Model-based Credit Assignment for Model-free Deep Reinforcement Learning