Dual Active Learning for Reinforcement Learning from Human Feedback
This paper first proposes a dual active reward learning algorithm for the simultaneous selection of conversations and teachers, and applies pessimistic RL to solve the alignment problem, based on the learned reward estimator.