Constrained Intrinsic Motivation for Reinforcement Learning