Combating Reinforcement Learning's Sisyphean Curse with Intrinsic Fear

TL;DR

This paper introduces intrinsic fear, a learned reward shaping method that helps reinforcement learning agents avoid catastrophic states, improving safety and learning speed in complex environments.

Contribution

The paper proposes intrinsic fear, a novel approach that predicts imminent catastrophes and penalizes them, enhancing safety and efficiency in deep reinforcement learning.

Findings

Intrinsic fear improves safety by avoiding catastrophic states.

Intrinsic fear accelerates learning speed.

Intrinsic fear enhances performance on Atari games.

Abstract

Many practical environments contain catastrophic states that an optimal agent would visit infrequently or never. Even on toy problems, Deep Reinforcement Learning (DRL) agents tend to periodically revisit these states upon forgetting their existence under a new policy. We introduce intrinsic fear (IF), a learned reward shaping that guards DRL agents against periodic catastrophes. IF agents possess a fear model trained to predict the probability of imminent catastrophe. This score is then used to penalize the Q-learning objective. Our theoretical analysis bounds the reduction in average return due to learning on the perturbed objective. We also prove robustness to classification errors. As a bonus, IF models tend to learn faster, owing to reward shaping. Experiments demonstrate that intrinsic-fear DQNs solve otherwise pathological environments and improve on several Atari games.