Looking for documentation on Bayesian inference with surrogate model

Hello all,

I’m using UQLab with PCE surrogate models to speed up Bayesian inference. I looked at the beam calibration case (https://www.uqlab.com/copy-of-inversion-map-estimation), and got some idea about how this works. However, in the documentation on Bayesian inference (https://uqftp.ethz.ch/uqlab_doc_pdf/1.3.0/UserManual_Bayesian.pdf) I cannot find more detailed information related to this option. One concrete question is: is the sampling for the surrogate model based on the prior? If yes, how does the surrogate model know about this? Can this be changed? How to specify sampling schemes?

Best regards,

Benjamin Sanderse

Hi Benjamin and welcome to UQWorld!

When using surrogate models to accelerate Bayesian inversion, the idea is to simply replace the original forward model with a priorly trained surrogate model inside the Bayesian inversion analysis. How you train this surrogate is in principle independent of the prior, although typically the prior distribution is used to obtain a surrogate model that is globally accurate.

You can, of course, also train the surrogate model with a different distribution. To avoid confusion between the prior (e.g. myPrior) and the surrogate training distribution (e.g. myInput), it makes sense to manually assign both to the surrogate model and inverse analysis through:

MetaOpts.Input = myInput;
BayesOpts.Prior = myPrior;

before calling uq_createModel(MetaOpts) and uq_createAnalysis(BayesOpts).

Let me know if this works :slight_smile:

Hi Paul-Remo,

Thanks for the quick reply! This helps, I’ll play around with it.