Graph-based State Representation for Deep Reinforcement Learning