Problem in building a Hierarchical Kriging

Hi everybody,

this is my very first post on this community, so I would like to say thank you for the project. I am Michele Capriati, PhD candidate and I am currently using UQLab for my study.

I am building a hierarchical Kriging following ‘The Gaussian process modelling module in UQLab’ article, anyway I come across a problem when applying it to my case study.
When I build the multi fidelity using the original uncertainty input space, the model response is good on the training points, but it seems to do not change the trend on the rest of the space, as in the following figure:
(legend: f:real function, LF:low fidelity model on 80 points, HF: High fidelity model on 3 points, MF: Multi-Fidelity on the same 3 points )
Anyway I noticed that normalizing the space, leads to much better results (I cannot post more than one picture.)

I would like to ask if such a problem was already been reported, or it is most likely due to a mistake in my script.

Many thanks, Michele

Hi Michele and welcome to UQWorld!

There does not seem to be a bug in the plots you show. This is an expected behaviour when the correlation length parameter converges to a very low value.

To achieve a more meaningful hierarchical Kriging surrogate you could try e.g. to increase the computational budget of the optimisation algorithm (e.g. increase the max. number of iterations, or the population size, etc.).

All the Best,

Hi Christos,

thank you for the suggestions, I will give a try!
To close the topic, I find interesting to report that normalizing the space leads to a much better result (attached). Probably it makes the optimization problem less stiff.