Many techniques have been tried to improve the accuracy of the assignment process in travel demand models. As models are being used for more purposes, it is desirable to improve the accuracy of the link and turning movement assignments.

One problem that contributes to the error to an extent on links and more on turning movements is the zone structure and accompanying connections of the zone centroid to the network. A typical zone system using census boundaries furnishes zone boundaries on the arterial street system. All land areas are encompassed by streets or major physical boundaries. The zone centroids are located somewhere near the centers of the zones and zone connectors are used to connect these centroids to the modeled network.

The assignment algorithm uses a least cost/shortest path algorithm. The shortest or least cost path algorithm uses the centroid connector as a link in the network. Therefore, trips in and out of a zone will use this path. Depending on the geographic location of the land use within a zone, the subsequent assignment of traffic may not ideally reflect where the trips should be loading on the network.

If we look at an example of a rectangular zone bounded by four (4) arterials as shown in Figure 1. There are four zone connectors connecting the centroid, designated as node A, to each of the arterials. Assume that this zone is also located on the northwestern edge of the model. Trips to and from the southeast area of the model will use either the southern or eastern centroid connectors, designated as A-G and A-E. These trips will travel through the intersection on the southeastern corner of the zone designated at node F.

In this standard type of assignment, these trips won't usually be assigned through any of the other three intersections, designated at B, D, and H. In actuality, one would expect trips to or from the zone to show up on each of the four arterials. However, in this case the trips are only present on East Avenue and South Street.

In a model with approximately equal densities of land use in the zones this problem is usually not observed. This is because the adjacent zones to the north and west will contribute trips to the links on the north and west and the other intersections. The assignment error will tend to average out if the numbers and types of trips in to and from each zone are close to equal or the zone is not on one of the model edges.

The problem is really only noticed when the land use within the zones and sizes of the zones are not equal. It is even more noticeable when tracking the trips from a selected zone to analyze the impacts of growth.

MPA allows the modeler to specify multiple points of access for each zone. During the assignment process, these points act like centroids in that trips begin and end at each of the points. A point of access can be any node in the system. There is no limit to the number of points of access that you can specify for each zone. A point of access can be used for one or more zones.

This can be especially helpful if you have the typical census based zone structure with arterial street boundaries. You can specify mid-block locations for access with proportions for origins and destinations for each point for each zone. Because the assignment begins at more realistic points of access the turning movements are more discrete and the result is better link and turn movement assignments.

Using the same example, but with MPA you get an assignment as shown in Figure 2. This yields four points for trips to begin and end for this zone. Trips to the southeast begin on each of the links, shown as points C, E, G, and I, and travel will be through all of the intersections except for the intersection in the northwest corner as B. This is much closer to the way the trip makers would choose their paths to the southeast.

The result is similar to having many times the number of zone centroids, but without the accompanying zone disaggregation or splitting effort and resulting computation times and data storage requirements.

Select link, zone, and trip purpose features can be used to show the travel patterns throughout the network. In addition, select link, zone, and trip purpose turning movements can also be saved and analyzed. Your model can be refined to match travel behavior even closer than before.

We had mentioned that MPA also gives you the ability to differentially assign the trips to the different access points based upon land use densities. The designation of the access points and the percentages for origins and destinations is defined in an equivalency file.

Using the same example zone, percentages of origins and destination are assigned to each of the access points as shown in Figure 3. The retail to the north and the dense single family development on the western portion of the zone are assigned greater percentages.

This is coded in section 2.7.9. MPA Equivalency file creator. Here, the zone centroid, the percentages and the access points are defined. Entries are created for each zone and each access point. Designations for this example are shown in the following table:

It is important that each zone be defined. If you are not sure of the percentages you can start with some automatic apportionment tools available in section 2.7.9.2. You must have an equivalency for each zone, or it will be ignored in the assignment process.

Normally the origin percentages and destination percentages should each sum up to 100 percent. However, this can also be a tool to account for some differences that are not accounted for in the trip generation process. If you use totals different than 100 percent you must have logical, rational reasons for any differences.

A link file containing the MPA percentages can be exported so that you can plot it and check it visually. A

centroid connector is NOT required between the zone centroid and the point of access. However, each centroid must always be connected to the network with at least one link. The shortest path connection time is still used in the distribution (gravity model) portion of the model run.

Like all model rules, the base year assumptions should be questioned and revised as necessary for the forecasts. Local knowledge and team input is highly recommended.

Watch for a forthcoming paper showing model comparisons. Preliminary review shows the R² increased and the RMSE dropped for each model with the MPA.