Toward the Generalization of Reinforcement Learning