The Utility of Sparse Representations for Control in Reinforcement Learning