How Much Convergence Is Enough for Traffic Assignments Used in Feedback?

Equilibrium travel demand model systems solve for consistency between travel times across a network used by models of travel demand, and the user-equilibrium travel times from assignment of this demand to the network, by means of some form of “feedback” iteration. Within this outer loop is an inner loop of traffic assignment, itself an iterative process solving for consistency between times traveled on the network and on the shortest paths between each origin and destination. This study examines how fine a convergence of these inner loops is sufficient to not prevent, limit, or delay convergence of the outer loops, while avoiding excessive run-time from superfluous over-convergence of tentative solutions. Test results from several Frank-Wolfe assignments varying in congestion level, link flow-delay relations, and region found a fairly consistent relation, within an order of magnitude, between assignment relative gap and a comparable measure of the errors of the resulting origin-destination travel times, compared to those of well-converged assignment. Assignments of several of the same cases with origin-based, projected gradient, or other advanced methods showed larger travel time errors for a given relative gap, with wider variation. In test runs of a feedback model, these results was used to choose adaptive stopping criteria automatically for pre-final assignments, which were lenient in early feedback iterations, and finer as the run progressed. Substantial assignment run-time was saved compared to constant stopping criteria, with insignificant harm to the feedback convergence progress.

Language

  • English

Media Info

  • Media Type: Digital/other
  • Features: Figures; References; Tables;
  • Pagination: 22p
  • Monograph Title: TRB 93rd Annual Meeting Compendium of Papers

Subject/Index Terms

Filing Info

  • Accession Number: 01520033
  • Record Type: Publication
  • Report/Paper Numbers: 14-3213
  • Files: TRIS, TRB, ATRI
  • Created Date: Mar 26 2014 10:13AM