State Construction In Reinforcement Learning