Interpretable Control by Reinforcement Learning