Agenda

• Confirmation of abstract submittal to 2024 AMPO conference
• Recent Phase 9a activities

Meeting Notes

Admin

• After the meeting, Michelle confirmed that Joe submitted the abstract for the 2024 AMPO Conference.

Action Items

• Discuss unit testing for BayDAG PRs

Phase 9a Updates

• RSG is currently working on the BayDAG PRs, so they can get a handle on two-zone performance before going too far along with one-zone model optimization. Right now, it’s individual PRs that are self-contained and easy to review. Would it be better to review individually and then merge all and run tests?
  • Jeff suggested they consider setting up unit tests for individual PRs? There are several things to consider, and the appropriateness of a unit test may vary between PRs. TO DO: Talk about them individually at a separate meeting.
  • What are our priorities? If we determine that the unit testing is going to take a long time, but we need the PR for the two-zone tests, we need to decide the priority. Current 9a SOW does not include creating unit tests.
  • TO DO: Talk about them individually at a separate meeting.
• For next Tuesday's meeting:
  • What are the PR requests that need unit tests? Which ones won’t impact our benchmarking? For example, we haven’t done profiling on estimation mode.
  • Results of the most recent optimized one-zone model are slightly different, and tests are failing because of that. There are multiple reasons why there are differences. For example, the bug that Sijia identified regarding the number of school escortees was an incorrect implementation. That fix has consequences in the results. In the vehicle type choice model, we removed all the vehicles where there is no make/model available, which resulted in a significant improvement in memory and some runtime. However, the model spec before this change did not have non-zero probabilities for alternatives with no makes/models – it would have been unlikely but not impossible to choose a vehicle type that doesn’t exist. The fix is an improvement, but there are slightly different choices as a result. We’ll need to have revised targets because of the bug fixes.
• Long-term – where are we going with the tests? Are we just testing against the MTC one-zone example or others? Right now, the focus is on the external repository. In the future, test will be run against the reference model, not everyone else's models. For the ActivitySim core model code, we should run tests against all our reference models and grow unit tests that can test individual components.
• Multiprocessing version of the test is failing, even after rolling back all the changes, but the single processing version passes. CS will dig through and figure out where/why.
• Overflow protection – Jeff has not yet merged the changes to pass those controls to the user. Sijia to review this week.
• David opened a draft PR with updated trip scheduling configs. It will change the results slightly, so tests would need to be updated because of that.