The feedback mechanism for this model compares skim matrices from consecutive iterations and checks the %RMSE to determine convergence. The current parameters are:
Max Iterations: 5
Target Gap: 0.1%
In my latest base year run, here is the AM assignment report:
If we moved to 1% gap instead of .1%, the AM period would only run twice instead of 4 times. This would be a sizable reduction in run times without giving up too much on the tightness. Recall that the model is using warm-start link times which is why it doesn't need to feedback as many times.
On a related note, the various updates we've done to date in v2 (including moving to TC10 and others) are why the base year is feeding back rather than finishing after a single iteration. Before creating a new release, we always update the warm start times.
@dkyleward thank you for your research into this matter. Modificaiton of the closure criteria will be discussed and decided at the June 5 Tech Team meeting.
The feedback mechanism for this model compares skim matrices from consecutive iterations and checks the %RMSE to determine convergence. The current parameters are:
Max Iterations: 5 Target Gap: 0.1%
In my latest base year run, here is the AM assignment report:
If we moved to 1% gap instead of .1%, the AM period would only run twice instead of 4 times. This would be a sizable reduction in run times without giving up too much on the tightness. Recall that the model is using warm-start link times which is why it doesn't need to feedback as many times.
On a related note, the various updates we've done to date in v2 (including moving to TC10 and others) are why the base year is feeding back rather than finishing after a single iteration. Before creating a new release, we always update the warm start times.