SMS dit: Adaptive representations for reinforcement learning