[vrjdev]

Description

With a partial ml_param_space, DoubleML.tune_ml_models installed parameters even for learners it did not tune. For those, _dml_tune_optuna returns get_params(deep=True) of a clone fitted on the full data, and tune_ml_models set these as fold-specific nuisance parameters.

• For a Pipeline: the parameters contain that clone's fitted step objects, so all folds shared them. Every fold was then predicted by the last fold's fit, i.e. in-sample.
• For any learner: parameters set earlier with set_ml_nuisance_params were overwritten with the constructor values.

Changes:

• doubleml/double_ml.py: tune_ml_models only sets the parameters of learners that were actually tuned (tuned_result.tuned); untuned learners keep their current parameters. This matches the scalar API (DoubleMLScalar.tune_ml_models), which only applies results with result.tuned.
  • Tuned learners are handled as before.
  • The returned tuning results are unchanged.
  • Skipping tuned_result is None doesn't change behaviour: set_ml_nuisance_params(..., None) already kept the existing parameters.
• doubleml/tests/test_dml_tune_optuna.py: two new tests, both of which fail before the fix.
  • A partially tuned PLR with an untuned Pipeline ml_l gives the same predictions, coefficient and standard error as the same model with the tuned ml_m parameters set directly, on the same splits.
  • Parameters set earlier for the untuned learner are kept.

Reference to Issues or PRs

Fixes #444

Comments

This changes what params shows after a partial tune_ml_models call. An untuned learner now keeps its previous entry (None, or the values set earlier) instead of a full copy of its constructor parameters. The fitted model only differs when the bug was hit.

PR Checklist

• The title of the pull request summarizes the changes made.
• The PR contains a detailed description of all changes and additions.
• References to related issues or PRs are added.
• The code passes all (unit) tests.
• Enhancements or new feature are equipped with unit tests.
• The changes adhere to the PEP8 standards.