Reproduce “Goal-Oriented Bayesian Optimal Experimental Design for Nonlinear Models using Markov Chain Monte Carlo” on OpenResearch

About this paper

Shijie Zhong, Wanggang Shen, Tommie Catanach, Xun Huan

Researchers from Johns Hopkins University, the University of Michigan, and Sandia National Laboratories developed a computational framework for Goal-Oriented Bayesian Optimal Experimental Design (GO-OED) tailored for nonlinear models. This framework uses nested Monte Carlo, MCMC, KDE, and Bayesian Optimization to quantify expected information gain on predictive quantities of interest (QoIs), demonstrating that GO-OED yields optimal designs distinct from and superior to parameter-centric approaches in various test problems and a convection-diffusion sensor placement application.

Key insights

• Optimal experimental designs derived from Goal-Oriented OED (GO-OED) for predictive quantities consistently differ from and prove more effective than those from parameter-centric OED, especially when parameter-to-QoI mappings are non-bijective or nonlinear.
• The integrated framework, combining NMC, MCMC, KDE, and BO, establishes a computationally viable approach for performing Bayesian GO-OED in complex nonlinear systems, bridging a long-standing gap between theory and practical application.
• Computational cost analysis revealed that MCMC computations dominate for higher-dimensional parameter spaces, while KDE bandwidth selection is critical for accuracy and efficiency, indicating specific areas for future methodological improvements.

Summary by alphaXiv.