ApolloChoiceModelling forum

Being able to parallelize the search start execution would drastically increase computational time when evaluating a large number of starting points.

Hi

yes, we have considered this, but have not had time to think about implementation

Stephane

I have attempted this, in addition to some other modifications for the searchStart function for my project. One other change was to track why solutions were pruned & how many ended up converging to the global/best solution. I also performed my sampling for the LCL/MXL based on the MNL solution, rather than using MLHS or other sampling methods.

I share my modified script if desired; I tried to keep track of my modifications using comments.

Hi Johnathan,

As Stephane mentioned, we thought about making apollo_searchStart, apollo_bootstrap, and apollo_outOfSample perform parallel estimation at the model level before, as opposed to using parallel estimation within each model. However, we did not do it because sharing the pre-processing for each model presented some challenges, and we did not have enough time to look into it back then.

If you are happy to share it, we would love to look at your solution.

Best wishes,
David