Optimal medication dosing from suboptimal clinical examples: a deep reinforcement learning approach.