3
$\begingroup$

I had been reading this paper of multi-output gaussian processes. It is part of the literature review for a paper that I am writing. I download the software and was surprised that they are not using signals that are depending outputs between them. As soon I modified the functions to be dependent the algorithms do not want to fit the signals. If someone can take a look at this paper or the code and give me a comment, I will appreciate. I don't want to reject the algorithm without knowing other opinions.

https://github.com/trungngv/cogp

$\endgroup$
4
  • 1
    $\begingroup$ Are you only interested in the paper and code you cite, or would you be interested in other techniques for multi-output Gaussian Process Regression? $\endgroup$
    – DeltaIV
    Commented Feb 4, 2017 at 16:15
  • 1
    $\begingroup$ I am interested in know how can be correctly implemented because the code in the GitHub does not do what the paper claim. The GitHub has non-related outputs and as soon you relate the outputs it fails. The math looks ok. If you can give me your opinion about the method I will appreciate that. $\endgroup$ Commented Feb 6, 2017 at 2:50
  • 1
    $\begingroup$ to a first look, the methodology looks fine, even if IMO overcomplicated (there are simpler approaches to multivariate Gaussian processes). To tell you more I need to see what you did, i. e., you should post or link both the code with independent outputs, which ran ok, and the code with non-independent outputs, and see exactly what you mean by "failure". $\endgroup$
    – DeltaIV
    Commented Feb 6, 2017 at 14:03
  • 1
    $\begingroup$ Yes there are, I had the code from Multi-Output GP from Alvarez and Lawrence(works great). However the promise of lower times is too much....In the code they implement: fn1 = @(x) sin(x) + 1e-7; fn2 = @(x) -sin(x) + 1e-7; y1 = fn1(x) + sqrt(1e-4)*randn(N,1); y2 = fn2(x) + sqrt(1e-4)*randn(N,1); if you change to cos or sin(fn1) in the second function the method stop working. If the frequency is incfeazse it will not converge. I think that there is an error in the likelihood. $\endgroup$ Commented Feb 6, 2017 at 23:00

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.