Surrogate¶

Surrogate model training, ensemble management, and uncertainty quantification.

TrainingConfig¶

Bases: BaseModel

Configuration for surrogate model training.

Attributes:

Name	Type	Description
`n_trials`	`int`	Number of Optuna HPO trials per surrogate.
`cv_folds`	`int`	Number of cross-validation folds.
`calibration_fraction`	`float`	Fraction of data held out for conformal calibration.
`ensemble_size`	`int`	Maximum number of models in the ensemble.
`diversity_threshold`	`float`	Maximum correlation allowed between ensemble members.
`softmax_temperature`	`float`	Temperature for softmax ensemble weight selection.
`default_coverage`	`float`	Default conformal prediction interval coverage (0–1).
`estimator_families`	`tuple[EstimatorFamily, ...]`	Estimator families to search over (XGBoost, LightGBM).
`n_threads`	`int \| None`	Thread limit per model. None uses all available cores.
`study_timeout_s`	`int`	Optuna study timeout in seconds.
`min_r2`	`float \| None`	Minimum R² threshold for model quality. None disables the check.
`random_seed`	`int`	Random seed for reproducibility.

Manages trained surrogate ensembles for all target columns.

Provides prediction, uncertainty quantification, and persistence. Created via the train class method or loaded from disk via load.