Feature Control as Intrinsic Motivation for Hierarchical Reinforcement Learning