Parenting: Safe Reinforcement Learning from Human Input

Christopher Frye, Ilya Feige

Autonomous agents trained via reinforcement learning present numerous safety concerns: reward hacking, negative side effects, and unsafe exploration, among others. In the context o…