How to optimize kernel_size in gaussian process method?

[hiremasa]

Thank you for your great repogitory.

Let me ask one question:
According to your implementation of method="gp", I see that you specify the lower and upper bound on ‘length_scale’ as kernel_size_bounds = (0.5 * kernel_size, 2 * kernel_size) (https://github.com/hippke/wotan/blob/master/wotan/gp.py#L34).
However, when fitting(https://github.com/hippke/wotan/blob/master/wotan/gp.py#L62), you don't specify any 'n_restarts_optimizer' value and use default value zero. This means that initial kernel_size will not be changed, I think.
Reference of sklearn: https://scikit-learn.org/stable/modules/generated/sklearn.gaussian_process.GaussianProcessRegressor.html
So, when do you optimize kernel size in gaussian process method?

[hippke]

I'm no expert of GPs. What do you recommend to improve the situation?

[hiremasa]

I guess it would be better to use n_restarts_optimizer=5(5 is arbitrary) of GaussianProcessRegressor's argument.

Please see: https://scikit-learn.org/stable/modules/generated/sklearn.gaussian_process.GaussianProcessRegressor.html

Edit:
Is the kernel_size unchanged during a fitting in your initial implementation?