HoeffdingRaceRegressor¶

Hoeffding Race model selection for regression.

Uses the Hoeffding bound to progressively eliminate underperforming models with statistical confidence. All active models are trained on each observation. After each observation, models whose performance is statistically worse than the current best are eliminated from the race.

This is more efficient than greedy selection when there are many candidate models, because eliminated models no longer consume computation. It is more principled than successive halving because elimination is based on statistical evidence rather than a fixed schedule.

Parameters¶

• models

  Type → list

  The models to race against each other.

• metric

  Type → metrics.base.RegressionMetric | None

  Default → None

  Metric used for comparing models. Defaults to metrics.MAE.

• delta

  Type → float

  Default → 0.05

  Confidence parameter. Smaller values make elimination harder (more conservative). The algorithm is correct with probability at least 1 - delta. Defaults to 0.05.

• loss_range

  Type → float

  Default → 1.0

  The range of the per-observation loss values. For unbounded metrics you should set this to a reasonable upper bound. Defaults to 1.0.

Attributes¶

• active_models

  The models still in the race.

• best_model

  The current best model.

• models

• n_active_models

  The number of models still in the race.

Examples¶

from river import datasets
from river import evaluate
from river import linear_model
from river import metrics
from river import model_selection
from river import optim
from river import preprocessing

models = [
    linear_model.LinearRegression(optimizer=optim.SGD(lr=lr))
    for lr in [1e-5, 1e-4, 1e-3, 1e-2]
]

dataset = datasets.TrumpApproval()
metric = metrics.MAE()
model = (
    preprocessing.StandardScaler() |
    model_selection.HoeffdingRaceRegressor(
        models,
        metric=metric,
        delta=0.05,
        loss_range=50.0,
    )
)

evaluate.progressive_val_score(dataset, model, metric)

MAE: 1.319678

You can inspect how many models remain in the race:

model["HoeffdingRaceRegressor"].n_active_models

4

Methods¶

References¶

• Maron, O. and Moore, A.W., 1993. Hoeffding Races: Accelerating Model Selection Search for Classification and Function Approximation. Advances in Neural Information Processing Systems, 6.