May 2025 Meeting

Natalia Brown · April 30, 2025

Meetings

Date/Time: Thursday May 8, 2025 10:00-11:30am
Location: WFRC - 41 N Rio Grande St, Salt Lake City, UT 84101
Download PDF Agenda here

Virtual option available through the calendar appointment. Email utahMUG@gmail.com to request the calendar appointment.

Agenda

Welcome & Introductions

Anyone attending for the first time?
User Spotlights

Discussion Topics

Utah-Specific resources:
- Volume traffic map (link)
- Household/Job viewer (link)
- Model one-pagers and documentation (link)
Model Status Update (link)
- Utah Statewide Travel Model: Natalia Brown, UDOT
- Wasatch Front TDM: Chad Worthen, WFRC
- Cache TDM: Isaac Gardner, CMPO
- Dixie TDM: Radhika Bhandari, DMPO
- Summit-Wasatch TDM: Natalia Brown, UDOT
- Iron TDM: Natalia Brown, UDOT
Presentation solicitations (link) – due by July 1st
Others?

Presentations

Adapting Machine Learning Algorithms to Longitudinally Calibrate the UrbanSim Model System (link)
Paul Waddell, UrbanSim Inc.
Summary: This presentation describes the application of techniques developed in the Machine Learning domain to use back-propagation and automatic differentiation to longitudinally calibrate the UrbanSim cloud-based land use modeling platform. It enables longitudinal calibration of all model parameters to maximize the ability of the model to accurately predict spatial development patterns, household and employment location patterns, housing construction, and housing prices and rents over time. We also highlight the application of these modeling techniques to the GNRC region.
Half Time Break
Jared Lillywhite, Mountainland Association of Governments
Planners’ Perspective on Using Big Data (link)
Keith Hangland, Bentley Systems, Incorporated
Summary: This presentation will dig into the value of big data for transportation planners, including the challenges and opportunities it presents. We will get into details on sources of data, how modes are derived, potential biases, and practical uses cases for transportation planners.

Next meeting

Date/Time: Thursday, September 11, 10:00-11:30 am @ MAG.

🍽 lunchiMUG

Gather with us at HallPass for lunch.

Notes

Model Status Update

Utah Statewide Travel Model: Natalia Brown, UDOT
- Prepping for 2027 LRP
- Working on base data sets (SE, TAZ, Network)
- Updating freight and long-range components
- Prepping traffic counts for model validation
Wasatch Front TDM: Chad Worthen, WFRC
- 9.1 is current with 2019 base year
- We’ll be having some SE changes in the MAG area and network changes in the WFRC.
- Mid to end of May is anticipated
- 9.2
  - 2027 RTP
  - 2023 base year
  - Calibration to be complete by the end of the year
  - New SE and control totals
  - Draft form until official release in 2027
- Activity Based Model
  - Working on implementation plan
  - Solicitation on consultant help
  - Work to begin next year
Cache TDM: Isaac Gardner, CMPO
- SE data updates
- TAZ data to Natalia
- Traffic count database to calibrate the model
Dixie TDM: Radhika Bhandari, DMPO
- Resident and parcel updates
- TAZ realignments
- University component
Summit-Wasatch TDM: Natalia Brown, UDOT
- MAG & UDOT
- Freight component update
- TAZ and parcel data
Iron TDM: Natalia Brown, UDOT
- TAZ & network complete
- Working on base year data
- Freight component update

Presentation - Adapting Machine Learning Algorithms to Longitudinally Calibrate the UrbanSim Model System Paul Waddell, UrbanSim Inc.

Notes (set 1)
- Initial work was exceptionally labor intensive (10 models in 15 years)
- Keys to improving model development  Data development can take 3 years, calibration another 3 years
- Calibration is the bottleneck for all models even after all of the optimization techniques to improve development speed
- Heavy calibration of constants erodes model sensitivity
- UrbanSim: Forward Mode
  - Machine learning algorithm around the simulation over time and then backwards over inference
  - Model knowns in 2010 and 2020. Find the differences between outputs and knowns.  A loss function measures the quality of the UrbanSim model’s output. The loss score is used as a feedback signal to adust the weights
  - Do this about 2000 times.
  - In 2024: A 6-month project – full model development in Edmonton
    - Pearson Correlation of .98 or HIGHER
    - Census tract comparisons are predicted and observed levels after 10 years of simulation from 2006 to 2016. Modeling is at dissemination level
      - Did this for predicted rent price
      - Still got .92 or higher
  - Optimistic because it’s only a 10 year difference
  - Another test comparing a predicted and observed CHANGES over 10 years of simulation, from 2006 to 2016.
    - .86 and above
    - Rent correlation is much worse here
  - Computational Performance – Side Benefit
    - Calibration of large regions using 2000 iterations of a 10-year simulation complete within 40 minutes
      - This is for both forward and backward
    - Full 30-year simulation of large regions like Chicago run in around 40 minutes compared to 8 hours for a previous generation model  Most regions complete a 30-year simulation within 10-20 min
    - Simulation results will further improve by an order of magnitude if we turn on hardware acceleration using GPU
  - What’s next?
    - Add independent test period, following calibration
    - Calibration toolkit to be put on the cloud
    - Invite collaboration with broader research community to test alternative model specifications, different models
Notes (set 2)
- There is a circular relationships between transport & land use
- Model challenges for land use modeling include data, run times, development times, skepticism
- UrbanSim made lots of improvement to their model software: templates, factories, which reduce development and run times
- Calibration is the slow stuff - thousands of parameters, highly recursive dependencies.
- Decided to use AI to aid in the calibration
  - Forward simulation → loss function → backward propagation
  - Massive Parallezition w/GPUs
  - Allows for calibrated and validated parameter sets in hours

Presentation - Planners’ Perspective on Using Big Data Keith Hangland, Bentley Systems, Incorporated

Lots of data available: Roadway volumes, OD, etc.
Important to remember that counts ≠ trips : this can make validation hard or tricky
- Counts are not objective truth either
Getting modality is hard
Marketing hype can get in the way of understanding
Data science corrects and also fills gaps in data.