Index: Modern BED Review

Modern BED Review

Routing Summary

Rainforth, Foster, Ivanova & Bickford Smith (2023), Modern Bayesian Experimental Design (Statistical Science). A review of how ML advances transformed BED into a deployable framework. Contains 5 notes + overview.

• New here / structure of the review / the three big messages? → Modern Bayesian Experimental Design - Overview
• EIG vs Fisher information, alphabetic optimality, why Bayesian? → Information-Theoretic Design Objectives
• Nested estimation, MLMC debiasing, variational & implicit bounds? → The Computational Revolution in EIG Estimation
• Black-box optimizers vs stochastic-gradient design? → Optimization and Gradient Schemes for BED
• Deep adaptive design (DAD), policy networks, total EIG? → From Designs to Policies (Deep Adaptive Design)
• Misspecification, active-learning/RL links, scaling, applications? → Open Challenges and Future Directions

Concept Map

Concept	Note	Type	Depends On	Key Result
Review structure; three messages; relation to Foster papers	Modern Bayesian Experimental Design - Overview	overview	Sequential and Adaptive BED	ML advances make EIG cheap to estimate & optimize; policies are the leap
EIG vs FIM; alphabetic optimality; why Bayesian	Information-Theoretic Design Objectives	concept	Expected Information Gain	FIM is a matrix that depends on unknown $\theta$; Bayesian avoids both flaws
Nested estimation; MLMC debiasing; variational; implicit	The Computational Revolution in EIG Estimation	concept	Nested Estimation and Nested Monte Carlo	MLMC → unbiased $\mathcal{O}(C^{-1/2})$; normalized $q$ → variational bound
Black-box vs gradient optimization; unified SGA; bias propagation	Optimization and Gradient Schemes for BED	concept	Unified SGD BOED - Overview	SGA on a lower bound (Eq. 15) replaces BO/grid; debiasing makes it consistent
DAD; policy network; total EIG; non-myopic; iDAD/RL	From Designs to Policies (Deep Adaptive Design)	concept	Sequential and Adaptive BED	${\xi }_t={\pi }_{\varphi }(h_{t-1})$ trained offline on TEIG; real-time, non-myopic
Policy scaling; active learning/RL; misspecification; applications	Open Challenges and Future Directions	concept	From Designs to Policies (Deep Adaptive Design)	BED especially sensitive to misspecification; embrace richer models

Notes

• Modern Bayesian Experimental Design - Overview — CONTAINS: review section→note map; three big messages; how the review synthesizes/generalizes Foster 2019 & 2020 and goes beyond to policies.
• Information-Theoretic Design Objectives — CONTAINS: EIG as one of a family of posterior-utility objectives; FIM definition + two flaws (matrix; depends on $\theta$); alphabetic A/D/E-optimality; why-Bayesian arguments; model-dependence caveat.
• The Computational Revolution in EIG Estimation — CONTAINS: MLMC debiasing (Goda 2022, Eqs. 9–11, antithetic coupling, $\mathcal{O}(C^{-1/2})$); variational upper/lower bounds (Eqs. 12–14); implicit-model estimation; debiasing-vs-variational trade-off table.
• Optimization and Gradient Schemes for BED — CONTAINS: estimator-wrapping optimizers (BO, evolutionary, coordinate exchange); unified SGA lower bound (Eq. 15); implicit-model & MLMC gradient routes; Kleinegesse–Gutmann parallel; bias-propagation and discrete-design difficulties.
• From Designs to Policies (Deep Adaptive Design) — CONTAINS: design-policy definition; total EIG (Eqs. 16–17); offline-train/live-deploy (Fig. 2); non-myopia; lineage (Huan–Marzouk → DAD → iDAD → RL); empirical quality gains.
• Open Challenges and Future Directions — CONTAINS: policy-based BAD scaling; active-learning (BALD) & Bayesian-RL links; misspecification catastrophic-failure (linear-regression extremes example); likelihood-principle protection; implicit-simulator & richer-model directions.

Sources

• Rainforth et al 2023 - Modern Bayesian Experimental Design.pdf — Rainforth, T., Foster, A., Ivanova, D.R., Bickford Smith, F. (2023), Modern Bayesian Experimental Design, Statistical Science (accepted). arXiv:2302.14545.

See Also

• Foundations — the EIG, nested estimation, and adaptive-design core
• Variational EIG Estimators — the §3.3.1 variational bounds in detail
• Gradient-Based Unified BOED — the §3.4 unified gradient approach in detail