Bill liked that it is after work/school location choice, before auto ownership, by person type, and thinks we should predict a subsidy ratio rather than an absolute value
Important to include pass purchase cost in model too since some folks can't afford the monthly pass purchase even if it benefits them throughout the month and so the model is sensitive to purchase price
We don't want to just assume fares are 50% off with a pass
Access to transit passes are often tied to your employer as well so maybe we should include person occupation/industry
Still lots of outstanding questions
Next steps are to do some data analysis and report back what we're seeing
I'll also post the PPT and update the design wiki page with the draft design from the presentation
Alex would like something more formal to react to / review
Refactored estimation integration so it is now testable
Added test suite for estimation, which adds a dependency on larch
Consolidated code so it is more runnable
Updated example notebooks so much more streamlined
Should we roll larch code officially into activitysim now? It seems we should, but let's discuss more
PSRC says updates look great and are now using the larch notebooks too
Maybe we make the larch asim integration a separate package under the activitysim org account as a compromise between in asim or in larch?
Jeff tidying up and then I'll review in more detail
Discuss performance profiling results, see asim_timing
Ran 100k sample of MTC, PSRC, and SEMCOG (with 30 minute time period setup like ARC)
Clint to post ARC setup for us to download and run
Three areas for performance improvements noted:
A representative time period for calculating logsums (for the scheduling modeling) + de-duplication to cut down the number of time period combinations to calculate for each tour and trip. This is shown in the SEMCOG example. Each TOD alt gets mapped to a representative skim period and representative 30 minute period. Could be done with annotation pre-processor. Only the parking cost/duration really gets impacted by this approach.
TAZ location sampling for PSRC like we discussed last week
Trip purpose/destination failing a lot for PSRC? Often transit trips or no HHs in the size terms data/simulation inconsistency issues. We could add a report of a crosstab of failure by purpose and mode to help see what's the issue.
Also ARC model not using mode choice logsums for non-mandatory travel and so runtimes will be even higher for models that do
Joel to email Clint some things they found when implementing some of the ARC extensions