Whitekernel, noise bounds

[sqbl]

While modelling very few or noisy data, there is a risk of "underfitting" the model, while placing all signal as noise. This is believed to be linked to the whitekernel that is added to the regressor-kernel. This might be solved by looking into the parameters set in the whitekernel. E.g the noise-bound can be reduced from the curren broad bands 1e-5 to 1e5 to something more sensible - although this might be problematic in cases where the y-axis is not normalized.

This might be part a wider discussion around the overall choice of default kernel to "Constant * Matern + White"

[dk-teknologisk-cbb]

The pull request https://github.com/novonordisk-research/ProcessOptimizer/pull/101 adds a parameter for setting noise_level_bounds.

I have not linked it to this issue, since I think we should still have some discussions about how the bounds should be set. I imagine we should create another pull request with an example describing how to decide which bounds to use in certain cases.

[dk-teknologisk-mon]

In some of the benchmark code I am working on now, the results will allow me to assess how close our models come to the "real" noise in a given system. This could affect how we set the noise_level_bounds value here. I consider this issue as on-going and will return with more opinions once I've completed the benchmark analysis.