Master Lane Change Decision-Making by Integrating Imitation Learning with Expert Rules

Imitation learning has been extensively explored in the decision-making of intelligent vehicles, showing a promising human-like driving performance. However, it suffers significant safety challenges when further implemented in industry due to the lack of transparency. Therefore, expert systems currently dominate in the decision module of driving software, whereas the predefined rules inevitably lead to rigid driving behaviors. To benefit from both methods, in this paper we propose to learn the practical lane change decision making by integrating the imitation learning with expert rules. A lightweight Transformer network is firstly designed to extract the comprehensive interaction between dynamic agents and road elements, generating the normalized probability for lane keeping, left lane-change and right lane-change. This network is trained to predict the next-step behavior by imitation learning wherein a massive dataset of drivers is collected and annotated to provide the ground truth. Then the expert system checks the multiple outputs of the network within a sliding window to determine whether to initiate or cancel a lane change. With the rule-based state machine attached to the network, the unsafe decisions can be efficiently eliminated by road regulations and collision avoidance. Experiments show that our method can effectively capture experience from human data, automatically identifying suitable lane change opportunities. It can improve the driving speed by 16.15% compared with the expert system with respect to overtake and off-ramp scenarios. Our work provides a more trustworthy and intelligent driving system for practical vehicles.