Improving the error message on the Gaussian Process Regression

I’m trying to do my first Gaussian Process Regression using scikit-learn.

Here’s my MWE:

from sklearn.gaussian_process import GaussianProcessRegressor
from sklearn.gaussian_process.kernels import Matern

kernel=Matern(nu=0.5)
model=lambda x: x*np.sin(x)
xdata=np.array([1,3,5,6,8])
ydata=model(xdata)
xfit=xdata[:,np.newaxis]
gp=GaussianProcessRegressor(kernel=kernel,alpha=0.5)
gp.fit(xfit,ydata)

When I run the above, I get

(...)\sklearn\gaussian_process\kernels.py:1623: RuntimeWarning: invalid value encountered in true_divide
  K_gradient = K[..., np.newaxis] * D \

Why is that? If I change the nu in the kernel to 1.5, the error disappears… I though it could be due to an inversion in the covariance matrix when compution the gaussian process, but I’ve used an alpha=0.5.

Also, if I try to plot the graph, with the code as is, I get a really strange plot… I somehow need to use length_scale_bounds=(1e-1, 10.0) to create a more truthful graph.

This tells me that probably the length-scale parameter is trying to go to problematic regions of the parameter space…

[anoldmaninthesea]

@cmarmo that warning is consistent with the hypothesis I put forward of the inversion of the covariance matrix being done for values that are problematic.

[cmarmo]

Thanks @AidarShakerimoff for the follow-up! And thanks @anoldmaninthesea for the report. Closing then as fixed in #19396.