Adaptive Representations for Reinforcement Learning