Semantic State Representation for Reinforcement Learning