An interpretation framework for autonomous vehicles decision-making via SHAP and RF

Decision-making for autonomous vehicles is critical to achieving safe and efficient autonomous driving. In recent years, deep reinforcement learning (DRL) techniques have emerged as the most promising way to enable intelligent decision-making. However, DRL with ‘black box’ nature is not widely understood by humans, thus hindering their social acceptance. In this paper, we combine SHapley Additive exPlanation (SHAP) and random forest (RF) techniques to bring transparency to decision-making obtained by DRL. Specifically, we first implement decision-making of autonomous vehicles following in discrete action space based on DRL algorithm with the goal of safety and efficiency. Then we use the SHAP technique to simplify the feature space, which shows that relative distance, longitudinal speed of the ego vehicle, and longitudinal speed of the proceeding vehicle have a critical impact on vehicle following task. Finally, we collect the state-action pairs generated by the DRL model and perform feature filtering, and fit the decision model with an interpretable RF model. The simulation results show that the RF model achieves the behavioral explanation of autonomous vehicle following.

I. Introduction

In recent years, autonomous driving (AD) has attracted a lot of attention benefiting from the boom in artificial intelligence technology [1], [2]. In particular, the AD decision-making module receives the surrounding traffic information from the perception to make safe and efficient behavioral decisions that are then fed to the planning and control to enable AD [3]. Finite state machines [4], [5] are currently the most widely used behavioral decision models because of their simplicity and ease of implementation but ignore the dynamics and uncertainty of the environment. In addition, the division and management of states are tedious when there are many driving scene features, which are mostly applicable to simple scenarios and difficult to perform the behavioral decision tasks in complex road environments with rich structured features [6].