---

name: learners-and-committees description: "Router for ActiveLearner, Committee, and CommitteeRegressor workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

learners-and-committees

Use this sub-skill for:

• ActiveLearner workflows for pool-based and stream-based active learning.
• Committee workflows for query by committee, voting, class/proba reconciliation, and bagging.
• CommitteeRegressor workflows for ensemble regression and standard-deviation query selection.

Keep the task here when it is about learner lifecycle, single-row teach/query shape handling, bagging, or committee disagreement. If the task is mainly about query-strategy formulas, selectors, or combinators, route to ../query-strategies/SKILL.md. If it is Bayesian optimization or deep/skorch/Keras/PyTorch, route elsewhere.

Read first

• references/workflows.md — initialization, query/teach/fit loops, stream sampling, committee bagging, and regression recipes.
• references/api-reference.md — signatures, return shapes, estimator requirements, and single-row handling.
• references/troubleshooting.md — recovery notes for shape mismatches, unfitted estimators, missing predict_proba, on_transformed, return_metrics, and only_new.

Run

• scripts/active_learning_smoke.py — deterministic pool, stream, bootstrap_init, bootstrap, only_new, fit, return_metrics, and on_transformed smoke.
• scripts/committee_smoke.py — deterministic committee vote/proba alignment, rebag, bagging, and CommitteeRegressor standard-deviation smoke.

Route elsewhere

• Query-strategy formulas, selectors, or combinators: ../query-strategies/SKILL.md
• BayesianOptimizer and acquisition loops: ../bayesian-optimization/SKILL.md
• DeepActiveLearner, skorch, Keras, PyTorch, or dropout helpers: ../deep-and-optional-integrations/SKILL.md

Quick rules

• Use an estimator with fit and predict; add predict_proba when the strategy needs probabilities.
• Teach/query single rows with the same shape style as the original training data. For NumPy, reshape single samples with (1, -1) and labels with (1,).
• bootstrap_init=True bootstraps the first fit; teach(..., bootstrap=True) bootstraps the refit; rebag() refits each learner on a bootstrap sample of its own history.
• only_new=True fits only the fresh batch and does not append it to stored training history.
• Committee reconciles class labels across learners; CommitteeRegressor uses ensemble prediction standard deviation instead of class probabilities.
• There is no public bag() method in this release; use the bootstrap flags and rebag() instead.