GP not robust to NaN metric

[amueller]

I'm running MOBSTER with default parameters, and my model (for whatever reason) decided to produce np.inf, which took down the tuning process.
In particular, in transform_state_to_data, the normalization makes all the targets be NaN. One option would be to exclude inf from the normalization, but then probably the GP chokes. Setting inf to the largest observed value might make sense? Another option would be excluding the sample.
Any thoughts?

[geoalgo]

Hi Andreas, thanks for flagging this it seems like a good suggestion, @mseeger is the main author of this code and may be able to help you.

[mseeger]

Closed with #788