Hierarchical Abstraction in Reinforcement Learning Based on Feature Discovery