Reliability analysis with AKMCS

UQ with UQLab Community Q&A and How To
1

Hello everybody.
Thank you for read my question. i have some problems to deal with the akmcs, for example, how to replace MCS with subset simulation to improve the akmcs with UQLab. can it be achieved with UQLab, thank you for the help.

1 Like
2

Dear Youxiong

Thanks for your interest in UQLab.

In the current version of UQLab, only the original AK-MCS is available. We are currently preparing a new active learning module for reliability, which will allow users to combine any surrogate model, active learning technique and simulation technique seamlessly.

Meanwhile you can carry out AK + subset simulation manually, using a for loop where you:

  • build a Kriging model of your limit state function with your current experimental design;
  • use any simulation technique to solve the reliability problem with this current surrogate;
  • if not converged, use one enrichment criterion, for instance the U function (see UQLab user manual - Reliability page 16).

We will keep you informed about the development of the new module.
Best regards
Bruno

1 Like
3

Thank you for your prompt reply. I am looking forward to the new version. In the end, can you refer to the example of for loop or combination? Thank you very much.

4

Hi @youxiong,

Currently UQLab only has two Active Kriging strategies implemented, the original AK-MCS and APCK-MCS (where Kriging is replaced by PCK). There is no option to replace MCS with SuS. We are developing a module for a generalized active learning framework in UQLab where you would be able to combine Kriging and Sus. This should be available in the next release.

In the meantime if you want to use subset, I would suggest you create a uq_analysis in the current uq_akmcs.m file in place of the MC estimator (lines 125-128). You can either keep the xcandidate variable as candidate pool for enrichment but make sure its size is large enough in case the target failure probability is extremely small.

Please let me know if you can’t work it out. I can give you more detailed instructions as to how to implement this.

Cheers,
Moustapha

5

Thank you very much for your reply, you fully understand my question, the line125-128 is EstimatePfall = [EstimatePfall, (sum(gmean <= 0) + sum(g(Options.AKMCS.IExpDesign.N+1:end,oo) <= 0))/MCSampleSize];
EstimatePfallp = (EstimatePfallp, (sum(gmean-gsnorminv(1-alpha/2) <= 0) + sum(g(Options.AKMCS.IExpDesign.N+1:end,oo) <= 0))/MCSampleSize ];
EstimatePfallm = (EstimatePfallm, (sum(gmean + gsnorminv(1-alpha/2) <= 0) + sum(g(Options.AKMCS.IExpDesign.N+1:end,oo) <= 0))/MCSampleSize ];
EstimateNsamplesAdded = [EstimateNsamplesAdded, NsamplesaddedTotal]; I don’t know how to modify it to make ak-mcs or apck-mcs into ak-ss or apck-ss, and use it correctly. Thanks again for the answer, thank you.

6

Dear @moustapha

I am very interested in this module for a “generalized active learning framework” since I have also come across with the need of using AK-IS or AK-SS to compute the reliability of systems with a very low probability of failure (i.e. approximately 1.0E-9). When is this new version of Uqlab coming out with these new features?

Looking forward to hearing from you.

Best Regards,
Neryvaldo

7

@Neryvaldo_Galvao

The “active learning module for reliability”, that generalizes AK-MCS to all kinds of AK-IS, AK-SS, etc. will be part of the release of V1.4. scheduled end of December 2020. This new module was mainly developed by @moustapha.

Best regards
Bruno

4 Likes
8

Dear @bsudret

These are great news. Looking forward to the release of the V1.4 of UQLab.

Cheers,
Neryvaldo

9

Dear Prof. @bsudret

It’s such a great opportunity for us to enjoy and learn more in the context of UQ. Many thanks, sir.

Best regards
Ali