Fully Offline Reinforcement Learning

arXiv cs.AI · 6 hours ago

Researchers introduced SOReL, a fully offline Bayesian reinforcement learning method that learns dynamics posteriors and selects hyperparameters without any online interactions, and TOReL, which extends this tuning framework to other offline RL algorithms. The approach achieves minimax-optimal parametric regret rates under standard conditions. This enables safe deployment of RL systems without requiring undocumented online tuning interactions or performance estimates from online data.