Open Challenges and Future Directions

Summary

The review’s §5 lays out where BED should go next: (1) Policy-based BAD is powerful but fledgling — scaling and better architectures/objectives are the frontier. (2) Linking with related areas — Bayesian active learning and Bayesian RL are intimately connected to BAD but largely siloed; cross-pollination (e.g. importing active learning’s success with high-dimensional data) is promising. (3) Model misspecification & downstream analysis — BED is only as good as its model, and is especially sensitive to misspecification; under-studied theoretically and empirically. (4) Models & applications — implicit-likelihood and richer simulators unlock new domains.

Overview

Having surveyed the computational revolution (§3) and the policy leap (§4), the review closes by identifying the most important open problems. These are the gaps a researcher entering the field should target.

Main Content

1. Policy-based BAD (§5.1)

DAD-style policies have sharply improved deployment speed and design quality, but represent a “fledgling” approach with large headroom. The biggest challenge is scaling to larger, more complex problems — sensitive to the dimensionality and smoothness of the model and design space and to the experiment length $T$. Advances are needed on network architectures, objectives, and training mechanisms; discrete design components remain hard.

2. Linking with related areas (§5.2)

BED ↔ Bayesian active learning ↔ Bayesian RL

Active learning is an important special case of BAD: adaptively train a model data-efficiently by choosing which inputs to label. The BALD score is the per-step EIG of the model parameters. Two transfer opportunities: (a) active learning has succeeded with much larger, higher-dimensional datasets (computer vision, NLP) — those techniques could scale BED; (b) conversely, BED’s EIG estimators and policy methods could help active learning beyond small-class classification. The review also notes BALD is a sub-optimal application of BED for active learning when the real goal is downstream predictions, not parameters (prediction-oriented design). Bayesian RL: BAD is a Bayes-adaptive Markov decision process with the incremental EIG as reward; information-based exploration in RL is analogous. Links exist but are “largely unexplored.”

3. Model misspecification and downstream analysis (§5.3)

BED's Achilles' heel (Rainforth 2023 §5.3)

BED is especially sensitive to misspecification because it uses the model both to fit data and to choose new data. When no ${\theta }^{\text{*}}$ makes $p(y\mid {\theta }^{\text{*}},\xi )$ match reality, BAD can suffer catastrophic failure — getting “stuck” querying uninformative designs and producing poor datasets. An extreme case: in linear regression the EIG of the coefficients is always maximized at the extremes of the inputs regardless of prior, so a misspecified linear model never explores the interior of the design space. Mitigation is under-studied (only limited theoretical and empirical work). A related issue: data gathered by BED is often re-used downstream for non-Bayesian purposes (empirical risk minimization, model selection) — the likelihood principle offers some protection for Bayesian downstream analysis (if not itself misspecified), but practical guarantees beyond inference-in-the-chosen-model are lacking.

4. Models and applications (§5.4)

BED’s performance is only as good as the underlying model, so modeling advances are the most important vector. Two general points:

• Implicit-likelihood models (The Computational Revolution in EIG Estimation) let BED use accurate simulators — often easier to build than closed-form likelihoods — spanning quantum logic gates to large-scale climate simulations.
• The computational advances make richer, more complex models feasible; practitioners should embrace flexible, accurate models that reflect their beliefs rather than restricting to easy-to-estimate ones, and should consider BED for larger and more complex problems than it has historically been used for.

Connections

• Caps the review narrative begun in Modern Bayesian Experimental Design - Overview.
• Model-dependence caveat picks up the thread from Information-Theoretic Design Objectives (BED is only as good as its model).
• Active-learning and RL links connect this topic to broader ML; misspecification connects to robust/Bayesian-model-checking literatures.

See Also

• From Designs to Policies (Deep Adaptive Design) — the policy frontier being scaled
• Information-Theoretic Design Objectives — the model-dependence caveat
• Model Checking / Model Comparison — tools relevant to detecting misspecification
• Bayesian Workflow - Overview — building and validating the models BED depends on