Research map / Optimization / Optimizers
On Estimating the Gradient of the Expected Information Gain in Bayesian Experimental Design
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes two estimators of the gradient of expected information gain for Bayesian experimental design: UEEG-MCMC (posterior MCMC samples) and BEEG-AP (atomic priors with sample reuse), with theory linking BEEG-AP to srNMC.
Problem
Optimizing expected information gain (EIG) for Bayesian experimental design requires gradients of EIG w.r.t. design variables, which are hard to estimate accurately and efficiently when the marginal likelihood is intractable and EIG values are large.
Approach
Derives a posterior-expected representation of the EIG gradient (using the tractable likelihood score in place of the intractable marginal score), then estimates it either with MCMC posterior samples (UEEG-MCMC) or with a finite set of reused parameter samples as atomic priors (BEEG-AP), which equals the gradient of the srNMC estimator.
Key results
- On Bayesian linear regression, UEEG-MCMC shows lower gradient bias than BEEG-AP and PCE as the ground-truth EIG grows, while BEEG-AP matches PCE bias at lower simulation cost.
- In the small-EIG STAT5 scenario, BEEG-AP outperforms other methods in posterior entropy; in the large-EIG scenario UEEG-MCMC gives the best posterior entropy and BEEG-AP falls below GradBED.
- Theory shows srNMC lower-bounds EIG, is monotone in M, has MSE O(1/M) under bounded likelihood, and requires M growing exponentially in EIG to reach a fixed error.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UEEG-MCMC instead of BEEG-AP or the other compared methods | Toy algebraic model · small noise (large EIG) | posterior entropy | Helped |
| use BEEG-AP or UEEG-MCMC instead of ACE, PCE, or GradBED | Toy algebraic model · small noise (large EIG) | posterior entropy | Helped |
| use BEEG-AP or UEEG-MCMC for design optimization | Toy algebraic model · large noise | design quality | Helped |
| use BEEG-AP instead of the other compared methods | PK model · small EIG | convergence rate | Helped |
| use UEEG-MCMC instead of the other compared methods | PK model · large EIG | posterior entropy | Helped |
| use ACE with a posterior inference network | PK model · large EIG | posterior inference network learning | Hurt |
| use BEEG-AP or UEEG-MCMC instead of the other compared methods | STAT5 model · small EIG | posterior entropy | Helped |
| use UEEG-MCMC instead of BEEG-AP or the other compared methods | STAT5 model · large EIG | posterior entropy | Helped |
| use BEEG-AP instead of GradBED | STAT5 model · large EIG | posterior entropy | Hurt |
| use ACE | STAT5 model · small and large EIG | design quality | Hurt |
| use UEEG-MCMC instead of BEEG-AP or PCE | EIG gradient estimation · Bayesian linear regression; increasing ground-truth EIG | gradient estimation bias | Helped |
| use BEEG-AP instead of PCE | EIG gradient estimation · Bayesian linear regression | gradient estimation bias | No clear effect |
Surprising findings
- ACE fails to learn a posterior inference network when EIG is large due to strong observation-parameter dependence.
- BEEG-AP's performance falls below GradBED in the large-EIG STAT5 scenario.
Limitations
- UEEG-MCMC is biased due to finite-length MCMC samples from the posterior.
- BEEG-AP is biased and its srNMC-based estimate needs exponentially many samples in EIG when EIG is large.
- General-purpose conditional density networks (MDN, normalizing flows) fail for large EIG in ACE.
Experimental setup
Benchmarks: Bayesian linear regression EIG gradient bias toy algebraic model PK model STAT5 model
Metrics: estimation bias EIG (NMC estimate) posterior entropy convergence vs. number of simulations
Baselines: PCE ACE GradBED
Builds on: nested Monte Carlo srNMC (sample reuse NMC) prior contrastive estimation adaptive contrastive estimation GradBED
Closest papers by meaning
- Bayesian Experimental Design Via Contrastive Diffusions (ICLR 2025 · similarity 0.90)
- Bayesian Adaptive Calibration and Optimal Design (NeurIPS 2024 · similarity 0.88)
- PASOA- PArticle baSed Bayesian Optimal Adaptive design (ICML 2024 · similarity 0.88)
- Maximin Robust Bayesian Experimental Design (NeurIPS 2026 · similarity 0.87)
- Nesting Particle Filters for Experimental Design in Dynamical Systems (ICML 2024 · similarity 0.86)
- Informed Initialization for Bayesian Optimization and Active Learning (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.