Route Set Generation Within a Dynamic Modeling Framework: a Comparison of Methods and GPS Route Choice Data
J Zantema, D H van Amelsfort, TU Delft, NL
This paper investigates different route set generation methods for a dynamic model. A Monte Carlo approach that combines higher and lower level route alternatives performs best in the model. The validity of route sets is verified using GPS data.
To assess the network effects of road pricing measures, a dynamic modeling framework was developed, which determines a route choice, departure time choice and elastic demand equilibrium. A macroscopic dynamic multi-user class model called INDY forms the center of this modeling framework. INDY consists of three parts: route generation, route choice and dynamic network loading. In order to correctly model the route choice behavior, the quality of the route generation process is of vital importance. If important routes are not generated in this process, than they will never be used in the assignment. Thus the number of routes needs to be sufficiently large. However, larger route sets increase the computation times drastically which is an undesired side effect. This paper explores the different methods that are available for route generation, e.g. Monte Carlo approach and static equilibrium assignment and compares the resulting route sets to actual route choice behavior measured from a GPS survey. Based on the analysis a two-step Monte Carlo approach is recommended which excludes a large number of irrelevant routes while it does take into account principal routes between origin and destination as well as smaller route changes around the origin and destination.
Two approaches of route set generation are possible in INDY, a Monte Carlo approach and an approach using a static traffic assignment. The static equilibrium assignment does not take into account junction delays and it seems to underestimate delays because of the lacking blocking back. The static assignment does not seem to produce a satisfying route set without becoming too large for practical application.
The Monte Carlo approach uses two parameters: the number of iterations and the maximum path overlap. Different route sets were created using these parameters. A large overlap parameter generates routes that differ only slightly and the differences mainly occur around the origin and destination of the trip. This does not provide a good route set. Using a small overlap generates routes that are substantially different, but this causes problems in urban networks where traffic need to be spread over multiple routes with smaller diversions to model congestion correctly.
Number of iterations Maximum overlap Highway variant Interurban variant Number of paths Total paths
4 0.95 Yes No 3 85k
30 0.5 Yes Yes 2 82k
30 0.4 Yes Yes 2 63k
20 0.7 Yes No 2 90k
Combined route set Yes Yes 3 45k
Table 1 Overview of path table characteristics for one OD pair
As seen in table 1, path tables consisted of many paths, already leading to long calculation times. The existence of variants and the number of paths are taken for one specific OD pair. Sets that were created with a high maximum overlap missed the interurban variant, while low overlap variants had much delay in the urban area, as traffic jammed as soon as city level was reached. This while parallel roads in the city, which detour would not give much delay, weren?t used at all.
Using an average approach on route set generation would not give satisfying results either, so chosen was to create a route set using a combination of both methods. At first a route set was created with a low allowed overlap and many iterations, though fewer then in a single set, to catch the interurban routes in the network. Second a route set was created with high overlap and very few iterations, in this way some urban choices were made available for all routes. These two sets were in size both smaller then the first sets used. Added together they would give choice on both urban and interurban level with acceptable calculation times.
Looking at possible routes, the combined set gives much more satisfying results in the sense that it has routes for each part of the network. When congestion occurred on the highway, there was the possibility of taking the interurban roads, and on city level, traffic was divided so that all roads are used equally. Thus showing that important routes are captured, while most irrelevant routes are left out.
The project for which the calibration of the network was needed was the ?SpitsMijden? project, which is a negative pricing program in order to give an incentive for people to avoid the morning peak. In the pilot program, GPS data was obtained for those participants who?s reward would be a route planner with online traffic data. The routes chosen by these 109 participants are being reconstructed and determining the overlap with the route sets will be finished April 2007.
Association for European Transport