tuning-hyperparameters

Tuning Hyperparameters

Overview

This skill provides workflow-level guidance for hyperparameter tuning in machine learning experiments. It covers strategy selection (grid, random, Bayesian, halving), Bayesian optimization with Optuna, search space design patterns, budget estimation, and result analysis -- the decisions and patterns that sit on top of the API-level tuning tools provided by scikit-learn.

Role in the ds plugin: This skill is the hyperparameter tuning workflow reference for the plugin. It is invoked by /ds:experiment at step 3 (Methodology Design) for strategy selection and search space design, at step 6 (Code Scaffold) for Optuna boilerplate generation, and at step 7 (Results) for tuning-specific result analysis and convergence diagnostics. Boundary with scikit-learn: scikit-learn provides the API patterns for GridSearchCV, RandomizedSearchCV, and HalvingGridSearchCV (what methods exist, their parameters, return values, and code examples). This skill provides the workflow guidance (when to use each strategy, how to design effective search spaces, how to estimate budgets, and how to analyze results). For sklearn search class API patterns, use the scikit-learn skill's references/model_evaluation.md. Boundary with experiment-tracking: experiment-tracking logs final hyperparameters and results. This skill manages the search process that produces those results. Boundary with statsmodels: time-series model order selection (ARIMA p,d,q) uses information criteria (AIC/BIC), not cross-validated search. Use the statsmodels skill for order selection; use this skill only when wrapping statsmodels models in Optuna for automated search.

When to Use This Skill

• Choosing between grid search, random search, Bayesian optimization, or successive halving
• Designing hyperparameter search spaces (distributions, ranges, conditional parameters)
• Estimating computational budget for a tuning run
• Setting up Bayesian optimization with Optuna (study, trial, objective, pruning)
• Analyzing tuning results (convergence, parameter importance, search coverage)
• Tuning XGBoost, LightGBM, or other models with Optuna's native integration

Do NOT Use This Skill

• For GridSearchCV/RandomizedSearchCV/HalvingGridSearchCV API patterns -- use scikit-learn
• For ARIMA order selection via AIC/BIC -- use statsmodels
• For feature selection (different from hyperparameter tuning) -- use scikit-learn
• For experiment logging and tracking -- use experiment-tracking

Quick Start

Strategy Selection Decision Tree

Estimate total fits: search_space_size x cv_folds = total_fits

Total Fits	Recommended Strategy	Tool
< 100	Grid search	scikit-learn GridSearchCV
100 -- 500	Random search	scikit-learn RandomizedSearchCV
> 500 or expensive evaluations	Bayesian optimization	Optuna study.optimize()
Large grid, limited budget	Successive halving	scikit-learn HalvingGridSearchCV

For details, see references/strategy_selection.md.

Bayesian Optimization with Optuna (Minimal Example)

import optuna
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score

def objective(trial):
    n_estimators = trial.suggest_int("n_estimators", 50, 300)
    max_depth = trial.suggest_int("max_depth", 3, 20)
    min_samples_split = trial.suggest_int("min_samples_split", 2, 20)

    clf = RandomForestClassifier(
        n_estimators=n_estimators,
        max_depth=max_depth,
        min_samples_split=min_samples_split,
        random_state=42,
        n_jobs=-1,
    )
    score = cross_val_score(clf, X_train, y_train, cv=5, scoring="accuracy")
    return score.mean()

study = optuna.create_study(direction="maximize")
study.optimize(objective, n_trials=100)
print(f"Best params: {study.best_params}")
print(f"Best score: {study.best_value:.4f}")

For full Optuna patterns (pruning, multi-objective, sklearn integration), see references/optuna_guide.md.

Search Space Design

# Common distribution patterns for Optuna
learning_rate = trial.suggest_float("learning_rate", 1e-5, 1e-1, log=True)  # log-uniform
n_estimators = trial.suggest_int("n_estimators", 50, 500)                    # uniform integer
max_depth = trial.suggest_int("max_depth", 3, 15)                           # uniform integer
dropout = trial.suggest_float("dropout", 0.0, 0.5)                          # uniform float
kernel = trial.suggest_categorical("kernel", ["rbf", "linear", "poly"])      # categorical

For per-model cheat sheets and conditional parameters, see references/search_space_design.md.

Common Workflows

Workflow 1: Standard Supervised Tuning

1. Estimate search space size (count parameter combinations)
2. Select strategy using the decision tree above
3. Design search space using per-model cheat sheets (references/search_space_design.md)
4. Run tuning with appropriate CV strategy (StratifiedKFold for classification, KFold for regression)
5. Analyze results (references/result_analysis.md)

Workflow 2: Bayesian Optimization for Complex Models

1. Define Optuna objective function (references/optuna_guide.md)
2. Use suggest_* methods with appropriate distributions (references/search_space_design.md)
3. Enable pruning for early stopping of unpromising trials
4. Run study with sufficient trials (50-100 minimum for convergence)
5. Analyze convergence and parameter importance (references/result_analysis.md)

Workflow 3: Budget-Constrained Tuning

1. Calculate budget: n_trials x avg_train_time x cv_folds = total_time
2. If budget tight: use successive halving (scikit-learn HalvingGridSearchCV)
3. If budget adequate: use Bayesian optimization with timeout
4. Use Optuna pruning to stop unpromising trials early

Paradigm-Specific Tuning Notes

Paradigm	CV Strategy	Notes
Supervised	StratifiedKFold (classification) or KFold (regression)	Standard tuning with all strategies
Unsupervised	KFold with internal metrics	Tuning optional; silhouette/Davies-Bouldin as objective; cluster count is most critical parameter
Temporal supervised	StratifiedKFold	Standard tuning; aeon models follow sklearn API
Temporal unsupervised	KFold with internal metrics	Same as unsupervised
Time-series forecasting	TimeSeriesSplit	MUST use temporal CV; never random k-fold; consider AIC/BIC for model order selection (statsmodels)
Anomaly detection	Custom (train on normal data only)	Tune window_size and threshold; few parameters; grid search usually sufficient

Resources

• Strategy Selection Guide -- decision tree for choosing the right strategy
• Optuna Guide -- Bayesian optimization patterns with Optuna
• Search Space Design -- per-model search space cheat sheets
• Result Analysis -- convergence diagnostics and parameter importance