Reinforcement Learning with Competitive Ensembles of Information-Constrained Primitives