Learning to Interrupt: A Hierarchical Deep Reinforcement Learning Framework for Efficient Exploration