An Explainable Q-Learning Method for Longitudinal Control of Autonomous Vehicles

Various artificial intelligence (AI) algorithms have been developed for autonomous vehicles (AVs) to support environmental perception, decision making and automated driving in real-world scenarios. Existing AI methods, such as deep learning and deep reinforcement learning, have been criticized due to their black box nature. Explainable AI technologies are important for assisting users in understanding vehicle behaviors to ensure that users trust, accept, and rely on AI devices. In this paper, an explainable Q-learning method for AV longitudinal control is proposed. First, AI control of AVs is realized by constructing a deep Q-network (DQN) with an intelligent driver model, with the control objective maximizing vehicle speed while preventing collisions. Then, a deep explainer for humans is developed via a Shapley additive explanation (SHAP), and a novel positive SHAP method that defines new base values is proposed to explain how individual state features contribute to decisions. Finally, statistical analyses and intuitive explanations are quantified based on SHAP tools to improve clarity. Elaborate numerical simulations are conducted to demonstrate the effectiveness of the proposed algorithm. The code is available at https://github.com/limeng-1234/Pos_Shap.

I. Introduction

Automated driving (AD) of autonomous vehicles (AVs) is considered as the main future trend in intelligent road mobility to reduce human labor and costs and enhance safety and reliability [1], [2], [3]. Recent academic publications have demonstrated the success of DRL in training AI agents to drive better than experienced human racers [4] or in developing intelligent environments (background vehicles) to accelerate testing and evaluation processes [5]. Despite recent technological advancements, the successful deployment of DRL-based AD techniques in real-world environments depends on users’ acceptance and trust. Due to reports on AV accident cases [6], [7] and the black box nature of DRL neural networks (NNs) [8], people are reluctant to accept the use of AVs. Therefore, effective approaches to increasing public confidence in DRL-based decision-making and control applications are necessary.