Where are All the Trucks? Getting Truck Costs Right Is Really Important
Several years ago, I was working with FAF (Freight Analysis Framework) truck assignment results for the upper Midwest. Those early FAF assignments routed trucks according to travel time. I noticed lots of trucks being assigned to the Chicago Skyway. The Chicago Skyway is a short toll road that runs from the south side of Chicago into northern Indiana. Not long afterward I had occasion to drive on the Chicago Skyway en route to Ann Arbor from my home. As I am driving along, I exclaimed to my wife, Shirley, “Where are all the trucks?!” There was not a truck in sight. It was clear to me then, and also after subsequent research on data from Ohio’s toll roads, that truckers are quite sensitive to tolls in their path choice decisions.
Adhering to an accurate estimate of truck costs is an imperative for doing any worthwhile freight forecasting at the road-segment level. Truck costs are not all that easy to calculate, so I recommend using something like the CFIRE Truck Cost Model, which contains everything needed.
There Are Many Failures of the Standard Model of Path Choice for Automobiles
There was a slide in the training course for NCHRP Report 187 that pretty much summarized the standard model of path choice for me. I had three pictograms for
TIME + COST + DISTANCE
The pictograms were a clock, a dollar bill and a ruler.
I always explained to my students that automobile operating costs were correlated with time and distance and could largely be ignored. “Costs” were tolls, for the most part. Distance correlated with time, which meant distance could be ignored, also. Values of time varied by driver’s income, so if the software is sufficiently capable, driver-class segmentation could be helpful when there are tolls.
Except I was confronted with all kinds of situations where the standard model did not work. Users of my software reported consistent over-assignment of freeway links. People either did not properly evaluate the travel times on freeways or they considered freeway driving to be inconvenient or stressful. Users also reported over-assignments on big bridges, regardless of tolls. I had questionnaire results showing people overreact to travel times in congested traffic and will avoid known bad pavements. People consistently fail to divert, when they should, from highway work zones. One software user told me he even needed to adjust for curvy roads. These are all “systematic” rather than “random” errors in the standard model.
The early literature on path choice made reference to many “random” influences. Drivers presumably make mistakes in their travel-time perceptions. Drivers have preference for different scenery. Drivers sometimes need to make unusual stops along the way, taking them off their preferred route. These “random” influences could be accommodated, I was told, by multipath stochastic traffic assignment techniques, which simply looked for second, third, fourth, etc. options and threw some trips at them, as well.
There are other random effects, which are under-appreciated and correctable. Models are known to have serious errors in travel time estimates along links, at bottlenecks and at intersection movements. Big errors stem from the common assumption that all trips for a zone originate or terminate at a centroid, a single point within the zone.
However, a lot of modelers find stochastic methods to be redundant. User-equilibrium (UE) traffic assignment is inherently multipath. In congested portions of a network, trips are spread out across lots of different paths. But a UE assignment does little to help rural areas and other uncongested parts of networks.
Stochastic assignment has some behavioral underpinnings, but it is mostly a kluge. It improves some traffic assignments by wallpapering over a lot of flaws in the standard model, but it does little for those systematic errors I mentioned earlier.
The standard model of path choice fails in a number of predictable ways. It is far better fixing those problem areas, where possible, than candy coating them with network theory and swallowing hard.
Turning Delays are Important to Path Choice
Drivers are sensitive to delays at intersections when choosing paths. Delays on urban arterial streets are caused, primarily, by intersections. Turning delays are caused, primarily, by opposing and conflicting traffic, by the type of type of traffic controls, and by how those traffic controls are set. The relationships between delays, traffic levels, and traffic controls are pretty well understood. These relationships are still mostly ignored in large regional models in the US.
Clever modelers will often add penalties to left turns at intersections, but those penalties are static and nearly impossible to forecast well into the future without a separate traffic model. While we can hope that these rather sizable errors will be unbiased and wash out during the equilibrium process, it is more likely they will introduce substantial mistakes to the path building calculations.
In addition, correct traffic control delay relationships have different, usually larger, sensitivities to traffic increases than those VDFs (volume-delay functions) typically found. A direct effect of a relatively insensitive VDF is insufficient diversion from badly congested facilities. Assignments will be wrong, and MOEs created from VDFs will likely grossly underestimate the impacts of traffic growth scenarios.
If you are keenly interested in having accurate future estimates of an environment impact, such as carbon emissions, you should be very skeptical of what is coming out of most MPOs these days.
Nobody should trust any project-level results from regional models that do not have reasonably correct delay relationships.
Overreaction to Congestion Can Affect Drivers’ Path Choice
In the late 1970s, I conducted a psychophysical scaling experiment in Chicago to determine drivers “subjective value of time”, using the technique of “magnitude estimation”. In one part of the experiment drivers responded by giving large subjective values of time to driving in congested traffic, similar to those seen for transit out-of-vehicle time. It is generally assumed people do not like being stopped or slowed when traveling. Another explanation is people are unhappy about the uncertainty of an on-time arrival when there are traffic delays. We know travel times vary more as traffic becomes more congested. We also recently learned drivers place a fairly large weight on uncertainty when making travel choices.
One method, perhaps, of modeling drivers’ overreactions to congestion is to add an uncertainty term to link impedance. But adding an uncertainty term makes sense only if the travel times are correctly estimated in the first place. Typical regional models, which eschew traffic dynamics or use multi-hour peak periods or rely entirely on VDFs, will perform badly in estimating congestion delays, thereby undercutting any improvements once might get from including uncertainty.
Simply adding uncertainty to impedance dooms most models from finding true shortest paths between origins and destinations. I am mostly OK with this. It is better to have the impedance roughly correct than to get the path building exactly correct. There are good, efficient algorithms for finding shortest paths with uncertainty in the impedance, but most MPOs don’t have access to them.
But I am not OK with sloppy path building in technical papers submitted to the scientific literature.
Poor Spatial Precision Causes Major Inaccuracies in Path Choice Calculations
You may or may not remember the Lexington data set. In the very early days of GPS, that is 1996, FHWA sponsored the collection of path trajectories of vehicles from 100 households in Lexington KY. I asked a graduate student, Oliver Jan, to look at variations in path choice within that data set. When origins and destinations were exactly similar, paths were mostly similar. When origins or destinations were in slightly different locations, paths could be vastly different. It seems, path choice is very sensitive to the exact locations of the trip ends.
I probably did not need the Lexington data set to figure this out, had I been looking for it. All one needs to do is to compare shortest paths from different places in a zone other than the centroid. In a subsequent experiment, I compared a standard centroid-based assignment with a spatially-disaggregated assignment, where every trip’s path began or ended at the nearest intersection. The differences in the assigned link volumes were huge. Clearly, a major source of error in travel models is the use of zones that are too big. Ideally, at least for the traffic assignment step, zones should be no larger than the catchment area of a single intersection node.
Increasing spatial precision is especially helpful in rural areas and small cities where the multipath nature of UE assignment does little to remove lumpy link loadings.
OD Table Estimation from Traffic Counts Requires Good Path Choice Calculations
Computed paths are of little interest to planners, who care much more about computed volumes, turning movements, and delays. My software, in order to reduce run times and storage requirements, does not even save computed paths.
However, OD table estimation methods need to know how many trips between any origin and any destination use any particular link. To get this information, the software briefly needs exact paths between OD pairs. So good path choice modeling is essential to good OD table estimation.
Turning movement data could greatly improve OD table estimates, but most cities do not have very much high-quality turning movement data. I have recommended, instead, to employ whatever movement data there is to validate the OD table. If the assignment gets the percentages of lefts and rights somewhat correct, then OD table can’t be too awful.
People doing OD table estimation without good path choice assumptions will surely be disappointed
I have heard from network theorists that paths in UE assignments are not unique. That is, it is possible to satisfy Wardrop’s first principle equally well with two or more entirely different sets of paths. You could force a model to create a different set of paths by, perhaps, starting the assignment with free travel times instead of LOS C travel times. It is not clear to me whether it makes much difference, given all the other sources of error in the estimation process.
Closure
A good understanding of path choice is essential for a good travel forecast. Unfortunately, there are a lot of travel models failing to understand path choice.
Alan Horowitz, Whitefish Bay, August 3, 2020
Thanks for writing this up, Alan! We love hearing of your specific examples for route choice & prediction issues.