Learning Interpretable, High-Performing Policies for Autonomous Driving