Reinforcement Learning with Multiple, Qualitatively Different State Representations