Predicting driver’s lane-changing decisions using a neural network model

Lane changing has a significant impact on traffic flow characteristics and potentially reduces traffic safety. However, literature relating to lane changing is not comprehensive, largely owing to the inherent complexity of lane changing and a lack of large-scale data to analyze such behavior. In an effort to cope with these obstacles, this study adopts a neural network (NN) model to capture the complexity of lane changing, and large-scale trajectory data are employed for model estimation and validation. For comparison purposes, a multinomial logit (MNL) model that was frequently accepted as a framework for lane changing in previous studies is also built. Although for non-lane-changing samples, both models perform well in model estimation and validation processes, for lane-changing samples, there are significant differences in their performance. The NN model is able to correctly predict 94.58% of left lane-changing samples and 73.33% of right lane-changing samples in the model estimation process, whereas the percentage correctly predicted by the MNL model is only 13.25% and 3.33%, respectively. While the accuracy of both models noticeably drops in the model validation process, prediction results in the NN model are still acceptable. Finally, the impact of heavy vehicles on driver’s lane-changing decisions is quantitatively evaluated using the sensitivity analysis of the proposed NN model.