Reinforcement Learning with Information-Theoretic Actuation