Towards Optimal Offline Reinforcement Learning