Can random regret minimization models predict traveler choices on freeways with managed Lanes?

What the paper says

• Examined the ability of several choice models to accurately capture real-world traveler choices on freeways with managed (express) lanes. • Models included multinomial logit (MNL), nested logit (NL), and cross-nested logit (CNL) models combined with both random utility maximization (RUM) and random regret minimization (RRM) frameworks. • The models performed well and were remarkably similar in the best independent variables to include, the coefficients of those variables, and the accuracy of the models. • Travelers who always used the general purpose lanes tended to pay more consistent toll rates and traveled less frequently. Travelers who always used the managed lanes showed the opposite pattern. Surprisingly, travelers who had knowledge of slower managed lane performance for a higher proportion of their trip history were less likely to always use the general purpose lanes. Managed lanes (MLs) typically offer faster and more reliable travel compared to adjacent toll-free general-purpose lanes (GPLs). Successful implementation of MLs requires accurate modeling and a comprehensive understanding of travel behavior. However, individual lane decisions are often more complex than the economically rational decision-making processes that focus on toll rates versus travel time savings. Empirical evidence revealed that many travelers always choose the same lane type regardless of toll or travel time savings. Travelers who always pay a toll to use MLs are classified as ML non-choosers, while those who always use GPLs are GPL non-choosers. Those who alternate between MLs and GPLs are referred to as choosers. The existence of non-choosers challenges traditional assumptions of rational decision-making and complicates the modeling of toll facility use. Most prior research has applied random utility maximization (RUM) models to predict lane choice, while random regret minimization (RRM) models have not been explored in the context of non-choosers and choosers. Moreover, nested logit structures have not been previously applied in this context. This study addresses these gaps by using multinomial logit (MNL), nested logit (NL), and cross-nested logit (CNL) models combined with both RUM and RRM frameworks to classify choosers, ML non-choosers, or GPL non-choosers and to examine influencing factors. Using over 30 months of Katy Freeway travel data, this is the first case study to employ RRM and nested logit structures in modeling lane choice decisions within the chooser context. Results of the case study indicate that the differences in model performance are not substantial across all models, with an overall accuracy percentage of approximately 66%. For the goodness of fit, RRM-based models performed slightly better than their RUM-based counterparts. CNL models slightly outperformed NL models, which then marginally outperformed MNL models. All in all, RUM-based models showed a slight advantage in identifying choosers, whereas RRM-based models performed marginally better in identifying non-choosers. Notably, within both frameworks, the CNL models exhibited better accuracy in classifying GPL non-choosers, outperforming other models by approximately 1.5%. Travelers with higher ML-GPL speed differentials or fewer entry location variations were more likely to be non-choosers. GPL non-choosers tended to pay more consistent toll rates and traveled less frequently, whereas ML non-choosers showed the opposite pattern. Surprisingly, travelers who had knowledge of slower ML performance for a higher proportion of their trip history were less likely to be GPL non-choosers.