Uncertain Decisions Facilitate Better Preference Learning

Cassidy Laidlaw, Stuart Russell

Existing observational approaches for learning human preferences, such as inverse reinforcement learning, usually make strong assumptions about the observability of the human's env…