On Estimating the Gradient of the Expected Information Gain in Bayesian Experimental Design

Ziqiao Ao, Jinglai LiAAAI 2024 · DOI

Introduction

In experimental sciences, from drug discovery to particle physics, the cost of acquiring data is often high. Bayesian Experimental Design (BED) provides a framework for selecting experimental conditions that maximize the information gained about unknown parameters. The standard utility function for BED is the Expected Information Gain (EIG). However, calculating EIG is computationally challenging because it involves an intractable marginal likelihood. While modern machine learning has introduced various estimators, most focus on approximating the EIG value itself. This paper argues that for optimization purposes, estimating the gradient of the EIG is more efficient and robust.

Gradient-Based Estimators: UEEG-MCMC and BEEG-AP

The authors propose a "posterior expected representation" of the EIG gradient, leading to two novel estimators:

  1. UEEG-MCMC (Unbiased Estimation of EIG Gradient with MCMC): This method utilizes Markov Chain Monte Carlo (MCMC) to sample from the posterior distribution. It is designed to be unbiased and remains robust even in "high-information" scenarios where other methods typically struggle.
  2. BEEG-AP (Biased Estimation of EIG Gradient with Atomic Priors): This estimator is optimized for simulation efficiency. It reuses a fixed set of parameter samples from the prior, making it significantly faster than MCMC-based approaches, though it introduces some bias.

The "Information Trap"

A key theoretical contribution of this work is the identification of the "Information Trap." The authors provide a rigorous proof (Theorem 4) showing that for many common estimators, such as Prior Contrastive Estimation (PCE), the number of simulations required to maintain a constant level of error grows exponentially with the true EIG. This explains why existing methods often fail to converge or provide poor designs when the potential for learning is high.

Figure 6: Comparison of estimator bias across different EIG regimes. The top row shows that as the ground-truth EIG increases (from left to right), the bias of BEEG-AP and PCE (y-axis) grows significantly relative to the low bias maintained by UEEG-MCMC (x-axis).

As shown in the bias analysis, UEEG-MCMC maintains low bias across all regimes, whereas BEEG-AP and PCE exhibit increasing error as the Mean EIG rises, illustrating the practical impact of the Information Trap.

Experimental Results

The proposed methods were evaluated across several simulation models, including Bayesian Linear Regression, a pharmacokinetic (PK) model, and a complex STAT5 signaling pathway model governed by ordinary differential equations (ODEs).

In the STAT5 model, which involves optimizing 10 measurement timings, UEEG-MCMC outperformed all benchmarks. It successfully identified designs that minimized posterior entropy, indicating the most precise parameter estimates. In contrast, adaptive methods like ACE failed to converge, likely due to the strong dependencies within the high-dimensional design space.

Figure 5: Final posterior entropy achieved by different optimization methods on the STAT5 model. UEEG-MCMC achieves the lowest entropy, indicating the most informative experimental design.

Conclusion

By shifting the focus from EIG value estimation to gradient estimation, this work provides a more powerful toolkit for Bayesian Experimental Design. UEEG-MCMC offers a robust solution for high-stakes, high-information experiments, while BEEG-AP provides a fast alternative for noisier settings. These methods overcome the fundamental "information trap" that limits existing contrastive estimators, paving the way for more efficient automated discovery in complex scientific domains.

Summary updated

BibTeX
@inproceedings{ao2024estimating,
  author = {Ziqiao Ao and Jinglai Li},
  title = {{On Estimating the Gradient of the Expected Information Gain in Bayesian Experimental Design}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i18.30012},
}

Loading related content...

Figure