Abstract
Deep reinforcement learning (DRL) methods have been widely applied to the discretionary lane change task to improve driving safety and efficiency. However, most of the existing DRL models apply random sampling in the exploration stage, encountering challenges of low sampling efficiency and poor sample quality. To address these issues, we propose a rule-integrated sampling algorithm to enhance the training of DRL models. In the exploration stage, we use a reward prediction model to estimate the reward of action demonstrated by a rule-based lane-change decision model. If the reward is higher than a threshold, the demonstrated action will be applied. Otherwise, we will exclude this action and select another randomly. Analysis of the samples demonstrates the efficiency of our method in providing high-quality lane-change samples, which accelerates the learning of the DRL model. Extensive experiments in a complex simulation environment illustrate that the improved DRL model can make safe, efficient, rational lane-change decisions.