Modeling lane changes using parallel learning

Abstract

This paper introduces an innovative approach to model the lane-change (LC) process of vehicles by employing parallel learning, seamlessly integrating conventional physical or behavioral models with data-driven counterparts. The LC process is divided into two distinct steps: the LC decision and the LC implementation, each independently modeled. For the LC decision model, a utility-based model is embedded into a neural network. Simultaneously, the LC implementation model incorporates a conventional car-following model, replicating the behavior of the new follower of the lane-changer, within the training process of a long-short-term memory model. Empirical trajectory data collected from unmanned aerial vehicles, which provides detailed information on the vehicles’ lane-changing process, serves as the basis for training and testing the proposed models. Additionally, data from a different site is employed to assess model transferability. Results demonstrate that the proposed models adeptly predict both LC decisions and implementations, outperforming baseline physical and behavioral models, as well as pure data-driven models, in terms of prediction accuracy and transferability. These findings highlight the significant potential of these models in improving the precision of microscopic traffic simulators.