The Principle of Unchanged Optimality in Reinforcement Learning Generalization