL;DR

This paper addresses the challenge of domain generalization in goal-conditioned deep RL, proposing a theoretical framework and a new method PA-SkewFit to improve robustness against environmental changes, with significant empirical gains.

Contribution

It introduces a theoretical framework for goal-conditioned RL in Block MDPs and proposes PA-SkewFit, a practical method to enhance domain invariance and generalization.

Findings

PA-SkewFit improves performance by 50% over baselines.

The framework characterizes policy generalizability in new environments.

Empirical results demonstrate robustness in unseen test environments.

Abstract

Deep Reinforcement Learning (RL) is successful in solving many complex Markov Decision Processes (MDPs) problems. However, agents often face unanticipated environmental changes after deployment in the real world. These changes are often spurious and unrelated to the underlying problem, such as background shifts for visual input agents. Unfortunately, deep RL policies are usually sensitive to these changes and fail to act robustly against them. This resembles the problem of domain generalization in supervised learning. In this work, we study this problem for goal-conditioned RL agents. We propose a theoretical framework in the Block MDP setting that characterizes the generalizability of goal-conditioned policies to new environments. Under this framework, we develop a practical method PA-SkewFit that enhances domain generalization. The empirical evaluation shows that our goal-conditioned RL agent can perform well in various unseen test environments, improving by 50% over baselines.