Imperfect reward functions in real-world settings
This article introduces reward misspecification in reinforcement learning: reward functions are only proxies for our true goals, and powerful optimizers exploit gaps between what we write and what we …