Interpretable End-to-end Autonomous Driving with Reinforcement Learning