search

markchil / gptools

Gaussian processes with arbitrary derivative constraints and predictions.
GNU General Public License v3.0
44 stars 17 forks source link

Output of Predict #9

Open altaetran opened 10 years ago

altaetran commented 10 years ago

Hi,

I noticed that the output of GaussianProcess.predict is often complex with a small imaginary component. Do you think this this something you could look into? Currently I am just using np.real() to address this, but I'm not sure if that's appropriate. Thanks

Best,

Han

markchil commented 10 years ago

Han:

I have noticed this on occasion when my hyperparameters stray into unphysical regions. If your hyperparameters are doing this while in a valid region, then this could point to a more subtle bug.

So I can investigate further, could you please answer the following: What covariance kernel are you using? What are some typical values of the hyperparameters that exhibit this bug? How many points are in the training data? What is the highest-order derivative you have in your training data? What is the highest-order derivative you are trying to predict? Do all orders of derivative you have looked at exhibit this behavior?

-Mark

On Aug 29, 2014, at 11:11 PM, Han Altae-Tran notifications@github.com wrote:

Hi,

I noticed that the output of GaussianProcess.predict is often complex with a small imaginary component. Do you think this this something you could look into? Currently I am just using np.real() to address this, but I'm not sure if that's appropriate. Thanks

Best,

Han

— Reply to this email directly or view it on GitHub.