New ActivitySim Mandatory Location Choice Constraint Mechanism: A Possible Way to Address the Issue of Longer Half-Tour Length

[NFerdous1]

At MWCOG’s Travel Forecasting Subcommittee Meeting on 11/18/2022, Joel Freedman (RSG) presented a new simulation-based constraint mechanism in ActivitySim mandatory location choice models (the presentation is available here - https://www.mwcog.org/file.aspx?&A=KslkdpTuPFRp8%2bpTQndxFJxYaAIs6OFyFS%2bhGAxFlqk%3d). The test carried out on the MTC full model seems to indicate that the simulation method tends to overestimate average work half-tour lengths and average school half-tour lengths relative to CT-RAMP method (see slide 29). Although a re-calibration of the mandatory tour destination choice model after incorporating this new mechanism could address the over-estimation of work/school tour lengths, a simple and intuitive tweak to the simulation method that RSG proposed may better match the observed data and reduce the need for calibration. Here are some thoughts for ActivitySim Consortium to consider in the future:

1. Run all workers through location choice model (same as the simulation method).

2. For each zone, compare estimated workers to total input employment. If zone is overestimated, randomly select workers (equal to the number over the total jobs). These workers will be re-run in the next iteration after removing all over-estimated zones (same as the simulation method).

3. The re-run should be carried out with this implicit assumption that all things being equal, workers prefer work locations that are closer to their homes (different from the simulation method). Several tests can be carried out to identify the most suitable specification. For example:

   • After making the over-estimated zones unavailable, for each of these workers create two sets of available zones – one set that includes zones that are closest to the worker’s home and the second set that includes zones that are closest to the overestimated zone. Test both sets of available zones and assess their performance.
   • When creating a set of available zones, test different sample size. And/or use distance as weight or create a decay function, etc.

If ActivitySim Consortium is interested in implementing such a tweak to the new simulation-based constraint mechanism to address the tour length overestimation issue in the future, MWCOG staff will be happy to discuss this with you further. Please feel free to email Nazneen Ferdous (nferdous@mwcog.org) with cc to Mark Moran (mmoran@mwcog.org) and Feng Xie (fxie@mwcog.org). Thanks.

[stefancoe]

My two cents- I think it will be easier to do some additional calibration. The workplace location model, for example, has several distance based variables. I would start by adjusting the coefficient for the general distance variable, util_distance and then maybe the next 3 (util_dist_squared, util_dist_cubed, util_dist_logged) if needed.

https://github.com/ActivitySim/activitysim/blob/main/activitysim/examples/prototype_mwcog/configs/workplace_location.csv

[NFerdous1]

My two cents- I think it will be easier to do some additional calibration. The workplace location model, for example, has several distance based variables. I would start by adjusting the coefficient for the general distance variable, util_distance and then maybe the next 3 (util_dist_squared, util_dist_cubed, util_dist_logged) if needed.

https://github.com/ActivitySim/activitysim/blob/main/activitysim/examples/prototype_mwcog/configs/workplace_location.csv

Thank you for the comment Stefan, we appreciate it. We agree with your work location choice model calibration related suggestion. We want to clarify something though. As you know, as part of ActivitySim mandatory tour location enhancements, RSG proposed a simulation method as an improvement over the CT-RAMP method. We (i.e., MWCOG) opened this issue to share our thoughts on some minor changes that could improve the proposed simulation method. Our suggestions are meant for the proposed simulation method only and are not meant to address any model calibration related issue.

[stefancoe]

Thank you @NFerdous1. I don’t think the longer distances you are seeing in the results are necessarily inherent to the new simulation-based methodology (that address school and work location capacity/availability constraints) but are instead driven by the coefficients used in these models **EDIT to add: coupled with a more efficient/effective way of meeting constraints/targets.

**EDIT- both slides 33 and 29 (printed numbering, not actual) are showing results from the work location choice model. The use of 'half tour length' in the title had me confused.

These differences are most likely the result of using work/school location coefficients that were calibrated using the old shadow pricing methodology. When choosers are re-simulated because their first choices are not available, the same coefficients are used to determine their new choice. So, calibrating the distance coefficients should impact their choice at this phase and in the original (first iteration) simulation. I feel like adjusting the distance coefficients downward kind of gets at what you are looking for when you say:

“The re-run should be carried out with this implicit assumption that all things being equal, workers prefer work locations that are closer to their homes”

It would be interesting to see what happens if you make util_distance in the Work Location model a bit more negative.

[jfdman]

Hi all,

It makes sense that tours get longer when running either shadow pricing or the simulation-based constraint method. In the initial iteration, there's no job competition, but in subsequent iterations, the competition for jobs makes the utility worse or the alternative unavailable for certain zones, this causes the chooser to select a lower-utility alternative. The problem is not so much that the tour length is getting longer; the problem is that the longer constrained tour length may not be consistent with the actual tour length. This occurs because the model was estimated without consideration of competition for jobs. I believe that the appropriate way to handle this would be to introduce a term in estimation which reflects the competition for jobs at each destination zone. This can be expressed as a zonal accessibility term which reflects access to households (ie labor) from each destination zone. In application, the term would be dropped from the specification, which should result in a mandatory location choice utility after constraint that is more consistent with the observed data. It would be interesting to experiment with an observed dataset to see whether this theory is correct.