Investigating the Properties of Neural Network Representations in Reinforcement Learning