Problem:
when using per_float_feature_quantization as one of the hyperparameters to optimize in grid_search, the parameter is exported wrong:
{'params': {'depth': 6, 'verbose': 200, 'l2_leaf_reg': 1, 'iterations': 500, 'learning_rate': 0.1, 'per_float_feature_quantization': [0.0]},
This causes a problem if running the grid_search with refit=True. However, it is similarly problematic if you extract the params and try to refit manually (or do anything else with the optimal hyperparameters).

A workaround is to inspect the catboost_training.json file and find the run with the best performance, then copy the parameters from there. However, this is quite tedious.

catboost version: 0.26
Operating System: linux
CPU: Tried on a local laptop Intel Core i5, as well as a google compute engine n1-standard-4

E: I switched to sklearn GridSearchCV instead, and then I get an error related to the fitting which was probably suppressed in the catboost gridsearch that I cannot use None for per_float_feature_quantization. Perhaps this is related?
In which case, I guess a related issue is: how can I include None as a potential value for per_float_feature_quantization in a grid search? I tried using an empty string, but that isn't accepted either.