Non-recurring disruptions to a traffic system caused by incidents or bad weather can result inuncertain travel times. Meanwhile, real-time information is and will be increasingly available totravelers so they can adapt to actual traffic conditions and reduce the negative effects ofuncertainties. As uncertain disruptions account for a significant portion of the total traffic delays,it is imperative to incorporate them in a traffic prediction model. This paper presents the resultsof interactive route choice experiments, where subjects made 120 “days” of repeated routechoices in a hypothetical, competitive network subjected to random capacity reductions. Twoscenarios are studied, one with real-time information regarding a probable incident and the otherwithout. Graphic analysis and non-parametric statistical tests are conducted to compare the twoscenarios. Results show that real-time information provides a positive effect on increasing thenetwork efficiency and reliability. A reinforcement learning model is developed to capturesubjects’ route choice characteristics in the information case. Two parameters, a scale factor anddiscounting rate of previous experience, are estimated from the data by minimizing the deviationof the predicted route flows from the observed ones on a “day-to-day” basis with a combinationof brute force enumeration and a subsequent stochastic approximation method suitable foroptimization problems with noises. The learning model accurately captures the “day-to-day”traffic flow evolutions and is a potential alternative and/or complement to an equilibrium trafficassignment model for assessing the impacts of a traveler information system as an integral partof an intelligent transport system.
展开▼
