Knowledge Transfer in Multi-Task Deep Reinforcement Learning for Continuous Control