Is it possible to use an active learning (sequential experimental design) approach to metamodel based on external datasets?


Most of the current metamodeling in Uqlab is based on MC\LHS, etc. In these experimental designs, all design points are selected in advance before any (computer or real-life) experiments are performed, and no other design points are selected afterwards. Although the metamodel accuracy can meet the requirements, it is prone to over-sampling and/or under-sampling.

I observed that active learning strategy is used in reliability analysis, so can it be applied in metamodeling, especially based on a given number of external data sets?

Dear Felix,
I am a UQLab user … and thank you for bringing this up! If it is possible to use the active leaning approach this would reduce the time/cost and reduce the uncertainty!
In a recent study we applied an adaptive sampling technique, similar to what is described here
Surrogate-assisted global sensitivity analysis: an overview | SpringerLink

But for my case I already connect UQLab to COMSOL so I have a wrapper function and I managed to get the adaptive sampling within the wrapper function for the sensitivity analysis. Maybe this could give you an idea of a better way of implementing!
Best regards

1 Like

Dear Tam:

Thank you very much for your suggestion, and I have studied the linked article carefully, which lists several high-dimensional and complex functions, and the results of the sensitivity analysis are very good.

However, what I want to achieve is to construct a metamodel using an active learning strategy based on an external dataset obtained from probabilistic finite element analysis[Black Box].

Similar to the following code, which performs metamodeling with a specific number of DOEs and validation sets, but this approach cannot adaptively build surrogate models.

FILELOCATION = fullfile(uq_rootPath, 'Examples', 'SimpleDataSets',...
load(fullfile(FILELOCATION,'Supportbeam_DOE40.mat'), 'X', 'Y');
load(fullfile(FILELOCATION,'Supportbeam_VAL30.mat'), 'Xval', 'Yval');

You said that you have connected UQLab to COMSOL in your case, are you building a metamodel based on an external dataset like me?

Thank you very much!
Have a nice day