Structured Computation and Representation in Deep Reinforcement Learning