L G ] 1 M ay 2 01 9 Information-Theoretic Considerations in Batch Reinforcement Learning