Improving Deep Reinforcement Learning with Knowledge Transfer