Combining Model-Based and Model-Free Reinforcement Learning Policies for More Efficient Sepsis Treatment.