Policy Optimization Achieves Data-Dependent Regret Bounds in MDPs with Unknown Transitions | ArxivCSExplorer