Unified Algorithms for RL with Decision-Estimation Coefficients: PAC, Reward-Free, Preference-Based Learning, and Beyond

arXiv:2209.11745v3 Announce Type: replace
Abstract: Modern Reinforcement Learning (RL) is more than just learning the optimal policy; Alternative learning goals such as exploring the environment, estimating the underlying model, and learning from preference feedback are all of practical importance. While provably sample-efficient algorithms for each specific goal have been proposed, these algorithms often depend strongly on the particular learning goal and thus admit different structures correspondingly. It is an urging open question whether these learning goals can rather be tackled …

abstract algorithms alternative arxiv beyond cs.ai cs.lg decision environment feedback free importance math.st modern policy practical reinforcement reinforcement learning sample stat.ml stat.th the environment type while