Deep Reinforcement Learning Policies Learn Shared Adversarial Features Across MDPs

TL;DR

This paper investigates the shared adversarial features in neural policies across different MDPs, revealing that high sensitivity directions are correlated and support the presence of shared non-robust features, advancing understanding of policy robustness.

Contribution

It introduces a framework to analyze decision boundary and loss landscape similarities across states and MDPs, highlighting shared adversarial features in deep reinforcement learning policies.

Findings

High sensitivity directions are correlated across MDPs.

Shared non-robust features are present across training environments.

Results support the hypothesis of environment-invariant adversarial features.

Abstract

The use of deep neural networks as function approximators has led to striking progress for reinforcement learning algorithms and applications. Yet the knowledge we have on decision boundary geometry and the loss landscape of neural policies is still quite limited. In this paper we propose a framework to investigate the decision boundary and loss landscape similarities across states and across MDPs. We conduct experiments in various games from Arcade Learning Environment, and discover that high sensitivity directions for neural policies are correlated across MDPs. We argue that these high sensitivity directions support the hypothesis that non-robust features are shared across training environments of reinforcement learning agents. We believe our results reveal fundamental properties of the environments used in deep reinforcement learning training, and represent a tangible step towards building robust and reliable deep reinforcement learning agents.