Learning Retrospective Knowledge with Reverse Reinforcement Learning