Bayesian Optimal Experimental Design

machine learning
decision making under uncertainty

In research settings where experiments are expensive, time-consuming, or potentially hazardous, it’s crucial to optimize experimental design to maximize information gain. This research advances Bayesian optimal experimental design (BOED) methodology to address these challenges.

Approaches

BOED has employed several key methodologies:

  1. Gaussian Approximation Method
    This approach leverages Gaussian approximations of posterior distributions to evaluate uncertainty reduction through posterior covariance matrices.

  2. KL Divergence Method
    This method involves calculating the expected Kullback-Leibler (KL) divergence between posterior and prior distributions to quantify information gain.

While these approaches are theoretically sound, they often involve significant computational overhead, limiting their practical application.

Schematic overview of BOED process: The diagram illustrates how BOED selects optimal experiment(experiment_1) to maximize information for model parameter estimation