Combating Reinforcement Learning's Sisyphean Curse with Intrinsic Fear

To use deep reinforcement learning in the wild, we might hope for an agent that would never make catastrophic mistakes. At the very least, we could hope that an agent would eventually learn to avoid old mistakes. Unfortunately, even in simple environments, modern deep reinforcement learning techniques are doomed by a Sisyphean curse. Owing to the use of function approximation, these agents eventually forget experiences as they become exceedingly unlikely under a new policy. Consequently, for as long as they continue to train, state-aggregating agents may periodically relive catastrophic mistakes. We demonstrate unacceptable performance of deep Q-networks on two toy problems. We then introduce intrinsic fear, a method that mitigates these problems by avoiding dangerous states.