Leaving out out-of-design arms, when adapting searchspace constraints between iterations #2258
-
Hey there, I am experimenting with lowering limits in searchspace constraints between iterations using some heuristic, and I get the message e.g.: I have located in the modelbridge code base, the option to fit on data out of design or not. My question is, what would be the pros and cons of doing it or not? I can imagine if we are talking about an optimization with e.g. trust regions and many arms, then maybe it will reduce the models ability to model data locally if its fit on out of design data? I am working with a quite small setup of e.g. <60 total arms. I imagine SAASBO might gain from fitting on out of design data to fit lengthscales? Cheers |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 3 replies
-
Hi @SebBT01. The default behavior of filtering out out-of-design observations is intended as a safety feature if those designs are significant outliers that may adversely affect model quality (this includes manually attached baseline arms that are not valid parameterizations). In your case, it sounds like the out-of-design observations are just outside of the search space, and including them would likely improve the model quality. Adding
|
Beta Was this translation helpful? Give feedback.
Hi @SebBT01. The default behavior of filtering out out-of-design observations is intended as a safety feature if those designs are significant outliers that may adversely affect model quality (this includes manually attached baseline arms that are not valid parameterizations). In your case, it sounds like the out-of-design observations are just outside of the search space, and including them would likely improve the model quality.
Adding
fit_out_of_design=True
tomodel_kwargs
should solve the issue. Unfortunately, there isn't a convenient way of doing this, unless you're manually defining theGenerationStrategy
. If you're usingAxClient
, you could update themodel_kwargs
on it using somet…