Photo Credit: (Brian van der Brug / Los Angeles Times)

Congestion pricing – used on highways in Southern California and in cities such as Singapore and London – requires understanding how much drivers would be willing to pay to travel faster. Technically, this is called travelers' value of travel time (VOT), and it is fundamental not only for congestion pricing, but also for many analyses of transportation projects. Most existing methods for estimating VOT rely on survey data. Yet those data – surveying persons usually at one time and place – may not be suitable for dynamic pricing which requires understanding how VOT varies across different times and traffic conditions. Streaming traffic data consisting of travel times and traffic volumes, on the other hand, is relatively abundant and can potentially provide insight into travel decisions, and hence VOT, more readily. In this project, we developed an adaptive pricing algorithm to estimate VOT from observed traffic data and then investigated the efficacy of the algorithm in human subject experiments.  
  

The adaptive pricing algorithm iteratively updates tolls on routes according to traffic data observed in response to past tolls. We estimate traveler VOT by choosing the VOT that best fits the observed traffic and toll data. We show that setting toll differentials to be proportional to observed traffic volumes and a small additive randomization renders just enough resolution to the resulting traffic data to estimate VOT accurately.   

We compared our algorithm against an existing adaptive stated preference (i.e., survey) method for VOT estimation in human subject experiments on computer screens. The stated preference survey asked subjects to choose between a tolled and an untolled route for different combinations of tolls and travel time savings. Our algorithm, on the other hand, was implemented in the context of a traffic assignment experiment, where the subject was asked to role play an operator who assigns fleet vehicles to three routes based on different values of tolls and travel times on the routes. The subject was instructed to be fair to all the vehicles in the fleet to encourage the assignment to resemble user equilibrium. The VOT estimates obtained by our algorithm were found to be lower than the one estimated by the adaptive stated preference method for overwhelming majority of experimental subjects. Yet our algorithmically estimated VOT also was more effective in minimizing average travel times.  

The findings from this project have several practical implications. Our algorithm can be utilized to estimate VOT from traffic data even when there is uncertainty in supply side parameters such as effective road capacity. All that is needed are data on toll levels and traffic – available from in-road sensors and historical toll information. Our algorithm gave different and often lower estimates of VOT, yet tolls based on our algorithm managed congestion better, suggesting the importance of re-examining how we calculate dynamic tolls. We suggest that the data-gathering framework for VOT estimation should be as close as possible to the context in which it will be used, and our results from analyses of tolls and traffic flows promise a new approach. More information can be found in the final report on the METRANS project “Dynamic Incentive Design for Transportation Systems with Unknown Value of Time”, funded by the Pacific Southwest Region UTC and available to view here.