Bayesian Optimal Experimental Design

• Bayesian optimal experimental design is a decision-theoretic framework that chooses experiments to maximize expected information gain about latent parameters and models.
• It leverages nested Monte Carlo and variational methods to estimate mutual information, addressing challenges posed by intractable likelihoods.
• It extends to sequential, simulation-based, and likelihood-free settings, enhancing decision-making in fields like robotics, cognitive science, and material characterization.

Bayesian optimal experimental design (BOED) is a Bayesian decision-theoretic framework for choosing experiments that are expected to be maximally informative about unknown quantities of interest. In its classical form, BOED selects a design by maximizing expected information gain (EIG), equivalently the mutual information between latent variables and future observations under a prior and a design-dependent data-generating model (Foster et al., 2019). Contemporary work retains this core formulation while extending it to sequential policies, simulator-based models with intractable likelihoods, decision-focused utilities, robustness under model misspecification, and geometry-aware alternatives to Kullback–Leibler-based design criteria (Foster et al., 2021, Go et al., 25 May 2026, Wu et al., 23 Apr 2026).

1. Formal foundations

The standard BOED setup assumes latent variables or parameters $\theta$, a controllable design $d$ or $\xi$, and future observations $y$, with prior $p(\theta)$ and likelihood $p(y\mid \theta,d)$. The canonical utility is the expected reduction in uncertainty about $\theta$: $$EIG(d)\triangleq \mathbb{E}_{p(y\mid d)}\big[H[p(\theta)]-H[p(\theta\mid y,d)]\big].$$ Equivalent mutual-information forms are

$$EIG(d)=\mathbb{E}_{p(y,\theta\mid d)}\left[\log \frac{p(\theta\mid y,d)}{p(\theta)}\right] =\mathbb{E}_{p(y,\theta\mid d)}\left[\log \frac{p(y\mid \theta,d)}{p(y\mid d)}\right],$$

and the optimal design is

$d^*=\argmax_{d\in\mathcal D} EIG(d).$

This identifies BOED with maximizing $d$0 (Foster et al., 2019).

A useful generalization treats the “variable of interest” as an arbitrary $d$1, not only a parameter vector. In that notation,

$d$2

This makes BOED directly applicable both to model discrimination, where $d$3 is a discrete model indicator, and to parameter estimation, where $d$4 is a continuous parameter vector (Valentin et al., 2021).

BOED is naturally divided into static and sequential forms. In static BOED, the full design is fixed before data collection. In sequential BOED, designs are selected adaptively as observations arrive. A policy $d$5 maps the current history $d$6 to the next design, and the total information of an adaptive experiment can be written as

$d$7

This reframes adaptive BOED as maximizing the mutual information between $d$8 and the full experimental history $d$9, not merely the next-step gain (Foster et al., 2021).

2. Computational structure and intrinsic difficulty

The practical difficulty of BOED lies in the nested structure of EIG. Even for a fixed candidate design, one must integrate over latent parameters, hypothetical observations, and posterior distributions. Direct Monte Carlo is obstructed because the integrand involves either the posterior $\xi$0 or the evidence $\xi$1, both of which are usually unavailable in closed form (Foster et al., 2019).

A standard baseline is nested Monte Carlo. For explicit-likelihood models, one estimator is

$\xi$2

with $\xi$3 and $\xi$4. This estimator is consistent, but its RMSE scales as $\xi$5, and under optimal allocation its total-cost rate is only $\xi$6, much slower than ordinary Monte Carlo (Foster et al., 2019).

The difficulty compounds in simulator-based science. In many realistic settings one can sample

$\xi$7

without being able to evaluate $\xi$8 itself. In that regime, posterior densities, marginal likelihoods, and mutual information are all intractable. Moreover, design optimization becomes a nested problem: an inner problem estimates informativeness at a fixed design from simulations, and an outer problem searches over the design space. When observations are discrete, as in sequential behavioral tasks, gradients with respect to design through simulator outputs may not be defined, so even continuous-design stochastic-gradient BOED is unavailable without additional approximation machinery (Valentin et al., 2021).

These facts motivate most modern BOED research. One line of work builds better EIG estimators; another changes the design optimizer; a third changes the BOED utility itself.

3. EIG estimation and design optimization methodologies

A major development is amortized variational estimation of EIG. One family of estimators learns reusable approximations to the posterior, the marginal predictive, or an importance proposal. The variational posterior estimator uses

$\xi$9

with gap

$y$0

The variational marginal estimator uses

$y$1

with

$y$2

Variational NMC (VNMC) retains asymptotic exactness by increasing the number of proposal samples, while implicit-likelihood settings can be handled by jointly learning surrogate marginal and likelihood terms (Foster et al., 2019).

A second line replaces two-stage “estimate EIG, then optimize design” pipelines with unified stochastic-gradient BOED. In this formulation one directly optimizes a lower bound $y$3 over both design parameters $y$4 and variational parameters $y$5. The paper “A Unified Stochastic Gradient Approach to Designing Bayesian-Optimal Experiments” develops this idea around the Barber–Agakov bound, Prior Contrastive Estimation (PCE), and the adaptive contrastive estimation (ACE) bound

$y$6

ACE is a lower bound, tight if $y$7 matches the posterior, monotone in the number of contrastive samples $y$8, and asymptotically exact as $y$9 (Foster et al., 2019).

Recent SBI-based BOED work argues that poor design optimization itself is a major bottleneck. “Supercharging Simulation-Based Inference for Bayesian Optimal Experimental Design” introduces multi-start parallel gradient ascent for per-trajectory design search and reports that, with this optimization change, SBI-based BOED methods can match or outperform by up to $p(\theta)$0 existing state-of-the-art approaches across standard BOED benchmarks (Klein et al., 6 Feb 2026). A complementary direct-gradient route introduces a pooled posterior

$p(\theta)$1

which minimizes the weighted average forward KL to the individual posteriors and enables a single-loop sampling-optimization algorithm for direct EIG gradients (Iollo et al., 2024).

When forward simulation cost dominates, BOED can be accelerated at the model-evaluation layer rather than only at the EIG-estimator layer. In soft-material characterization, local radial basis function surrogates were used to replace most expensive forward solves while keeping a large prior sample set for EIG estimation; in a five-model problem in a two-dimensional design space, EIG estimates were obtained at just $p(\theta)$2 of the full computational cost (Chu et al., 19 May 2025).

4. Sequential, amortized, and likelihood-free BOED

Sequential BOED traditionally recomputes a one-step EIG objective after each observation. An alternative is amortization over entire experiments. Deep Adaptive Design (DAD) learns a deterministic policy $p(\theta)$3 offline and optimizes the total mutual information of the whole experiment rather than only the immediate next step. Its sequential contrastive bounds, including sPCE and sNMC, support training by simulated rollouts, after which live design decisions reduce to a single forward pass and can be made in milliseconds during deployment (Foster et al., 2021).

Likelihood-free BOED has been developed by combining mutual-information lower bounds with simulation-based inference. In cognitive simulator models, BOED has been implemented with the NWJ lower bound

$p(\theta)$4

paired with learned blockwise summary-statistic networks for sequential behavioral data and Bayesian optimization over discrete reward-schedule designs (Valentin et al., 2021). This yields a single learned object that serves both for design evaluation and for approximate posterior recovery.

In SBI settings with non-differentiable simulators, one strategy is to replace the intractable likelihood by an amortized conditional density model $p(\theta)$5 and insert that surrogate into a contrastive EIG lower bound. The resulting likelihood-free PCE objective supports joint optimization of the design $p(\theta)$6 and the inference network $p(\theta)$7 (Zaballa et al., 2023). Subsequent work generalized this perspective by showing that different EIG identities naturally align with neural posterior estimation, neural likelihood estimation, and neural ratio estimation, thereby turning SBI and BOED into a unified design-and-inference pipeline rather than two separate procedures (Zaballa et al., 11 Feb 2025, Klein et al., 6 Feb 2026).

A recurring theme in this literature is that simulator-based BOED often requires an additional exploration mechanism in design space. Design distributions, design checkpoints, Bayesian optimization over noisy EIG surrogates, and multi-start gradient ascent are all used to prevent optimization from collapsing onto poor local optima in sparse-reward or flat-EIG regimes (Zaballa et al., 11 Feb 2025, Klein et al., 6 Feb 2026).

5. Scientific applications and domain-specific formulations

BOED has been deployed in a wide range of scientific settings, and the design variable changes meaning across domains. In cognitive science, BOED has been used to design multi-armed bandit tasks for model discrimination and parameter estimation in human decision-making models. There the design variable consists of blockwise Bernoulli reward probabilities, and optimized schedules such as $p(\theta)$8, $p(\theta)$9, $p(y\mid \theta,d)$0, $p(y\mid \theta,d)$1, and $p(y\mid \theta,d)$2 were found to improve model recovery and parameter estimation relative to baseline literature-style designs (Valentin et al., 2023).

In decision-critical inverse problems, BOED can be embedded directly inside downstream control. GoBOED defines a robust posterior decision problem

$p(y\mid \theta,d)$3

and then chooses the experiment by

$p(y\mid \theta,d)$4

This has been demonstrated in source localization, epidemic management, and pharmacokinetic control, where near-optimal design windows were found to be substantially wider than those predicted by goal-agnostic EIG maximization (Go et al., 25 May 2026).

In robotics and nonlinear system identification, the “experiment” can itself be a trajectory. For uncertain nonlinear dynamical systems,

$p(y\mid \theta,d)$5

BOED has been used to choose parameterized trajectories that maximize expected information about $p(y\mid \theta,d)$6 while also satisfying reachability-based safety constraints. In this formulation, the design variable is a trajectory parameter $p(y\mid \theta,d)$7, and the utility is the sum of expected information gains along measurement times on that trajectory (Ewen et al., 2023).

Other applications emphasize that BOED can target quantities of interest beyond conventional parameter vectors. For black-box uncertainty propagation, one BOED formulation optimizes the expected information gain about the scalar functional

$p(y\mid \theta,d)$8

the statistical expectation of a physical response surface under an input distribution. In that case the induced prior and posterior over $p(y\mid \theta,d)$9 are Gaussian under a GP surrogate, allowing a semi-analytic expected-KL acquisition function (Pandita et al., 2018). In constitutive model calibration, BOED has been integrated into an interlaced characterization-and-calibration loop that chooses the next biaxial load step to maximize expected information about elastoplastic material parameters, leading to lower posterior uncertainty than naive static load paths in exemplar problems (Ricciardi et al., 2023).

6. Extensions, robustness, and current directions

A central current theme is that EIG is not always the right BOED target. Goal-driven BOED argues that maximizing information gain about all parameter directions can be misaligned with robust downstream decisions. Under a task-relevant subspace assumption, the GoBOED gradient is insensitive to parameter directions irrelevant to the decision constraints, formalizing why decision-focused design can admit broader near-optimal sets than parameter-centric EIG design (Go et al., 25 May 2026).

Another direction studies BOED under model misspecification and train-test covariate shift. In that setting, generalization error for the learned predictor can be decomposed into misspecification bias, estimation bias, and a cross-term called error (de-)amplification: $\theta$0 This leads to a representativeness-adjusted acquisition

$\theta$1

intended to mitigate misspecification by favoring designs that are both informative and representative of the target distribution (Tang et al., 9 Jun 2025).

A more radical response is to replace KL altogether. “Beyond Expected Information Gain” proposes an IPM-based BOED framework in which the prior-posterior discrepancy is measured by an integral probability metric, including Wasserstein distance, Maximum Mean Discrepancy, and Energy Distance, rather than by a log-density ratio. The paper states that IPM-based utilities provide stronger geometry-aware stability under surrogate-model error and prior misspecification than classical EIG-based utilities, and extends the same sample-based BOED template in a plug-and-play manner to geometry-aware discrepancies beyond the IPM class (Wu et al., 23 Apr 2026).

Across these extensions, a consistent caveat remains: no design is optimal in isolation from the scientific target, the prior, the model class, the utility, and the allowable design space. BOED does not eliminate modeling assumptions; it makes them operational. That is why contemporary BOED research increasingly treats utility choice, estimator choice, optimizer choice, and robustness assumptions as coequal parts of the design problem rather than as secondary implementation details (Valentin et al., 2023).