UQWorld

Model-based sampling for more accurate metamodels?

Hi everyone,

I’ve been using UQLab for quite some time now to perform sensitivity analysis on an expensive COMSOL model and to generate predictors for my complex model. But now I switched to a matlab model and I’m trying to evaluate as less points as possible to build an accurate metamodel.

I was wondering if one specifies an m-file as the input model to UQLab, can UQLab generates model-based samples that can then be used to build a better metamodel. Because the way I understood it, the function defined in an m-file can only be used to compute a Y, which is an image of a sample X generated independently of the function.

I looked in the documentation and on the forum/google to see if such a feature is avalaible in UQLab but I didn’t find anything so I’m just asking if I missed it or if I have to code it myself.

Cheers,

Paul Lagouanelle

Hi @plagouanelle, welcome to UQWorld!

If I understand your question, I don’t think this kind of generic active learning to construct a more accurate metamodel is available in UQLab. For specific applications, like reliability analysis, there is indeed an enrichment scheme where new points are added iteratively using some learning criteria, such that the metamodel representing the limit state function is more accurate (e.g., AK-MCS available in UQLab). I think there is also such a scheme for global optimization using metamodels (e.g., EGO).

As far as I know, there is no such scheme for constructing a globally accurate metamodel, at least in UQLab. But I think, once the learning criteria is determined, you would be able to code it with UQLab. Anyone else has a thought on this?

PS: Looking at the search results of “active learning metamodel” quickly, there might be some promising leads (you most probably have your idea already though :slight_smile: )

Hi @damarginal !

Thank you for answering ! You do understood well my question. I already looked the documentation about reliability analysis and all… But as you said I got my own idea on how to code it and I found some others on the internet. So I guess I’m gonna start coding this for myself,

Thanks again for your help and the info

Great! If you need any help don’t hesitate to come back here.

And, if you’d like, you may also share with us later about your work using UQLab to generate a model-based sample for your purpose, some people here would surely be interested :wink:.

Thanks!

Great ! I will be happy to help.

Thanks !

1 Like