Provably Good Batch Reinforcement Learning Without Great Exploration