FanovaImportanceEvaluator mutates study.trials[-1].distributions when a parameter has dynamic range

[sgrubas]

Hello,
Thank you for your great work!

I ran into an issue with using FanovaImportanceEvaluator. For some reason, it changes the last trial of a study when the search space of a parameter depends on another parameter. See the code snippet below:

import optuna
from optuna_fast_fanova import FanovaImportanceEvaluator
optuna.logging.set_verbosity(optuna.logging.WARNING)

def test_objective(trial):
    x = trial.suggest_float('x', 1, 4)
    y = trial.suggest_float('y', x, 5)
    return x**2 + y**2
    

test_study = optuna.create_study(study_name="Test", direction='minimize')
test_study.optimize(test_objective, n_trials=100, show_progress_bar=True)

importances = optuna.importance.get_param_importances(test_study, evaluator=FanovaImportanceEvaluator())
for trial in test_study.trials:
    if trial.distributions.keys() != trial.params.keys():
        print(trial)

The output:

FrozenTrial(number=99, state=TrialState.COMPLETE, values=[3.7186055946265806], 
datetime_start=datetime.datetime(2024, 2, 22, 18, 43, 50, 977310), 
datetime_complete=datetime.datetime(2024, 2, 22, 18, 43, 50, 981316), 
params={'x': 1.3032071826948155, 'y': 1.4213573208729824}, user_attrs={}, system_attrs={}, intermediate_values={}, 
distributions={'x': FloatDistribution(high=4.0, log=False, low=1.0, step=None)}, trial_id=99, value=None)

P.S. Usage of the default evaluator evaluator=None does not give this issue.

[sgrubas]

A follow-up question.
Currently, by default, Optuna does not properly calculate parameters importance, it skips those with dynamic ranges (the reproducible code is exactly as I attached above). This issue has been raised quite a few times and has not been implemented yet in Optuna (stale and closed issue example optuna/optuna#1856).
I am wondering if it is possible to resolve it at least in your FanovaImportanceEvaluator, so that the importances are calculated for all the parameters including those with dynamic ranges.

Thank you!