Notes on param range in trees
(param: type, range)
max_depth: int, if not none, then [1, )
min_samples_split: int -> [2, number samples], float-> (0, 1] ---> int updated to [2, inf)
min_samples_leaf: int -> [1, number of samples], float-> (0, 0.5] ---> updated to [1, inf) and (0, inf) respectively
min_weight_fraction_leaf: float-> [0, 0.5]
max_features: int -> [1, number of features], float (0, 1], or string (not checked with check_scalar)
max_leaf_nodes: int, if not none, then [2, )
min_impurity_decrease: float-> [0, )
ccp_alpha: float-> [0, )

Edit: updated ranges based on the ranges checked in the code
Edit2: note the upper bound for min_samples_split and min_samples_leaf was not explicitly specified in the code, but I included it in the new checks.

Hi @glemaitre, the range for these parameters is different in the existing code and in my notes:
min_samples_leaf: float-> (0, 0.5] vs (0, 1)
min_weight_fraction_leaf: float-> [0, 0.5] vs [0, 1)
max_leaf_nodes: [2, ) vs [1, )

Could you help to clear this up? I am already writing new checks based on the range in the notes, but if it makes more sense to keep the ranges consistent with the existing code, I will update them. Thank you!

It makes sense for max_leaf_nodes because if you have 1 split then you automatically have at least 2 leaves.
For the other one, it is less straightforward but if the code makes these checks so these bounds are correct.

For this PR we do need to update the messages in test_gbdt_parameter_checks since we are changing the error messages:

scikit-learn/sklearn/ensemble/tests/test_gradient_boosting.py

Line 119 in 4769a42

def test_gbdt_parameter_checks(GradientBoosting, X, y, params, err_msg):

Thanks @thomasjpfan, I've added a commit the updates the messages causing the issues.

The new release of numpydoc-1.2 is making test_check_docstring_parameters fail but this is unrelated to this PR.