2025-12-16 Product/Community Team #53
Description
Agenda
- SEMCOG Agency Report Out
- Phase 11C Task Summary
- ActivitySim Application & Analysis Guide - Confirm scenario selection
- Data Model / Data Dictionary - Confirm expansion of data model scope to include entire model systems
- Explicit Error Terms: Updated Random Number Generation - status of TransLink research
- Other Tasks
- Consultant Review and Cost Estimates
- AI Taxonomy
Meeting Notes
1. SEMCOG ABM Evaluation (Primary Topic)
Background
SEMCOG serves seven counties in southeastern Michigan with 2,800 TAZs, 1.9 million households, and 4.8 million residents. The agency is transitioning from a trip-based model to an activity-based model (ABM) using ActivitySim, currently under evaluation before production deployment.
Implementation Milestones
- Version 1.0 (2023): Initial implementation with ISG using ActivitySim 1.2
- Version 1.1 (2024): Upgraded to Cielo using ActivitySim 1.3
- Version 1.2 (2025): Transit model enhancements addressing transit subsidy and pass ownership issues
Performance Improvements
Model runs on AMD machine (64 cores, 512GB RAM) with 5 feedback rounds achieving approximately 12-hour runtime—meeting the goal of overnight processing. Significant improvements from pre-Cielo version (18 hours) in both runtime and memory usage.
Transit Model Enhancements
Initial ABM showed only 3% subsidized transit fares versus 35% from 2019 onboard survey, and 5% transit pass ownership versus 48% from survey. Improvements included:
- Added transit availability from home to work/school for subsidy model
- Added transit availability at home for pass ownership model
- Developed new calibration targets by county and person type using both 2005 household travel survey and onboard transit data
ABM vs. Trip-Based Model Comparison
Tested three years (2020 base, 2025, 2050) from 2050 RTP using nearly identical inputs. Key findings:
- Auto trips showed reasonable base year match and future trends
- Major difference: Trip-based model shows transit ridership increase 2025-2050; ABM shows flat trend
- Transit boarding differences warrant further investigation
Scenario Testing (6 Tests)
Tested extreme scenarios to answer planner questions about achieving 10% transit mode share:
- Free transit fares
2-3. Increased transit frequency (every 5 minutes) - Increased driving cost ($0.18 to $1.00 per mile)
- Increased bike speed (12 to 20 mph)
Key Findings:
- Both models showed minimal response to free fares
- Both showed sensitivity to frequency changes
- Trip-based model showed no sensitivity to driving cost changes (requires investigation)
- ABM demonstrated significantly larger sensitivities overall
Sensitivity Testing Concerns
Transit Fare Insensitivity Issue: ABM shows unexpectedly low sensitivity to transit fare changes. Two tests conducted:
- Changing fares in import files produced larger response than changing utility coefficients
- The magnitude of difference was unexpected and requires further investigation
Possible explanations for low sensitivity:
- University of Michigan provides free transit
- High percentage of subsidized fares in region
- 50% of transit riders are from zero-auto households (transit-dependent)
Additional Testing
Free-Flow Speed Changes:
Work tours showed highest sensitivity to speed changes, which was counterintuitive. Question raised whether teleworking/work-from-home options create new sensitivities not yet fully understood by the industry.
Capacity Reduction Tests:
- I-696 highway closure (29 miles, 50% capacity reduction): Work tours decreased in nearby counties, which seemed unexpected
- Woodward Avenue transit corridor: Transit boardings slightly decreased; team expected larger impact from congestion
Aging Population Test:
Shifted 10% of population to senior age while holding regional total constant. Results showed logical patterns: increased total trips but decreased VMT, shorter average trip lengths, shifts from commute to other trip purposes, and increased walk/shared ride modes.
Project Applications
ABM used for two exploratory transit projects:
- Detroit to Ann Arbor (D2A2): 40-mile corridor with alternative stop configurations
- Bus Rapid Transit: Four major transit corridors evaluation
D2A2 findings: Fast, direct routes performed best; extra stops did not boost overall ridership; Wayne State University stop added more value than Michigan Central stop.
Current Status & Next Steps
- Created 34 GitHub issues; 6 remain open working with ISG
- Further investigation needed on transit fare sensitivity and transit trend differences
- Remaining tasks: conformity testing with ABM output, potential telecommuting scenarios
- Target: Version 1.3 for actual project use
- Formal evaluation memo forthcoming with consultant partner (CS) assistance
2. Phase 11C Scoping Update
Major Changes
Joe presented revised Phase 11C task document incorporating feedback from partner meetings. Key modifications:
ActivitySim Application & Analysis Guide Task:
Multiple suggestions received including land use changes, telecommuting scenarios, network/supply-side changes, land use model integration, and micromobility. Scope and resource allocation still being determined.
Data Model / Data Dictionary Task:
Merged previously separate "model documentation/data dictionary" and "data model" tasks due to significant overlap. New combined scope addresses entire model documentation while including data model considerations.
Other Tasks:
- Random number generation: Unchanged, awaiting TransLink testing results
- Output analysis: Unmodified
- Telecommute implementation: Minor changes for calibration guidance
- Model calibration: Moderate changes warranting review
Procurement Approach
Removed initial proposed task assignments. Plan to:
- Share document broadly with all consultancies
- Solicit scope / schedule / budgets from consultancies
3. AI Taxonomy Development
Brief discussion on developing AI use case taxonomy for transport modeling workflows, requested by Transport for New South Wales. Lisa suggested creating framework for AI use cases and technologies.
Suzanne noted two distinct areas:
- Developing models using AI (early stage)
- Using AI for model development/scripting (more advanced)
Joe committed to drafting framework document this week for collaborative input from both consortium members and consultants.
Action Items
- Joe/Zephyr: Finalize Phase 11C scope with key agency staff and distribute to consultants
- Joe/Zephyr: Draft AI taxonomy framework document for consortium input