Project Meeting 2021.01.19 - ActivitySim/activitysim GitHub Wiki

Technical meeting

  • Update from Joe on the bench contract

    • May be a bit awkward for one of the consultants to bid on task orders and set the agenda
    • Will likely have an owner's agent task order to continue to help administer the effort
    • This may end up being a bridge solution for the time being
    • @Joe to share draft scope for next round of work with the partners to discuss
    • Includes removal of orca and other draft 6b work items
  • Update from Doyle on performance issues profiling and PSRC support

    • Setup and ran PSRC example
    • Went ahead and multiprocessed accessibility since it was quite slow for PSRC and it was sort of already working
    • Run times went from something like 5 or 6 hours to 20 minutes
    • PSRC to test; code on the develop branch
    • Need to setup and run ARC next and then profile
  • Update from me on location sampling improvements design ideas

    • Jeff and I discussed the issue in detail and have two approaches that we're considering
    • Chooser approach and matrix approach
    • For chooser, location_sample always runs at the TAZ level
      • And size terms are aggregated to TAZ level when first calculated
      • n TAZs are picked (say 30) and then an MAZ for each TAZ is picked based on the MAZ's share of TAZ size term
      • In addition, we de-duplicate the calculation within chunks to save on runtimes and we also sort by origin zone to increase the amount of de-duplicate
      • Fits well with existing design and hopefully is fast enough
    • For matrix, we do the calculation as a pre-computed probability matrix by market segment
      • Users need to configure ahead of time and then reference in expression files
      • We need to build new data structures that are persisted? and referenced through the system
      • Expression files are OD level rather than chooser
      • Is just the number of calculations that are required + probability lookups with random numbers to make choices
      • May require more code changes and may not be as flexible moving forward
      • For estimation integration, we'd need to convert the outputs to a chooser like format or update the larch code to handle the matrix
    • We're still thinking it through, but we're hoping the chooser approach will be ok
    • Stefan drafted a quick chooser style approach and its runs quite fast so this approach may work just fine
    • For ActivitySim, we're often charting a course that draws from the hard-coded/fast DaySim approach and the UEC/slower/but more flexible CT-RAMP approach. This is a good example.
    • All software moving forward should support:
      • estimation
      • documentation
      • tracing
      • multiprocessing
      • multiple zone systems
  • Update from Newman on estimation enhancements

    • Plan to work on it this week
  • Planning for Joel to join next week's call to share the draft transit pass model design

    • Let's share the plan (or outline) ahead of time so partners have time to review
    • We need a generic model that can be used in multiple regions - not just a model for semcog
  • Update on telecommuting

    • Model programmed and being setup for semcog now
    • Is this being used in mode choice? Not yet. Any downstream model can be conditioned with the new person variables though
    • Partners planning to discuss getting better example data together for a more realistic example
  • Update on draft agreements shared

    • Plan to discuss again Thursday
    • New MOU solves everyone having to re-sign when a new partner joins
    • Funding agreement explains what you are paying for
    • The term of the agreement is still undefined
    • Want to get everything executed by July 1
    • Please review as soon as you can
    • Alex to share his comments