Modern Bayesian Experimental Design - Overview

Summary

Rainforth, Foster, Ivanova & Bickford Smith (2023), Modern Bayesian Experimental Design (Statistical Science). A review arguing that recent ML advances have transformed BED from a computationally crippled niche into a practically deployable framework. It organizes the field around: the EIG objective and why it beats classical Fisher-information criteria; a “computational revolution” in EIG estimation (nested estimation, debiasing/MLMC, variational bounds, implicit models); stochastic-gradient design optimization; the leap from designs to policies (deep adaptive design); and open challenges (misspecification, active-learning links, scaling).

Overview

BED chooses experiment designs $\xi$ to maximize the information gathered about a target $\theta$. Despite elegant information-theoretic foundations dating to Lindley (1956), its uptake was long held back by crippling computational bottlenecks — especially in the adaptive setting where decisions must be made in real time. This review (from the group behind Foster 2019 and Foster 2020) is a guide to the developments that broke those bottlenecks and to where the field is heading.

Main Content

Structure of the review and note map

Review section	Content	Note
§2 Information-theoretic design	EIG (Eqs. 1–3), BAD (§2.2), why Bayesian vs frequentist/FIM (§2.3)	Information-Theoretic Design Objectives + Expected Information Gain + Sequential and Adaptive BED
§3 A computational revolution	nested estimation (§3.1), debiasing/MLMC (§3.2), variational & implicit (§3.3)	The Computational Revolution in EIG Estimation
§3.4 Optimization	stochastic-gradient schemes, the unified gradient approach	Optimization and Gradient Schemes for BED
§4 From designs to policies	deep adaptive design (DAD), policy networks, total EIG	From Designs to Policies (Deep Adaptive Design)
§5 Future directions	policy-based BAD, active-learning links, misspecification, models/applications	Open Challenges and Future Directions

The three big messages

1. The EIG is the right objective, and modern estimators finally make it cheap to estimate and optimize — see The Computational Revolution in EIG Estimation and Optimization and Gradient Schemes for BED.
2. Policies, not just designs. The biggest recent leap is deep adaptive design (DAD): learn a design policy ${\pi }_{\varphi }(h_{t-1})\mapsto {\xi }_t$ upfront, so adaptive experiments need no inference or optimization during deployment, and can be non-myopic — see From Designs to Policies (Deep Adaptive Design).
3. BED is now viable for large, complex, implicit-likelihood models — practitioners should embrace flexible accurate models rather than restrict to easy-to-estimate ones.

Where this review sits relative to the other two ingested papers

The review synthesizes and generalizes both Foster papers: Foster 2019’s variational estimators appear as the §3.3.1 variational bounds; Foster 2020’s unified SGA appears as the §3.4 stochastic-gradient scheme (their Eq. 15). It then goes beyond both into policy-based methods (DAD, iDAD, RL approaches) that postdate them.

Connections

• Generalizes Foster 2019 and Foster 2020 into a unified narrative of the field.
• Frames BED’s relationship to Bayesian active learning, Bayesian RL, and classical (alphabetic-optimality / FIM) design — see Information-Theoretic Design Objectives and Open Challenges and Future Directions.
• Forward pointer: policy-based adaptive design (From Designs to Policies (Deep Adaptive Design)) is the review’s signature “what’s new since 2019” contribution.

See Also

• Information-Theoretic Design Objectives — EIG vs FIM, why Bayesian
• The Computational Revolution in EIG Estimation — nested → debiased → variational → implicit
• Optimization and Gradient Schemes for BED — stochastic-gradient design
• From Designs to Policies (Deep Adaptive Design) — amortized policies
• Open Challenges and Future Directions — misspecification, active learning, scaling