Sharing Knowledge in Multi-Task Deep Reinforcement Learning