MBVI: Model-Based Value Initialization for Reinforcement Learning