GAMESlab

Availability of individual-level longitudinal data provides the opportunity to better understand travelers’ day-to-day learning behavior, enabling more accurate predictions of traffic patterns in a network with random travel times. In this paper, an instance-based learning (IBL) model that can capture the recency, hot stove and payoff variability effects embedded in travelers’ day-to-day learning processes is developed for route-choice based on the power law of forgetting and practice. Experiments based on synthetic datasets show that the true parameter values of the IBL model can be consistently retrieved and the model can potentially predict different traffic patterns compared to non-learning models. The IBL model is compared with a baseline learning model using an experimental dataset of repeated route-choice. Estimation results show that the IBL model reveals higher sensitivity to perceived travel time and achieves better model fit. Cross validation experiments suggest that the forecasting ability of the IBL model is consistently better than the baseline learning model. Practical considerations for choice modeling are further discussed.