Traffic Assignment in General (but intersection delay in particular)

2 posts / 0 new
Last post
Traffic Assignment in General (but intersection delay in particular)

Alan's observation last week about "path choice" (i.e., the actual link to link to link path that modeled and real-world drivers make through a road network) is a good one, and has led to some interesting off-the-listserv conversations with Alan Horowitz, Sam Granato, and others about not just "path choice" (which gets even messier because of the nature of the equilibrium assignment process, as well as the real-world situation where all auto and truck drivers going from point A to point B at a particular start time will not always follow the exact same path) but the issue of using "average link speeds" for links that have a signal at the downstream end as the starting point for validation-related testing, where drivers making left, through and right movements at that downstream signalized intersection will experience different delays. I still think average speeds/times on links (where the averages include all vehicles, regardless of what is their next link to traverse) are useful as part of validation tests, but this may be one of those "necessary but not sufficient" tests if someone is "really" trying to get a traffic assignment "right."

I highly suspect this "gets complicated rather quickly" issue is a major part of the reason why simple assignment approaches (e.g., representing intersection delay via a simple look-up table, plus coding of specific turning movement prohibitions, but not much more than this) will continue to be found in regional models for many years to come, particularly those models that are not "really" being directly used to provide information to decision-makers who are trying to figure out exactly what tolled and/or non-tolled changes (investments) to make to an existing road system, which would include new and/or modified intersections.

While the idea of "one travel model to support all possible decision-making purposes" sounds appealing, it is not so clear that this is generally practical, particularly in an environment where the forecasters in large regions don't want to have lots of really small zones because of the additional model run time involved, plus there is often a recognition (although not so often publicly stated) that the model-produced truck and non-truck trip tables even for an observed year may not be a good representation of actual truck and non-truck travel patterns, and goodness gracious as to the likely accuracy of 20 year-plus tables. So garbage going into a wonderful region-wide network assignment procedure with perfect representations of intersections and 5,000-plus zones is still going to be garbage out.
Ken C.

From: Alan J Horowitz []
Sent: Wednesday, July 22, 2020 2:04 PM
To: Cervenka, Kenneth (FTA) <<>>; TMIP <<>> Subject: Re: [TMIP] Intersection Delay in Travel Demand Model


Regarding your "Regarding #2" comment, second paragraph, NCHRP Report 765 reiterates multiple times that good delay estimates are essential to project-level forecasts. Those better delay estimations methods may be incorporated into regional models, hybrid models or subarea models, as appropriate. Ideally, if an agency wanted to use regional-model outputs for project-level work, that regional model should have a dynamic traffic assignment or another equally good way of calculating delay at controlled intersections and at other locations where substantial queueing can occur.

We cannot necessarily assume that good OD travel times lead to good path choice estimates. Path choice involves link/turn-level decision making.


From:<> <<>> on behalf of KenCervenka <<>> Sent: Wednesday, July 22, 2020 10:12 AM
To: TMIP <<>> Subject: Re: [TMIP] Intersection Delay in Travel Demand Model


Consider these as nothing more than personal observations, which are presented in the hopes this may stimulate some clear thinking on the topic.

I think two of the reasons why regional travel models often wind up with traffic assignment procedures that may be over-simplified in their representation of link-to-link-to-link auto or commercial vehicle travel times between two zone centroids are:
1) The lack of meaningful model validation/testing procedures (which includes "validation over time" checks).
2) The uncertainties as to "so what to do next" if the testing reveals some big differences between predicted and observed.

Regarding #1:
Let's consider the model-based generation of directional auto travel times for different time periods, for an observable year: there is lots of "Big Data" information available these days, that should enable any MPO (or any metropolitan area in the world) to produce a fairly robust database of average observed link travel times that represent what's happened on "typical" days (i.e., days without any big special events happening, or other delays caused by traffic accidents). This opens up opportunities to compare modeled vs observed travel time contours (isochrones) to and from specific points of interest for different time periods of a weekday, as well as for other aggregations of a series of links, e.g., roadway segments that could be several miles long.

Regarding #2:
If some big differences are found in #1, this opens up a need to get a clear understanding of the underlying causes. For example, it "might" be a direct result of a poor representation of intersection approach capacities, but it also might be a direct result of major inaccuracies in the OD vehicle trip tables that are input to the traffic assignment module, and the subsequent traffic volume assignments. So having a very robust traffic assignment procedure in place may not necessarily appear to have any clear connection to "better accuracy" in the prediction of traffic volumes and travel times, if the OD matrices are inaccurate.

Someone please correct me if they really think I am wrong, but the "solution" that is sometimes implemented is to not even directly use the outputs of a regional travel model for any sort of traffic operations-focused study, but instead develop a subarea-based assignment model (DTA, or even a "full" microsimulation) that has a greater level of intersection detail. The zone-to-zone vehicle trip tables from the regional model "might" be used as a seed matrix in some form of "matrix estimation" exercise that "forces" a close match between observed counts and the assigned volumes on links, but that is fraught with dangers because while the assigned volume-to-count and travel time validation results may look good, the underlying vehicle trip tables could still be way off from reality. The same concern may apply to use of Big Data-derived zone-to-zone trip tables, that can also be subjected to a similar "OD matrix estimation" routine that will "force" the assigned volumes to compare closely to the available counts (even flawed counts), but this doesn't mean the resulting OD vehicle trip tables are automatically a good-enough representation of the real-world OD tables.

Quite a dilemma, but regardless of how one attempts to find workable solutions, it seems a bit silly to not take a "really good look" at what the current models are doing well and not-so-well, when it comes to their prediction of auto and truck travel times. And since any model can be "adjusted" until it appears to fit the known reality fairly well, to then take this even one step further, which is to see what the current models are doing well and not-so-well, when it comes to their prediction of changes in auto and truck travel times (as well as auto and truck directional volumes by time-of-day), from one year to the next.

Ken C.

From: Xinbo Mi
Sent: Wednesday, July 8, 2020 11:21 AM
Dear Prof. Horowitz,

As a recent graduate with 2 years' experience in the area, I have only played with one model so far. My vision is still too limited to represent the state of the practice. That is why I am asking around how others are dealing with a specific issue in their models.

I have been actually trying developing our model from scratch recently. For the network initialization stage, I coded the free-flow speed and capacity exactly matching the HCM 2016 procedures for Freeway, Multilane Highway and Two-Lane Highways. However, when it comes to Urban Street Segments, it just becomes too complicated with all those intersection lane configuration and signal timing plans as input data required. In the first place, I thought using HCM-based calculation in a macroscopic model is too much work so I did not include it as a choice in my initial post. However, like what you indicated, there are quite a few responses in my questionnaire that uses detailed HCM calculation (maybe partially for major intersections), which surprises me more or less. While, there are still quite a few responses do not include intersection delay at all or just using a generalized turn penalty table.

Of course, the HCM calculation will be the finest way in detail to get the most accurate delay, but to some areas that haven't built a robust connection between signal timing update and modelers systematically, I was trying to find a way to calculate intersection delay somehow between "doing nothing" and "going fully to HCM".

Another idea I always think about is that whether we can create a signal plan file compatible with all using purpose and all software, something maybe have extension of ".sig", so that people can simply pass around the files from the original equipment (e.g. controller installed at an intersection) through the intermediate users (e.g. signal controller people) to the end users (e.g. travel demand modeler). That will make life much easier.

Again, any discussion and feedback will be extremely beneficial to me. Thanks!

FYI, questionnaire link:


Xinbo Mi

From: ajhorowitz
Sent: Wednesday, July 8, 2020 9:51 AM
Have we reached at point in our thinking where we are totally ignoring doing intersection delay correctly, with consideration for conflicting and opposing traffic? Has the state of the practice become entirely the use of VDFs and associated fudges?

From: Xinbo Mi
Sent: Tuesday, July 7, 2020 2:09 PM

Dear all,

I am wondering how people are dealing with intersection delay in their travel demand models, especially for signalized intersections.

While I was doing some literature review recently, I found pretty decent simplified volume-delay functions considering volume-dependent signalized intersection delay, which requires only green time and cycle length as extra input data, which can be even simplified by assigning default values based on functional classes comparison at an intersection. However, it seems they are not widely used in travel demand models in reality, or maybe just my misconception.

Therefore, I would like to reach out to everyone to know more how your models work as far as handling intersection delay. Following options may be considered:
1. Not at all. Only using link capacity and link volume-delay function
2. Using generalized default turning movement delay values for all nodes, regardless of link volume
3. Using a generalized default node delay value for a certain node, regardless approaches and their link volumes
4. Using a specialized volume-delay function for a certain node, regardless approaches
5. Using a specialized volume-delay function for a certain approach, regardless of under-capacity or over-capacity
6. Using piecewise specialized volume-delay functions for a certain approach, part for under-capacity situation and the other part for over-capacity situation

I created a poll containing the choices above. The link is as follows:

My email address is<>

To respond, please feel free to directly respond within TMIP, email me or using the poll, whichever you prefer. Any feedback will be appreciated.


Xinbo Mi

Transportation Engineer

Evansville Metropolitan Planning Organization


I have been goaded into providing a more complete response to the original question.  The result is maybe too long to post here, but I have prepared a blog post titled "Six Short Essays on Highway Path Choice".  The six essays have their own titles:

Where are All the Trucks?  Getting Truck Costs Right is Important

There Are Many Failures of the Standard Model of Path Choice for Automobiles

Turning Delays are Important to Path Choice

Overreaction to Congestion Can Affect Drivers’ Path Choice

Poor Spatial Precision Causes Major Inaccuracies in Path Choice Calculations

OD Table Estimation from Traffic Counts Requires Good Path Choice Calculations

You can find the blog post at: