Does D-optimal mean maximizing uncertainty?

No. It maximizes information determinant and minimizes coefficient covariance determinant under the relevant model.

Can constraints be included?

Yes, by restricting candidates or encoding requirements in the search.

Will one optimal design serve every model?

No. The selected model terms and objective are central to its definition.

AI Glossary / Statistics and probability / Intermediate

Optimal design

Definition

Optimal design is an approach to selecting experimental runs by optimizing a stated statistical criterion under a specified model and practical constraints. Computer algorithms often search for the run settings that provide the most useful information within a budget. Optimal means best for the chosen criterion and assumptions, not universally best for every question or guaranteed to reveal the best operating condition.

What it means

An experiment does not have to test every possible combination of settings. If the goal and model are clear, a smaller, carefully chosen set of runs may estimate the quantities of interest efficiently.
The word optimal is conditional. A design that is strong for estimating model coefficients may be weaker for detecting an omitted interaction or predicting at a particular location. Changing the feasible settings, number of runs, or assumed model can change the preferred design.

How it works

Specify the response, factors, feasible settings, run budget, and model terms. Candidate settings may be restricted by equipment, cost, safety, or combinations that cannot be implemented. A design algorithm can work within such constraints rather than requiring a standard rectangular experimental region.
Choose an objective that matches the information needed. In a linear model with independent errors of equal variance, the information matrix is proportional to X-transpose-X, where X contains the model terms for the selected runs.
D-optimality maximizes the determinant of that information matrix. Equivalently, under the model assumptions, it minimizes the determinant of the estimated coefficients' covariance matrix. These are inverse quantities; maximizing coefficient uncertainty would be the wrong direction.
Other criteria emphasize different goals. A-optimality concerns the sum of coefficient variances under the specified parameterization, while prediction-oriented criteria can emphasize average or worst-case prediction variance. State the criterion instead of treating all optimal designs as interchangeable.
Search algorithms select or exchange candidate runs, but a returned solution may be locally rather than globally optimal. Multiple starts, comparison with known designs, and checks of estimability help assess the result.
Review the proposed design for replication, lack-of-fit assessment, blocking, and run-order randomization. These features may require constraints or deliberate additions. An efficient information matrix does not protect against time trends, poor measurement, or a misspecified model.

Three practical examples

Process studies: choose machine settings when some combinations are infeasible and only a limited number of runs is available.
Simulation experiments: select input combinations that estimate a surrogate model efficiently under a computational budget.
AI-assisted workflow tests: allocate a planned set of configuration comparisons to estimate prespecified effects, while retaining independent evaluation of the final choice.

Think of it this way

Think of choosing where to place measuring stations on a bridge. The best locations depend on whether you want to estimate overall bending, local stress, or an unexpected defect. A mathematically efficient placement answers the stated measurement question, not every possible question about the bridge.

Worked example

Consider an original four-run experiment for the model y = a + bx + error. Allowed settings are -1, 0, and 1, replication is permitted, and errors are independent with common variance sigma squared.
Design A uses -1, -1, 1, 1. Its input sum is zero and squared-input sum is four, so X-transpose-X is diagonal with entries 4 and 4. Its determinant is 16.
Design B uses -1, 0, 0, 1. Its corresponding diagonal entries are 4 and 2, giving determinant 8. Under the stated model, slope variance is sigma squared/4 for A and sigma squared/2 for B. A therefore halves slope variance relative to B.
For any four allowed settings, the determinant is 4 times the sum of squared inputs minus the squared input sum. The squared-input sum cannot exceed four, so the determinant cannot exceed 16. Design A reaches that bound and is D-optimal for this problem.
However, A cannot distinguish an x-squared effect from the intercept because every run has x-squared equal to one. Adding curvature to the intended model changes the design problem.

Illustrative case study

Fictional case: Alder Materials plans a small calibration study with limited instrument time. A design program selects extreme settings that efficiently estimate a straight-line response.
An engineer raises a credible concern about curvature, so the team revises the model and includes intermediate settings. It compares information criteria while preserving replicated runs and randomized execution. The revised plan sacrifices the original straight-line optimum to answer the broader scientific question more adequately.

Common misconceptions and mistakes

Optimal design does not mean choosing settings already known to produce the highest response. Learning about a response and maximizing it operationally are different objectives.
Do not assume a computer-generated design is globally optimal. Search details, starting points, and numerical tolerances can matter.
An orthogonal design can sometimes be optimal under a particular criterion; optimal-design methods and classical designs are not mutually exclusive categories. Likewise, a nonorthogonal design is not automatically defective.
Most importantly, an optimum for the wrong model may provide misleadingly precise estimates. Include plausible terms and consider robustness before relying on an efficiency score.

Business use

Businesses can use optimal design to spend experimental resources where they provide the most relevant information. The planning discussion should connect the criterion to decisions, practical constraints, and assumptions about the process. Compare the proposed savings with the need for validation, replication, and diagnostic checks. Record the design specification so later analysts understand why particular settings were chosen.

Check your understanding

Why is Design A better for the example's slope estimate?

Answer: Its greater spread in x increases information about slope under equal error variance.

Why is it unsuitable for separately estimating curvature?

Answer: Its x-squared column is identical to the intercept column, so those effects cannot be separated.

Related terms

Replication (statistics)
experimental design
D-optimality
information matrix
A-optimality
estimability
design space
replication
model misspecification

Learn next

compare designs under both a linear and a quadratic model, checking matrix rank before comparing any optimality scores.

Model checking

Sources

Sources: NIST/SEMATECH, D-Optimal designs: https://www.itl.nist.gov/div898/handbook/pri/section5/pri521.htm . NIST glossary seed: https://www.itl.nist.gov/div898/handbook/glossary.htm . The detailed chapter correctly states information maximization; the glossary reverses the covariance-determinant direction. That discrepancy is corrected here. Worked design and fictional case are original.

Previous: Optimal controllers Next: Optimal discriminant analysis and classification tree analysis