Project Meeting 2021.03.09

Technical Call

• Discuss new issues and PRs
  • Brice to add more info, context, maybe a stack trace and then Doyle will review
  • I'll pull Matthew's doc updates when we merge develop2
• Discuss develop2 merge
  • Tests not passing again so Doyle to review
  • I plan to merge once working
  • I'll work on documenting the additional features after that
• Discuss representative logsums progress with multi-region timing sheet and additional config file
  • Feature working and tested with work/school location for MTC and runs 6x faster
  • Will test on example_semcog, which has 30 minutes time periods so many more alts
  • Will update the representative time periods based on MTC model
  • The mode choice model still has access to the actual start and end period
  • Need to add a config setting to enable/disable rep-logsums
  • Will run with and without and compare resulting time-of-day distributions
  • Mandatory tour scheduling was also the high water mark and so by using less memory, maybe we can get more throughput and faster overall runtimes
  • We should think more about how to manage the multi-region timing workbook since its for benchmarking and we don't want to lose the context of each entry after we move on to the next topic. Basically, what's our plan for maintaining/using this new resource?
• Discuss estimation notebook examples
  • All submodels working except mandatory_tour_scheduling and stop_frequency
  • We're looking into the mandatory_tour_scheduling issue/bug
  • stop_frequency coefficients are duplicated across purposes but needs to be in one big file to be estimated together
  • Newman to merge into one big file for us to use for the example going forward; Doyle update code as needed
  • Joint models use non-mandatory specs and the estimation code can now handle this cross referencing
  • Newman to do some more testing and notebook clean-up
  • Once develop2 merged to develop, Newman will merge in the estimation code
• Discuss PSRC/SANDAG testing progress
  • PSRC working through setting up estimation mode for the submodels
  • Will want to use latest code from Newman in a few weeks
  • SANDAG doing some performance testing with the MTC TM1 model design but with their inputs
  • Testing runtimes and memory use
  • Chunksize setting doesn't matter much because it is now adaptive
  • Mandatory tour scheduling is the high water mark, but some other submodels are also close
  • Will add machine specs to their reports and add to the multi-region timing google doc
  • Accessibility module running in 10 hours, uses TVPB; PSRC was 20 minutes but 2 zone system
  • SANDAG still working through setup issues and will probably add an example_sandag test example since its good to have a 3 zone system version too
  • Presampling cut submodel runtimes in half
  • Let's discuss progress again next time
• Discuss updated telecommute and transit pass design pages