Prioritized Sampling with Intrinsic Motivation in Multi-Task Reinforcement Learning