Supercharging Simulation-Based Inference for Bayesian Optimal Experimental Design

Bayesian optimal experimental design (BOED) seeks to maximize the expected information gain (EIG) of experiments. This requires a likelihood estimate, which in many settings is intractable. Simulation-based inference (SBI) provides powerful tools for this regime. However, existing work explicitly connecting SBI and BOED is restricted to a single contrastive EIG bound. We show that the EIG admits multiple formulations which can directly leverage modern SBI density estimators, encompassing neural posterior, likelihood, and ratio estimation. Building on this perspective, we define a novel EIG estimator using neural likelihood estimation. Further, we identify optimization as a key bottleneck of gradient based EIG maximization and show that a simple multi-start parallel gradient ascent procedure can substantially improve reliability and performance. With these innovations, our SBI-based BOED methods are able to match or outperform by up to $22%$ existing state-of-the-art approaches across standard BOED benchmarks.

💡 Research Summary

This paper tackles a central challenge in Bayesian optimal experimental design (BOED): the computation of the expected information gain (EIG) when the likelihood is intractable and only a simulator is available. While simulation‑based inference (SBI) has become the de‑facto tool for posterior inference in such “likelihood‑free” settings, prior attempts to combine SBI with BOED have relied on a single contrastive bound on the EIG, limiting both theoretical insight and practical performance.

The authors first observe that the EIG admits two equivalent formulations – one based on the reduction of posterior entropy (Eq. 1) and one based on the log‑likelihood ratio (Eq. 2). Each formulation naturally aligns with a different class of SBI density estimators: neural posterior estimation (NPE), neural likelihood estimation (NLE), and neural ratio estimation (NRE). By explicitly mapping each SBI method to a variational bound on the EIG, they create a unified framework that allows practitioners to pick the most suitable estimator for their problem.

A key technical contribution is a novel direct EIG estimator built on NLE. Instead of using a contrastive estimator that requires many prior samples, they train two conditional normalizing‑flow models: one for the conditional likelihood qϕ(y|θ,ξ) and one for the marginal likelihood qϕ(y|ξ). The EIG is then approximated as the expectation of log