Quantifying generalization in reinforcement learning