Configuring base models

Pass a configured model instance through estimator= to model-based selectors and EPM. Model parameters belong in the estimator constructor; selector parameters such as budget and maximize belong in the selector constructor.

from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor
from asf.selectors import PairwiseClassifier, PerformanceModel
from asf.epm import EPM

selector = PairwiseClassifier(
    estimator=RandomForestClassifier(n_estimators=200, max_depth=12, random_state=42),
    budget=100,
)

performance_selector = PerformanceModel(
    estimator=RandomForestRegressor(n_estimators=200, random_state=42),
    budget=100,
)

epm = EPM(estimator=RandomForestRegressor(n_estimators=200, random_state=42))

Cloning and model lifetime

ASF uses sklearn.base.clone to create a fresh model for each training task: each algorithm pair, individual algorithm, ensemble member, or EPM fit. Repeating fit creates new models. The supplied estimator is a configuration template and is never fitted by ASF. For standard scikit-learn estimators, cloning copies constructor parameters without copying learned state, even if the supplied estimator has already been fitted. Passing a fitted estimator does not enable inference without training or warm-start training in ASF.

An integer random_state is preserved in every clone. ASF does not automatically assign different seeds to different models.

Compatible estimators

The estimator must support sklearn.base.clone and provide the methods required by the selector. Standard scikit-learn estimators, compatible pipelines, and ASF's SklearnWrapper subclasses support cloning. Custom estimators can implement the scikit-learn get_params constructor contract (usually via BaseEstimator) or __sklearn_clone__ on scikit-learn 1.3 or later. A custom clone implementation must return a fresh, unfitted model without sharing mutable training state. ASF does not fall back to deep-copying non-cloneable estimators.

Cloning does not make every estimator suitable for every selector:

• Classification selectors need classifiers, and regression selectors need regressors.
• PairwiseClassifier uses sample weights by default. Its estimator must accept sample_weight in fit, or use use_weights=False to disable weighting.
• Multi-target performance models need a regressor supporting multiple target columns.
• SimpleRanking and DyadRanking require ranking-specific fit arguments such as qid. JointRanking uses the matrix and algorithm-feature interface of RankingMLP. Survival selectors require survival prediction methods.

Raw PyTorch modules are not accepted: they do not provide the required fit/predict and cloning interface. Use a cloneable training wrapper that constructs its network and optimizer during fit, inferring dimensions from the training data. ASF's RegressionMLP and RankingMLP support cloning through their constructor parameters. Their trained networks and optimizer state are not retained in a clone. When supplying a custom network with model=, its initial weights are part of the configuration: each wrapper trains a separate copy and leaves that supplied network untouched. RankingMLP can infer its input dimension during fit when neither a network nor input_size is supplied. Custom network and loss objects must support deep copying, as required for these constructor parameters by scikit-learn's cloning mechanism.

from asf.predictors import RegressionMLP, RankingMLP
from asf.selectors import PairwiseRegressor, JointRanking

regression_selector = PairwiseRegressor(
    estimator=RegressionMLP(epochs=100, compile_model=False),
)
ranking_selector = JointRanking(
    estimator=RankingMLP(epochs=100, compile=False),
)

Built-in predictor compatibility

Predictor family	Cloning	Additional requirements
Scikit-learn wrappers, including XGBoost and its ranker	Supported	Use the appropriate classification, regression, or ranking selector; install optional XGBoost dependencies when needed.
EPMRandomForest	Supported, including forest and EPM parameters	Use return_var=False for selectors expecting one prediction per row.
EPMExtraTrees	Supported, including log and forest parameters	Always returns (means, variances); requires a consumer or adapter that handles that output.
RandomSurvivalForestWrapper	Supported	Requires scikit-survival and structured survival targets; use SurvivalAnalysis.
RegressionMLP	Supported	Requires PyTorch; use regression targets and no sample weights.
RankingMLP	Supported	Requires PyTorch and algorithm features; use JointRanking.

Cloneability alone does not adapt prediction shapes or training interfaces. In particular, a (means, variances) predictor is not a drop-in replacement for a scalar regressor in EPM or the ordinary regression selectors.

For SATzilla, estimator= configures the instance-label classifier and epm_estimator= configures the per-algorithm regressors.

Existing code and tuning

Existing model_class= selector arguments and EPM's predictor_class=, predictor_kwargs=, and predictor_config= remain supported. These create models through their existing class/factory path. ConfigSpace-based tuning continues to use that path, including nested model hyperparameters.

When estimator= is supplied, it takes precedence over the class argument. In EPM, combining estimator= with predictor_config or nonempty predictor_kwargs raises an error; put the parameters on the estimator itself. For new manually configured models, use estimator= instead of a class or functools.partial.