Configuration

All genetic-algorithm behaviour is controlled by GAConfig, a dataclass passed to the selector. It validates its values on construction (raising ValueError on invalid input).

from evo_gafs import GAConfig

config = GAConfig(
    population_size=50,
    n_generations=100,
    alpha=0.8,
    cv_folds=5,
    random_seed=42,
)

Key parameters

Parameter	Default	Description
population_size	50	Number of candidate masks per generation. Larger explores more, costs more.
n_generations	100	Number of evolutionary iterations.
mode	'single'	'single' (weighted fitness) or 'multiobjective' (NSGA-II).
alpha	0.8	Performance vs compression weight in single-objective mode (see Concepts).
cv_folds	5	Cross-validation folds used to score each subset.
min_features	1	Minimum number of selected features (enforced by the repair operator).
crossover_prob / mutation_prob	0.8 / 0.15	Probabilities of crossover and mutation.
tournament_size	3	Tournament size for selection (single-objective).
elite_size	2	Best individuals carried over unchanged each generation.
early_stopping_rounds	None	Stop if the best fitness stagnates for this many generations.
n_jobs	1	Parallelism passed to scikit-learn’s cross-validation.
random_seed	42	Seed for reproducible runs.

See the full list and defaults in the GAConfig API reference.

Choosing alpha

Goal	Suggested alpha
Maximum accuracy, size irrelevant	1.0
Balanced (general use)	0.8
Edge/embedded, small models	0.5–0.7

If you are unsure which trade-off you want, use multi-objective mode and inspect the Pareto front instead.