Getting NaN values for PCE when using different methods such as OLS, etc

UQ with UQLab Community Q&A and How To
,
1

Hi, everyone.
I am facing issues with using PCE model on an existing dataset with 5 geotechnical inputs (5 random variables). When I try to use the OLS method, the result summary says:

Input to SVD must not contain NaN or Inf.

I tried looking into the input parameters but couldn’t find NaN values anywhere.

Further, when I use the LAR method the following warning appears on the screen:

Warning: Warning: numerical instability!! Gamma for LAR iteration 19 was set to 0 to prevent crashes. 
Warning: Warning: numerical instability!! Gamma for LAR iteration 35 was set to 0 to prevent crashes. 
Warning: Warning: numerical instability!! Gamma for LAR iteration 55 was set to 0 to prevent crashes. 
The estimation of PCE coefficients converged at polynomial degree 3 and qNorm 0.75 for output variable 1
Final LOO error estimate: 6.785325e-02

This is giving me a tough time. Other methods are also throwing similar error messages. Can someone please help me understand where I am going wrong…

Attached is the following:

  1. Basic code of the problem.
    UQlab_code_need_help.m (2.5 KB)

  2. Dataset : matrix X is named “100_samples.xlsx” and output Y is named “FOS_100.xlsx”
    100 datapoints.zip (16.3 KB)

  3. Validation set with 400 datapoints: Xval is named “400_samples.xlsx” and Yval is named “FOS_400.xlsx”
    400 val datapoints.zip (40.3 KB)

  4. Custom user defined Inversegaussian marginal distribution package (zip file).
    inversegaussian.zip (6.6 KB)

Any help or suggestion would be highly appreciated. Thank you.

Regards,
Waris

3

Dear @Abdul_Waris_Kenue

Thanks for the nice reproducer!

I noticed that the distribution type for “RV1” is missing. Based on the comment you left in the code, I assumed that it should be a gamma distribution. But then the given distribution does not match the data. Can you add this information?

Best regards
Styfen

1 Like
5

Dear Styfen,

Thank you for taking the time to look into the problem I presented. It’s wonderful to see UQworld as such an active forum for discussions.

I revisited the problem and identified the source of the error, thanks to your insights. Your input helped me diagnose that I had been using a distribution of RV1 with a different mean than intended.

You were correct in your assumption; RV1 is indeed a gamma distribution. However, the moments used in the code should have been a mean of 0.07 and a standard deviation of 0.014.

The PCE model is now functioning correctly.

Regards,
Waris