Choice of sampling type for regression-based PCE


I have a general question about regression-based PCE model building: are there any general guidelines on how to select the type of experimental sample (e.g. LHS, Sobol, Monte-Carlo, etc.) used when constructing a PCE model?

Some more detailed questions:

  1. Is the efficiency of a certain experimental sample type very application-dependent?

  2. Does it depend on the scale of the considered input variability (e.g. meta-modeling over large input variation domains vs uncertainty quantification)?

  3. Does it depend on the type of PCE coefficient calculation method (OLS, sparse methods (LARS, OMP, SP, etc.))?

I would be grateful for any insight on this topic, thank you.

Dear @achille

In 2021, @nluethen published an excellent paper titled “Sparse Polynomial Chaos Expansions: Literature Survey and Benchmark,” which provides insights into the questions you’re asking. I recommend taking a look at it!

Best regards

1 Like