Rule-based Recurrent Reinforcement Learning