BAIL: Best-Action Imitation Learning for Batch Deep Reinforcement Learning