Combating Reinforcement Learning's Sisyphean Curse with Intrinsic Fear