Project Meeting 2020.12.22

Technical Call

• Travis
  • Our test system, TravisCI is no longer being free
  • We'll sign up for the paid version, $70/month and expense the project moving forward
  • Easier and cheaper to do that than to switch to another free option
  • Check contract for billing expenses
  • https://github.com/rsginc/activitysim_resources also costs money, as does ActivitySim.org
• Release preparation
  • Working to finish pulling / reconciling various contributions to develop
  • Will release multizone and wait on estimation until everything done and QA/QC'd in Jan
  • I checked out Clint's PR to my fork and then resolved conflicts and merged to ActivitySim since I don't have write access to Clint's fork
  • Newman and Doyle to work on estimation branch moving forward
• ARC update
  • ARC not using the coefficient files - just a dummy file for now since that's easier for translating CT-RAMP UECs
  • Would be good to note things to watch out for in release notes
  • Updated to pandas 1.1.0 required
  • Dynamic chunking for ARC dropped runtimes from 16 hours to 8 hours!
  • Clint to share his activitysim presentation to ARC with the team at a later time
  • Clint tried the scheduling model improvement of just O,D pairs and not O,D,duration but didn't help
  • We still think it would be good to de-duplicate within chunks the logsums alternatives to improve runtimes
  • Doyle to work on performance tuning after estimation work
• TVPB validation
  • Fixed a few more drive transit expressions in the Marin examples and results looking good
  • Trying to balance directly translating the existing expressions versus cleaning up the example so its easier on someone who downloads our example and wants to learn from it
  • For distributed examples, we may want to take out all the calibration/adjustment type stuff specific to regions since it adds a lot of extra stuff to the expressions to understand and support, makes it more difficult to use for a new region, and highlights the dark corners of model calibration/fit
  • Basically, our examples should strive to be more examples and less a specific implementation for a specific region
  • I'll move the auto sufficiency expressions to the tour mode choice pre-processor to make it easier to understand
  • Its probably three phases to creating good examples - 1) Marin TM2 --> 2) ActivitySim software testing --> 3) ActivitySim example for distribution
• Licensing / copyright
  • Bench contract delayed due to inconsistent project history with licensing / copyright
  • There's conflicting license and copyrights in all previous contracts and GitHub resources
  • At issue is a work for hire clause and copyright, especially related to UrbanSim and the time period of the acquisition by Autodesk
  • RSG, WSP, and CS happy to do whatever the consortium/AMPO wants
  • Joe talk with AMPO attorney and may assert copyright back to the beginning of all contracts
  • Guy to help Joe with understanding old discussions related to the topic
• Discuss estimation integration progress
  • Our example model can produce some synthetic travel diaries that are not consistent with the CDAP patterns
  • Not many, but a few, and for example two preschoolers going shopping
  • So this makes testing estimation a bit difficult because our data set is a bit funny
  • CDAP model is already pretty complex so let's not add more to it
  • Basically it seems there is some leakage in the downstream logic
  • Would be good to know how rare this is?
  • It is the users responsibility to supply a well formed survey for estimation
  • Would be good to be very explicit in the docs about the survey data expectations, which of course vary by the submodels/spines
  • I'll create an issue so we don't forget
  • Let's talk with Doyle next time we chat
  • We could add re-simulating a household like DaySim does, if the household is determined to have bad results or can't complete