4
$\begingroup$

I used Gaussian Process Regression to predict a time series, what I have is sensor's readings that come every hour ( I have data for about 3 years) I chose the periodic kernel function mentioned here [1] which looks like this

$$ K(x,x′)=\sigma^2 \exp(−2\sin^2(π|x−x′|/p)/\mathcal{l}^2) $$ The period p simply determines the distance between repetitions of the function.

In my data I can observe 2 kind of periods (every 8 hours, every 24 hours) so P in this case can be 8, 24 1.how can I estimate the other hyper parameters $\sigma$, $\mathcal{l}$? 2.how can I plug the two periods in this kernel function?

[1]The Kernel Cookbook:Advice on Covariance functions http://www.cs.toronto.edu/~duvenaud/cookbook/index.html

$\endgroup$
4
$\begingroup$

answering them in reverse order..

2) let K be the sum (or multiple but I think in your case sum) of the two kernel functions. That is one with each period.

1) you want to minimise the negative log-likelihood as explained in sections 5.4.1 of GPML (link here).

$\endgroup$
  • $\begingroup$ Thank you for your answer I could not understand how exactly I can implement the log-likelihood thing with matlab :( Do u have any online tutorial to help me with it ! $\endgroup$ – malsioufi Feb 26 '17 at 2:27
  • $\begingroup$ Check out the GPML code that comes with the book (both free :) ). Warning though sometimes it'll through an error if you have a really ill-conditioned covariance matrix! (sometimes confusing to people learning at first) $\endgroup$ – j__ Mar 10 '17 at 10:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.