Bayesian Design of Experiments — Bayesian Optimal Experimental Design
Bayesian Optimal Design of Experiments · Also known as: Bayesian DOE, Bayesian optimal design, Bayesian experimental design, BDE
Bayesian design of experiments selects experimental runs by maximising a utility function — typically the expected information gain — computed over prior beliefs about model parameters. Unlike classical design, which optimizes algebraic criteria such as D-optimality under fixed assumptions, Bayesian DOE incorporates prior knowledge and uncertainty about the system, yielding designs that are optimal in expectation across all plausible parameter values.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Bayesian DOE when prior information about system parameters is available and ignoring it would waste resources; when the experiment is expensive (clinical trials, physical prototypes, simulation runs) and run count must be minimised; or when a sequential, adaptive strategy is needed — adding runs based on what has already been observed. It is especially powerful for nonlinear models where classical alphabetic optimality is poorly defined. Avoid it when no credible prior can be specified (a flat prior gives little advantage over classical D-optimal design), when the computational budget for optimising the utility integral is prohibitive, or when regulators require a pre-registered classical design.
Strengths & limitations
- Explicitly incorporates prior knowledge, reducing the number of runs needed to achieve a given level of inference.
- Handles nonlinear models and non-normal responses naturally, where classical optimality criteria break down.
- Supports sequential and adaptive designs: each batch of results updates the prior before the next design is chosen.
- Decision-theoretic framework allows the utility to reflect real costs and objectives beyond variance minimisation.
- Provides coherent uncertainty quantification through the posterior distribution, not just point estimates.
- Requires a well-specified prior; a poorly chosen or overconfident prior can produce a design that performs worse than a classical alternative.
- Optimising the expected utility integral is computationally intensive, often requiring Markov Chain Monte Carlo or other simulation methods.
- Less familiar to practitioners and reviewers accustomed to classical factorial or response-surface designs, which may complicate communication.
- Software support is less standardised than for classical DOE; implementations vary across R packages (acebayes, BOED) and Python libraries.
Frequently asked
How does Bayesian DOE differ from classical D-optimal design?
Classical D-optimal design minimises the determinant of the parameter-estimate covariance matrix under the assumption that the model is correct and parameters are unknown with no prior structure. Bayesian D-optimal (or Db-optimal) design averages the same criterion over a prior distribution on the parameters, so the chosen design is optimal in expectation across the range of plausible parameter values rather than for a single assumed value.
What is the most common utility function used?
The expected Kullback–Leibler (KL) divergence from prior to posterior — also called the expected information gain — is the most theoretically principled and widely used utility. It is equivalent to maximising the mutual information between the design outcome and the model parameters. Other utilities include expected variance reduction, predictive accuracy, or custom cost-weighted objectives.
Can Bayesian DOE be used with nonlinear models?
Yes, and this is one of its key advantages. Classical alphabetic optimality criteria (D, A, G) are derived from the Fisher information matrix, which requires a linear or locally-linearised model. Bayesian DOE works directly with the likelihood and prior, so it handles genuinely nonlinear models — pharmacokinetic, enzyme kinetics, or sigmoidal dose-response models — without the linearisation approximation.
Is a sequential Bayesian design the same as an adaptive design?
They are closely related. In a sequential Bayesian design, experiments are run in batches; after each batch the prior is updated to the posterior and a new optimal design is computed for the next batch. An adaptive design is the special case where batch size is one. Both allow the experiment to focus resources on the most informative regions of the factor space as data accumulate.
What software can I use to implement Bayesian DOE?
In R, the acebayes package provides functions for Bayesian optimal design via approximate coordinate exchange. The BOED Python library and the BoTorch framework support information-theoretic design for machine-learning-integrated experiments. For clinical adaptive designs, FACTS and East (commercial) or the R package bcrm are commonly used. General-purpose probabilistic programming languages such as Stan or PyMC can be combined with custom optimisation routines for bespoke designs.
Sources
- Chaloner, K., & Verdinelli, I. (1995). Bayesian Experimental Design: A Review. Statistical Science, 10(3), 273–304. DOI: 10.1214/ss/1177009939 ↗
- Ryan, E. G., Drovandi, C. C., McGree, J. M., & Pettitt, A. N. (2016). A Review of Modern Computational Algorithms for Bayesian Optimal Design. International Statistical Review, 84(1), 128–154. DOI: 10.1111/insr.12107 ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Optimal Design of Experiments. ScholarGate. https://scholargate.app/en/experimental-design/bayesian-design-of-experiments
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian Reliability AnalysisBayesian↔ compare
- Central Composite DesignExperimental design↔ compare
- Design of experimentsExperimental design↔ compare
- Response Surface MethodologyExperimental design↔ compare