UQWorld

Surprising results with LHS sampling and constant inputs

Hi all,

I’m a bit surprised to see the following UQLab behaviour. I’ve been setting up a PCE-LARS model with 1 input and specifying LHS as experimental design, e.g.

metamodelLARS.ExpDesign.Sampling = ‘LHS’;
metamodelLARS.ExpDesign.NSamples = 8;
myPCE_LARS = uq_createModel(metamodelLARS);

I look at the samples by studying myPCE_LARS.ExpDesign.X.

The surprise is that when I repeat this, but extending to 2 inputs of which one is a constant, I get different samples (note that the random seed is fixed). I had thought that the constant input would not have to be sampled, but it looks like an LHS of the 2 dimensional space is made, which is then transformed to be constant in the second dimension (somewhere in uq_GeneralIsopTransform). Is that correct? Could you explain me a bit why this is done?

Best regards,

Benjamin

Hi @bsanderse,

Thanks a lot for reporting this issue. You are right: when a metamodel generates the experimental design (from a random sampler), it does not disregard the constant variables. The constant variables are only treated by transforming all the corresponding samples to the constant value. We have fixed this bug in the code which will be released together with a set of hotfixes in a few days.

Best regards,
Xujia

Thanks Xujia!
I’d be happy to know when the hotfix is out.
I was also wondering, whether this issue has also been present in the earlier versions (1.3)?

Hi @bsanderse,

We are performing some final tests before releasing the new subversion (V1.4.2).
Regarding the early version, I think the issue was also present in V1.3.

Hi Xujia, do you have an idea when the hotfix will be available? I’ve been waiting to finalize some results for a paper, but in case you think it will still take some time I will just continue with the current results.

Hi @bsanderse,

Thanks for asking. The new subversion V1.4.2 with some hotfixes is available now. You will find an update notification when activating UQLab in Matlab.