Combating Reinforcement Learning's Sisyphean Curse with Intrinsic Fear
This addresses a practical issue for DRL applications in environments with catastrophic states, though it is incremental as it builds on existing reward shaping techniques.
The paper tackles the problem of Deep Reinforcement Learning agents periodically revisiting catastrophic states by introducing intrinsic fear, a learned reward shaping method that penalizes the Q-learning objective based on predicted catastrophe probability, resulting in agents solving pathological environments and improving performance on several Atari games.
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.