Agenda

• School escorting code review
• Follow-up to shadow pricing

Action Items

• Discuss trip scheduling at next meeting
• RSG to prioritize understanding run-time with current shadow pricing approach, before finalizing code and addressing other topics such as WFH and in-commuting.
• RSG to look into what happens when all MAZs in a TAZ are -999. Is the shadow pricing considered in the sampling mechanism or is it just the size term? If it is size term, they could also modify the size term, by making it zero.

Meeting Notes

Admin Things

• Draft scope for memory profiling task
  • Reminder to review draft scope
• Code review plan
  • Sharrow – both WSP and RSG reviewing
  • School escorting - Joe Flood has been reviewing
  • WSP still needs to discuss resources for code review with Joe
• Issue on time of day distribution used for stops
  • Scheduling is done based on time window available for the entire tour
  • Postpone discussion to Thursday, put a deck together to understand what’s currently being done now and what are alternatives (with general level of effort required)

Shadow pricing

• Under the current thinking, there is no shadow price other than -999 or 0. If the shadow price is -999 after the first iteration, no more workers would be added.
  • Never starts with a previous set of shadow prices.
  • There are no shadow prices, so you can’t save them for the next run, meaning there isn’t an initial set you could start with like in other shadow pricing approaches.
• RSG to look into what happens when all MAZs in a TAZ are -999. Is the shadow pricing considered in the sampling mechanism or is it just the size term? If it is size term, they could also modify the size term, by making it zero.
• Does scaled workers and scaled employment account for number of workers who choose home as regular workplace location or number of jobs consumed by people who in-commute from out of region?
  • Most models have telecommuting before workplace location, so they wouldn’t be in the sample set. SEMCOG model runs after workplace location. This needs more thought.
  • It is in the scope to look at in-commuting, they are just not there yet.
• Comment to do a reasonableness-assessment of run-time implications before the code is very polished. If it’s not minutes, there needs to be a discussion about whether the simulation-based approach is the best path.
  • Suspicion is that this approach is faster than what’s there now, but what’s there now isn’t needed to be run every time.
• Shadow pricing only kicks in with full sample (usually last iteration)
  • This is a concern, would like to have some shadow pricing for every iteration.
  • Could possibly use imported shadow pricing from other efforts for early (non-full-sample) iterations – this is an option that could be considered (turn off shadow pricing and use another method to generate initial shadow prices).
  • Need to look at how much of an impact the shadow pricing has, how different are things from the non-shadow price destination choice.
• A maximum number of iterations can be specified, no guarantee that the process would reach the conditional convergence
• Convergence criteria is dependent on the sample size
• Decision: prioritize understanding run-times (and convergence properties) and then address things like WFH and in-commuting. Readdress this topic after run-times are determined.