Empirical performance prediction

To configure an EPM directly, pass an estimator instance:

from asf.epm import EPM
from sklearn.ensemble import RandomForestRegressor

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

EPM clones the estimator on each fit and leaves the supplied instance untouched. See Configuring base models for cloning requirements, including custom and PyTorch models.

ASF allows to easily tune and train EPMs. For example, to tune an EPM:

features, performance = get_data()

# Initialize the selector
epm = tune_epm(
    features,
    performance,
    model_class=RandomForestRegressorWrapper,
    features_preprocessing=None,
)

# Fit the selector to the data
epm.fit(features, performance)

predictions = epm.predict(features)

# Print the predictions
print(predictions)

By default, ASF uses log scaling of the performance. Other performance scaling methods include standarization, and inverse sigmoid. For more details check the API