Abstraction for Deep Reinforcement Learning