Hypothesis-Driven Skill Discovery for Hierarchical Deep Reinforcement Learning