Feature Generalization In Deep Reinforcement Learning: An Investigation Into Representation Properties