Solving Markov Decision Processes via Largest-Size Average Estimator