In the past decade several new algorithms were proposed for accelerating the convergence of user equilibrium assignment models. These algorithms not only held the promise of reducing convergence error in planning models and project evaluation, but also offered the promise of faster computing times. While we and others published convincing evidence of the benefits of these new algorithms, most of that evidence was based on very limited empirical application and testing. In this presentation, we now can report on the application of accelerated traffic assignment models to at least a dozen separate deployed MPO models. These models have the real world complexities of turn prohibitions and penalties and multiple user classes characterized by class network link exclusions, varying passenger car equivalents, and differential tolls.

Our comparisons with prior methods demonstrate conclusively that the new methods work well on a wide range of regional planning models including the largest ones and that convergence issues can be meaningfully addressed leading to more accurate estimates of project impacts. We also explore and illuminate the benefits of multi-threading these algorithms.

As part of our research, we have conducted a new and very broad set of tests on the issue of how much convergence is enough. The results will surprise anyone who has followed the literature on this topic.

We also present new findings on the use of traffic assignments to model toll facilities by using multiple user classes to approximate traveler value of time distributions. A comparison with legacy toll diversion methods illustrates its shortcomings and the superiority of multi-class assignments to estimate toll revenues.

In spite of the generally positive findings we report, we also discuss some of the shortcomings and open questions surrounding new traffic assignment models. Where possible, guidance is offered to model developers in assignment model formulation, calibration, and validation.