Project Meeting 2020.11.24 - ActivitySim/activitysim GitHub Wiki
Technical Call
Update from Jeff Doyle on TVPB multiprocessing/performance, see working example
Running multiprocessed all the way through now on Mac
Performance testing mostly linear - ~16 min 1 process, ~8 min 2 process, ~4 min 4 process
Fixing a waiting for other processes to complete issue on Windows
Adding chunking to initialize_tvpb as well to better manage memory
We will review timing statements once we have the full scale Windows example run
During model development, probably want to use dynamic calcs since operating on a small sample, but during application, probably want to use static/pre-compute calcs to save runtime
Could probably speed up best path calc method even more if we spend more time on it
Feature is important to SANDAG, MTC, ODOT so faster the better
Let's get the full scale Windows run working, review timing statements, and discuss next steps
@Jeff fix issue and then work on estimation mode and then I'll re-run
Update from Jeff Newman on interaction_simulate estimation recipe and notebook example for larch, see issue
Non-mand tour freq notebook done
Like last time, instability in results since sample size is very small
Added optional ability to link coefficients (with the exact same value) for estimation
Run all feature working as well
Maybe want to do a coefficient template across person types (like we did for mode choice)?
@Doyle to update EDB for all person types and clean-up the CDAP coefficients as well so Newman can finish that too
@Newman and Doyle to chat later this week
Update from Clint on ARC progress
Multiprocessed full scale model run runs in about 16 hours
Using something like 36 of 40 processors on machine
Trip scheduling went from 20 hours to 4 minutes with tidying up the code, filtering out choosers not needed, etc.
This includes 5 iterations of work and school location choice shadow pricing
Mandatory tour scheduling is the slowest module at 7 hours. Uses 30 minute time periods and logsums. Could be sped up a pre-processor, which ARC may do in the next round of work
WSP just about done; plans to tidy up, write a verification memo, and then hand back to ARC
ARC planning transition in mid/late 2021
Great to see the same code base working for MTC, SEMCOG, and ARC. That's the mission of ActivitySim.
Update from me on work_at_home model integration based on discussion with Mark
DaySim has it as an oddball always available alternative in work location, rather than sampled
Makes shadow pricing and estimation complicated
Better for parcel level modeling since work_at_home means home, whereas for maz level modeling, people could work in the same maz but not work_at_home
Best to split these models out, including storing results in separate person variables - work_at_home = T/F, work_zone = XXXXX
The decision is then either before with simpler accessibility term or after with mode choice logsum
Let's do after since SEMCOG wants to use the mode choice logsum and its a more complete measure
The person work_at_home and telecommute_frequency attributes can be added to any downstream model without any code changes
Can we sensitivity test the simpler accessibility term versus the mode choice logsum?
Mode choice logsum will already be calculated so doesn't add much runtime to run the model after
I'll update the spec and then we can get started on programming
SANDAG talking with Boston MPO and DC MPO about telecommute modeling post COVID