Selecting the State-Representation in Reinforcement Learning