Using Model-Based Reflection to Guide Reinforcement Learning