Sample Efficient Reinforcement Learning via Model-Ensemble Exploration and Exploitation