Traffic Assignment and Feedback Research to Support Improved Travel Forecasting

Traffic Assignment and Feedback Research to Support Improved Travel Forecasting


Howard Slavin, Caliper Corporation


This research investigated whether or not current traffic assignment and model feedback practices are sufficiently accurate for calculating congested highway travel times and for quantifying the highway benefits of major transit projects.


This research study had the objective of assessing whether or not current travel demand modeling practices are sufficiently accurate for calculating congested highway travel times and for quantifying the highway benefits of major transit projects. This question is closely associated with the effectiveness of the highway traffic assignment procedures that are is used to predict traffic flows and congested travel times in regional travel demand models. In prior research, insufficient convergence of traffic assignment models was shown to lead to unreliable estimates of project impacts.
The study was performed in two phases. In the first, an inventory and review of the modeling practices employed at the 30 largest metropolitan planning organizations (MPOs) in the U.S. was performed. In the second, 5 of the better MPO models were selected for in-depth examination. These models were used to study their ability to measure the road traffic impacts of highway and transit improvement projects and to experiment with more stringent criteria for convergence. One highway project and one transit project from each region was evaluated using regional models with various improvements in computational approach. We also compared model-based estimates of congested travel speeds with measurements of travel speeds from commercial sources. Data from HERE, INRIX, and Google were utilized for that purpose.
The review of network modeling and feedback practices in use in 2011 by the 30 largest MPOs in the U.S. indicated that deficient methods were in widespread use. Many MPOs used inappropriate assignment algorithms and/or incorrect closure metrics. There was not a single MPO that performed feedback for each and every time period and also used tight closure criteria for feedback.
Working with the 5 selected models, we found that the maximum link convergence error was around 1000 vehicles for an AM peak period at a relative gap of .0001. At the .01 (1%) relative gap that was traditional historically and is still used in some MPO models, the maximum link flow error for the AM peak period was between 5000 and 10000 vehicles.
With respect to feedback practices, the empirical work established that a closer tolerance between input and output travel times can be computed than is currently being sought. It also established that the final answers in terms of link flows will be different with different stopping tolerances and that feedback methods can give the illusion of convergence when small changes in the same direction for each loop ultimately add up to larger differences in traffic flows after many loops.
With respect to traffic assignment convergence, a major finding was that very tight assignment model convergence was needed to assess project impacts. For some projects, convergence to .000001 (1.E-6) relative gap was necessary to get a plausible estimate of project impacts that was free from spurious artifacts. Importantly, using tightly converged assignment models, we found that it is possible to estimate the road congestion relief benefits of transit improvement projects as well as to estimate the travel time benefits of highway projects
Using data from several commercial sources, comparisons of modeled and measured speeds revealed that, in general, the travel demand models did not produce congested travel times that were in good agreement with independent measurements. For 4 of the 5 MPOs, the overall model travel speeds were slower than the measured travel speeds. The one MPO that has made more extensive use of speed data actually achieved a fairly close match between the model speeds and the reported measurements.
Given current practices, we concluded that validation should be accorded a greater priority in the model development process. It should be disaggregate in nature, and traffic assignment models should be validated at the link level by time period and direction. A sufficient number of directional counts by time period should be obtained by functional class to be statistically valid. In addition to link volumes, measured speed data should be used as part of the model development and validation process. Otherwise, estimates of vehicle hours of travel are not likely to be very accurate. Before-and-after studies of project impacts are also suggested to assess the external validity of travel demand models and their forecasts


Association for European Transport