sign error when using fmin_cobyla optimizer for gaussian processes #3180

filthysocks · 2014-05-22T11:22:19Z

Hello,

the gaussian process are by default fitted using the fmin_cobyal optimizer. The optimizer is started from various random inital conditions to avoid local extrema. The problem is, the worst of those extrema is chosen in the end.

optimal_minus_rlf_value, optimal_par = self.reduced_likelihood_function(theta=optimal_theta)
# sign error 
# reduced_likelihood_function return the likelihood
# not the negative likelihood
optimal_rlf_value = - optimal_minus_rlf_value(theta=optimal_theta)

The reason it did not show up in the unit test is, that they been to easy. Even the worst local maximum was good enough to pass the test. I added a test which is more "tough" and manage to prove the current solution wrong.
Here is the solution to the unit test plotted with the current implementation:

and fixed:

filthysocks · 2014-09-26T15:39:02Z

duplicate of #2632, which got merged.

filthysocks pushed a commit to filthysocks/scikit-learn that referenced this issue May 22, 2014

fix sign error in gp + unit test scikit-learn#3180

661d3c3

filthysocks mentioned this issue May 22, 2014

fix sign error in gp + unit test #3180 #3181

Closed

filthysocks closed this as completed Sep 26, 2014

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

sign error when using fmin_cobyla optimizer for gaussian processes #3180

sign error when using fmin_cobyla optimizer for gaussian processes #3180

sign error when using fmin_cobyla optimizer for gaussian processes #3180

sign error when using fmin_cobyla optimizer for gaussian processes #3180

Comments