A New Interpretation of the Certainty-Equivalence Approach for PAC Reinforcement Learning with a Generative Model