Skip to content

Stop After a Time Budget ​

Time: 5 min | Difficulty: Beginner

What This Solves ​

Use TimerStopping to stop launching candidate batches once the search budget expires. Allow extra wall-clock time for a running batch, final selection and refit.

Recipe ​

python
import time
from sklearn.datasets import load_breast_cancer
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import StratifiedKFold, train_test_split

from sklearn_genetic import GASearchCV, EvolutionConfig, RuntimeConfig
from sklearn_genetic.callbacks import TimerStopping
from sklearn_genetic.space import Categorical, Continuous, Integer

X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, stratify=y, random_state=42
)

BUDGET_SECONDS = 2

ga = GASearchCV(
    estimator=RandomForestClassifier(random_state=42, n_jobs=1),
    param_grid={
        "n_estimators": Integer(50, 300),
        "max_depth":    Integer(3, 20),
        "max_features": Continuous(0.1, 1.0),
        "class_weight": Categorical([None, "balanced"]),
    },
    scoring="roc_auc",
    cv=StratifiedKFold(n_splits=5, shuffle=True, random_state=42),
    evolution_config=EvolutionConfig(
        population_size=20,
        generations=100,   # large upper limit — timer will stop it sooner
        elitism=True,
    ),
    runtime_config=RuntimeConfig(n_jobs=1, verbose=False),
    random_state=42,
)

started = time.perf_counter()
ga.fit(X_train, y_train, callbacks=[TimerStopping(total_seconds=BUDGET_SECONDS)])
elapsed = time.perf_counter() - started

last_generation = ga.history["gen"][-1]
print(f"Last logged generation {last_generation} after {elapsed:.1f}s (budget: {BUDGET_SECONDS}s)")
print(f"Best CV ROC AUC: {ga.best_score_:.4f}")

Combine Timer with Plateau Detection ​

python
from sklearn_genetic.callbacks import ConsecutiveStopping, TimerStopping

# Stop when either condition is met
ga.fit(
    X_train, y_train,
    callbacks=[
        TimerStopping(total_seconds=BUDGET_SECONDS),
        ConsecutiveStopping(generations=5, metric="fitness_best"),
    ]
)

Key Points ​

  • The initial batch runs even with a very short budget, so generation zero can be the last entry.
  • Smaller candidate batches reduce the amount of work already in progress at the deadline. With n_jobs=1, a single expensive candidate can still take longer than the entire budget.
  • Compare ga.fit_stats_["cross_validate_calls"] across runs to see how many candidate evaluations the time window allowed. The elapsed time around fit also includes refitting.

See Also ​

Released under the MIT License.