An Interpretable Deep Reinforcement Learning Approach to Autonomous Driving

Zhihao Song, Yunpeng Jiang, Jianyi Zhang, Paul Weng, Dong Li, Wulong Liu, Jianye HAO

Abstract:

This paper investigates a deep reinforcement learning (RL) approach for autonomous driving. Since interpretability is essential for such a high-stake domain, in contrast to previous deep RL work, we exploit a recent neuro-symbolic model called differentiable logic machine to learn an interpretable controller in the form of a first-order logic program. As a proof of concept, we demonstrate the feasibility of our approach on two classical decision-making scenarios in autonomous driving: lane changing and intersection management. Our preliminary results obtained in a simple simulator suggest that learning an interpretable controller does not penalize performance. Moreover, since
the controller is a logic program, it is understandable and is amenable to analysis.