Machine Learning

Model Selection and Hyperparameter Search

Model selection turns many candidate settings into one chosen procedure; dev/CV may choose, while test stays untouched for final evidence.

status: publishedimportance: criticaldifficulty 3/5math: undergraduateread: 20mlive demo

Back to Machine Learning Next: Evaluation Pipelines

Concept Structure

Model Selection and Hyperparameter Search

01Intuition

Start with the picture, metaphor, or geometric mechanism.

02Math

Make the objects explicit and connect them with notation.

03Code

Mirror the equations with runnable implementation details.

04Interactive Demo

Manipulate the mechanism and watch the idea respond.

3prerequisites

2next concepts

3related links

Learning map

Model Selection and Hyperparameter Search

BeforeTrain/Dev/Test Splits, Cross-Validation, and LeakageNow4/4 sections readyTryManipulate one control and predict the visible change.NextEvaluation Pipelines

Object flow

4/4 sections readyAsk about this Research room

ConceptModel Selection and Hyperparameter SearchMachine Learning EquationModel Selection and Hyperparameter Search equation 1Exact equation object CodeModel Selection and Hyperparameter Search code witness 1Exact code witness DemoModel Selection and Hyperparameter Search interactive demoVisualization object ClaimHyperparameter search is model selection: candidates may be fit on tr...Exact claim check SourceCS229 Notes: Regularization and Model SelectionExact source object

ConceptModel Selection and Hyperparameter SearchMachine Learning

5 sources attachedLocal snapshot ready

concept:machine-learning/model-selection-hyperparameter-search

Codewitness nearby Predictbefore reveal Roomobject handoff

Conceptual Bridge

What should feel connected as you move through this page.

Carry inTrain/Dev/Test Splits, Cross-Validation, and Leakage

Bring the mental model from Train/Dev/Test Splits, Cross-Validation, and Leakage; this page will reuse it instead of restarting from zero.

Work hereModel Selection and Hyperparameter Search

Model selection turns many candidate settings into one chosen procedure; dev/CV may choose, while test stays untouched for final evidence.

Carry outEvaluation Pipelines

The next edge should feel earned: use the demo prediction here before following Evaluation Pipelines.

Test the linkManipulate one control and predict the visible change.Then continue to Evaluation Pipelines

01IntuitionStart with the picture, metaphor, or geometric mechanism.02MathMake the objects explicit and connect them with notation.03CodeMirror the equations with runnable implementation details.04Interactive DemoManipulate the mechanism and watch the idea respond.

Intuition

Build the mental picture first so the rest of the page has something to attach to.

Section prompt

You are here because training one model is rarely the real experiment. In practice, you try many settings: polynomial degree, regularization strength, feature count, threshold, architecture size, prompt template, learning rate, batch size, or decoding rule. Model selection is the discipline that keeps that search from quietly becoming test-set memorization.

Before this, know train/dev/test splits, regularization, and why metrics can hide different costs. By the end, you should be able to run a small grid search, explain what the development set is allowed to choose, and say why the test set must stay untouched until the selected procedure is frozen.

Think of a hyperparameter grid as a menu of possible procedures:

degree 2, ridge $\lambda=1$
degree 5, ridge $\lambda=0.1$
degree 8, ridge $\lambda=100$
many more

Each candidate gets fit on training data. Then a development set or cross-validation estimate chooses which candidate to keep. That chosen candidate is not "the model that happened to have the lowest visible number." It is the result of a search procedure.

The dangerous move is test peeking:

fit many candidates on train
look at the test scores
choose the candidate with the best test score
report that same best test score

That score is no longer a final estimate. It is the score you optimized over. With enough candidates, one can look unusually good on the test set by luck. The model may not be better; you may have selected a lucky error bar.

The repair is simple and strict:

train data fits parameters
dev or cross-validation chooses hyperparameters
test data estimates the already chosen procedure once

If you need to tune and estimate performance on scarce data, use an outer evaluation loop such as nested cross-validation or keep a separate audit set. The invariant is not "never search." The invariant is "never report the data you searched over as if it were untouched evidence."

Math

Translate the story into symbols, assumptions, and a derivation you can inspect.

Section prompt

Equation 1\hat f_\lambda=A_\lambda(D_{\text{train}}).Equation 2\widehat R_{\text{dev}}(\lambda) = \frac{1}{n_{\text{dev}}} \sum_{(x_i,y_i)\in D_{\text{dev}}...

Let $\Lambda$ be a finite set of hyperparameter settings. A setting $\lambda\in\Lambda$ might encode degree, penalty strength, feature count, threshold, or an architecture choice.

For each candidate, the training algorithm fits parameters using only training data:

\hat f_\lambda=A_\lambda(D_{\text{train}}).

The development risk estimate is

\widehat R_{\text{dev}}(\lambda) = \frac{1}{n_{\text{dev}}} \sum_{(x_i,y_i)\in D_{\text{dev}}} L(\hat f_\lambda(x_i),y_i).

Model selection chooses the setting with the lowest development estimate:

\hat\lambda = \arg\min_{\lambda\in\Lambda} \widehat R_{\text{dev}}(\lambda).

Only after $\hat\lambda$ is fixed do we evaluate once on test:

\widehat R_{\text{test}} = \frac{1}{n_{\text{test}}} \sum_{(x_i,y_i)\in D_{\text{test}}} L(\hat f_{\hat\lambda}(x_i),y_i).

This estimates the selected procedure, not every candidate separately. The test set did not choose $\hat\lambda$ ; it only measured the procedure after selection.

Cross-validation replaces one development split with several rotating validation folds. For $K$ folds $F_1,\ldots,F_K$ inside the training budget,

\widehat R_{\text{cv}}(\lambda) = \frac{1}{n_{\text{train}}} \sum_{k=1}^{K} \sum_{i\in F_k} L\!\left(A_\lambda(D_{\text{train}}\setminus F_k)(x_i),y_i\right).

Then choose

\hat\lambda_{\text{cv}} = \arg\min_{\lambda\in\Lambda} \widehat R_{\text{cv}}(\lambda).

The crucial detail is fold locality. Preprocessing, feature selection, early stopping, calibration, and threshold choice must be learned inside each training fold when they are part of the candidate procedure.

Test peeking changes the selection rule to

\tilde\lambda = \arg\min_{\lambda\in\Lambda} \widehat R_{\text{test}}(\lambda).

The reported value

\min_{\lambda\in\Lambda}\widehat R_{\text{test}}(\lambda)

is optimistically biased as an estimate of future performance because the minimum selected a favorable noise realization. More candidates usually create more chances to find a lucky low estimate. That does not mean large grids are forbidden. It means the grid must be selected using development/CV evidence and evaluated with untouched data.

Nested cross-validation separates the two jobs when no external test set is available. The inner loop chooses $\lambda$ ; the outer loop estimates the performance of the whole selection procedure.

Code

Keep the implementation aligned with the notation so the algorithm is legible.

Section prompt

Code witness 1import numpy as np rng = np.random.default_rng(4) degrees = range(1, 9) lambdas = [0.1, 1.0,...python

import numpy as np

rng = np.random.default_rng(4)
degrees = range(1, 9)
lambdas = [0.1, 1.0, 10.0, 100.0]

def true_risk(degree, lam):
    log_lambda = np.log10(lam)
    return 0.12 + 0.015 * (degree - 4) ** 2 + 0.012 * (log_lambda - 0.3) ** 2

rows = []
for degree in degrees:
    for lam in lambdas:
        risk = true_risk(degree, lam)
        dev = risk + rng.normal(0, 0.035)
        test = risk + rng.normal(0, 0.035)
        audit = risk + rng.normal(0, 0.008)
        rows.append((dev, test, audit, degree, lam))

by_dev = min(rows, key=lambda row: row[0])
by_test = min(rows, key=lambda row: row[1])

for name, row in [("dev-selected", by_dev), ("test-peeked", by_test)]:
    print(name, {
        "degree": row[3],
        "lambda": row[4],
        "dev": round(float(row[0]), 3),
        "test": round(float(row[1]), 3),
        "audit": round(float(row[2]), 3),
    })

print("test-peek optimism:", round(by_test[2] - by_test[1], 3))

The code simulates noisy estimates for a degree/lambda grid. The development-selected candidate is the clean protocol. The test-peeked candidate is the invalid protocol: it chooses the lowest test number and then tries to report that same number as final evidence. The audit column stands in for fresh data that did not participate in the search.

Interactive Demo

Use direct manipulation to connect the explanation to a moving system.

Section prompt

Choose the search width and estimate noise. Before revealing the grid, predict which protocol will survive the fresh audit: development selection, test peeking, or a small-grid tie.

The grid shows candidate degree/lambda settings while the scores stay hidden. After reveal, the heatmap appears: one marker shows the candidate chosen by development score and another shows the candidate chosen by repeatedly reading test scores. The point is not that test peeking always chooses a more complex model. The point is sharper: once the test set chooses, its score is no longer untouched evidence.

Live Concept Demo

Explore Model Selection and Hyperparameter Search

The stage is code-native and interactive. Use it to test the explanation against the mechanism.

difficulty 3/5undergraduatecode-aligned

Demo Prediction Checkpoint

Manipulate one control and predict the visible change.

Commit to what Model Selection and Hyperparameter Search should make visible before reading the result.

After The First Pass

Turn the concept into an inspected object.

Once the invariant is visible in the intuition, math, code, and demo, use these panels to inspect the mechanism visually, check source support, practice the idea, and attach a grounded research question.

Mechanism Storyboard

See the idea move before the page explains it

Model selection turns many candidate settings into one chosen procedure; dev/CV may choose, while test stays untouched for final evidence.

Prediction open01 / Intuition

Prediction lens

Start with the picture, metaphor, or geometric mechanism.

Commit first

Before reading further, choose the kind of change Model Selection and Hyperparameter Search should make visible.

Visual Inquiry

Make the image answer a mathematical question

Model selection turns many candidate settings into one chosen procedure; dev/CV may choose, while test stays untouched for final evidence.

4/4 stages readyLive demo connected

Visual cueWhich visible object should carry the first intuition?

Inspection depth2/4

Prediction

Which visible object should carry the first intuition?

Commit first

Pick the cue that should make Model Selection and Hyperparameter Search easier to reason about before the page gives the answer.

Source Grounding

Canonical references for the mechanism on this page.

course-notes · 2019CS229 Notes: Regularization and Model SelectionStanford CS229

Source for hold-out cross-validation, k-fold cross-validation, and choosing among model classes or hyperparameters using held-out empirical error.

Open source

book · 2023An Introduction to Statistical LearningJames, Witten, Hastie, Tibshirani, and Taylor

Source for validation-set and cross-validation approaches to estimating test error for model assessment and model selection.

Open source

documentation · 2026scikit-learn User Guide: Tuning the hyper-parameters of an estimatorscikit-learn developers

Source for grid search, randomized search, cross-validated estimators, and the idea that hyperparameters are set before fitting model parameters.

Open source

documentation · 2026scikit-learn Example: Nested versus non-nested cross-validationscikit-learn developers

Source for the warning that using the same data to tune and evaluate can produce optimistically biased scores.

Open source

paper · 2010On Over-fitting in Model Selection and Subsequent Selection Bias in Performance EvaluationCawley and Talbot

Source for model-selection overfitting and biased performance evaluation when selection variance is ignored.

Open source

Claim Review

Model selection turns many candidate settings into one chosen procedure; dev/CV may choose, while test stays untouched for final evidence.

Status1 substantive review recorded

Claims without a substantive review badge still need exact source-support review.

Sources5 references

cs229-regularization-model-selection, islr-resampling-model-selection, sklearn-grid-search, sklearn-nested-cross-validation, cawley-talbot-selection-bias

Witnesses4 local objects

Use equation, code, and demo objects to check whether the source support is operational.

Substantively reviewedHyperparameter search is model selection: candidates may be fit on training data and compared on development or cross-validation scores, but repeated test-set peeking turns the test score into a tuned selection signal and makes the reported performance optimistic.Claim metadata: source checked

CS229 and ISLR support held-out/CV model selection; scikit-learn supports grid/randomized search and warns through nested-CV examples that tuning and evaluating on the same data biases estimates; Cawley and Talbot support selection overfitting and subsequent selection bias.

Sources: CS229 Notes: Regularization and Model Selection, An Introduction to Statistical Learning, scikit-learn User Guide: Tuning the hyper-parameters of an estimator, scikit-learn Example: Nested versus non-nested cross-validation, On Over-fitting in Model Selection and Subsequent Selection Bias in Performance EvaluationThe code and demo use finite toy loss estimates for a degree/lambda grid; they do not claim a universal bias magnitude, an optimal search algorithm, or that nested CV is always required for every operational setting.A bounded review summary is present; still check caveats and exact source scope.

Substantive local review after three GPT Pro/Oracle stalls. CS229, ISLR, scikit-learn, and Cawley-Talbot support the model-selection/test-peeking contract. Claude alternate review found no math/source/code/demo blockers after documented desktop/mobile QA; the toy witness caveat remains. Evidence: responses/model-selection-browser-source-qa-20260628.md.

Reviewer: codex-source-audit+claude-alt-review; reviewed 2026-06-28

source-span-cs229-regularization-model-selection source-span-islr-resampling-model-selection source-span-sklearn-grid-search source-span-sklearn-nested-cross-validation source-span-cawley-talbot-selection-bias math-object-2 code-witness-1 interactive-demo

Source support candidates

course-notes 2019CS229 Notes: Regularization and Model Selection

Source for hold-out cross-validation, k-fold cross-validation, and choosing among model classes or hyperparameters using held-out empirical error.

book 2023An Introduction to Statistical Learning

Source for validation-set and cross-validation approaches to estimating test error for model assessment and model selection.

documentation 2026scikit-learn User Guide: Tuning the hyper-parameters of an estimator

Source for grid search, randomized search, cross-validated estimators, and the idea that hyperparameters are set before fitting model parameters.

documentation 2026scikit-learn Example: Nested versus non-nested cross-validation

Source for the warning that using the same data to tune and evaluate can produce optimistically biased scores.

Mechanism witnesses

Equation 1

\hat f_\lambda=A_\lambda(D_{\text{train}}).

Equation 2

\widehat R_{\text{dev}}(\lambda) = \frac{1}{n_{\text{dev}}} \sum_{(x_i,y_i)\in D_{\text{dev}}} L(\hat f_\lambda(x_i),y_i).

Code witness 1import numpy as np rng = np.random.default_rng(4) degrees = range(1, 9) lambdas = [0.1, 1.0,...Demo stateLive mechanism probe

Practice Loop

Try the idea before it explains itself

Model selection turns many candidate settings into one chosen procedure; dev/CV may choose, while test stays untouched for final evidence.

Readiness0/3 checks ready

Predict

Before touching the demo, predict one visible change that should happen in Model Selection and Hyperparameter Search.

Hint 1

Reveal when your model needs a nudge.

Hint 2

Reveal when your model needs a nudge.

Hint 3

Reveal when your model needs a nudge.

Your answer canvas

Local checks

Claim

A concrete answer is on the canvas.

Mechanism

The answer names why the claim should hold.

Bridge

It touches the page context or a neighboring idea.

Misconception check

Object research drawerClose

ConceptModel Selection and Hyperparameter SearchMachine Learning

Code witness comparisonModel Selection and Hyperparameter Search code witness 1rng = np.random.default_rng(4)Prediction before revealModel Selection and Hyperparameter Search interactive demoManipulate one control and predict the visible change.

Grounded room questionWhat is the smallest example that makes Model Selection and Hyperparameter Search click without losing the math?Local snapshot ready

Research Room

Attach the question to an exact object

Pick the concept, equation, source, code witness, claim, misconception, or demo state before asking for help. The handoff stays grounded to that object.

Next local actionNo local draft saved yet

Open the draft below to save one note and next action in this browser.

conceptMachine Learning

Model Selection and Hyperparameter Search

Anchored question

What is the smallest example that makes Model Selection and Hyperparameter Search click without losing the math?

Local action draftNo local draft saved yetExpand only when ready to capture one local next action

Local action draft

This draft stays locally in this browser for concept:machine-learning/model-selection-hyperparameter-search.

Draft noteNext action

No local draft saved.

Evidence to inspect

Source ids to inspect: cs229-regularization-model-selection, islr-resampling-model-selection, sklearn-grid-search, sklearn-nested-cross-validation, cawley-talbot-selection-bias
Definition, prerequisite, and contrast concept links
The equation or code witness that makes the concept operational
One demo state that shows the invariant instead of a slogan

What would resolve this

The learner can state the mechanism in their own words
The learner can name the prerequisite that would repair confusion
The learner can predict how the mechanism changes under one perturbation

Grounded AI handoff

I am working in Continuous Function's research reading room. Object: concept - Model Selection and Hyperparameter Search Object key: concept:machine-learning/model-selection-hyperparameter-search Context: Machine Learning Anchor id: concept/concept-notebook/machine-learning/model-selection-hyperparameter-search Open question: What is the smallest example that makes Model Selection and Hyperparameter Search click without losing the math? Evidence to inspect: - Source ids to inspect: cs229-regularization-model-selection, islr-resampling-model-selection, sklearn-grid-search, sklearn-nested-cross-validation, cawley-talbot-selection-bias - Definition, prerequisite, and contrast concept links - The equation or code witness that makes the concept operational - One demo state that shows the invariant instead of a slogan What would resolve this: - The learner can state the mechanism in their own words - The learner can name the prerequisite that would repair confusion - The learner can predict how the mechanism changes under one perturbation Answer as a careful research tutor: stay source-grounded, separate verified evidence from assumptions, name the relevant math objects, and end with one next action.

Open source object

concept/concept-notebook/machine-learning/model-selection-hyperparameter-search
concept:machine-learning/model-selection-hyperparameter-search

Learning Map

Before / Now / Try / Next

BeforeTrain/Dev/Test Splits, Cross-Validation, and Leakage

NowIntuition → Math → Code → Demo

TryManipulate one control and predict the visible change.

NextEvaluation Pipelines

Intuitionready
Mathready
Codeready
Interactive Demoready

Object Companion

Ask beside the selected object

Model selection turns many candidate settings into one chosen procedure; dev/CV may choose, while test stays untouched for final evidence.

Your question

GoalComfortStyleStuck on

Context prompt

You are my AI learning companion for Continuous Function. Current context: Machine Learning concept. Learning surface: Model Selection and Hyperparameter Search. What this page says: Model selection turns many candidate settings into one chosen procedure; dev/CV may choose, while test stays untouched for final evidence. Current section: Intuition, math, code, and interactive demo. Suggested next step: Manipulate one control and predict the visible change.. Learner goal: Understand the idea. Learner comfort level: New to this. Preferred explanation style: Visual first. Task: Explain the central idea in plain language, then restate it with the exact math objects from the page. Answer in a way that helps me learn: ask one clarifying question only if needed, use intuition before notation, and end with one thing I should try on the page.

Domain

Machine Learning

machine-learningmodel-selectionhyperparameter-searchcross-validationevaluation

Prerequisites

Train/Dev/Test Splits, Cross-Validation, and Leakage Regularization: Ridge, Lasso, and Elastic Net Classification Metrics, Thresholds, and Calibration

Leads To

Evaluation PipelinesBootstrap Jackknife Resampling (planned)

Bias-Variance Decomposition Linear Regression & Least Squares Logistic Regression

Within this domain