Goal-driven Bayesian Optimal Experimental Design for Robust Decision-Making Under Model Uncertainty

ICLR 2026 Conference Submission19243 Authors

Keywords: Bayesan optimal experimental design; Convex optimization; Robust decision making

TL;DR: GoBOED designs experiments that reduce the uncertainty that matters for decisions, linking BOED with robust optimal control via convex optimization and a differentiable decision layer.

Abstract: Bayesian optimal experimental design (BOED) aims to predict experiments that can optimally reduce the uncertainty in the model parameters. However, in many decision-critical applications, accurate parameter estimation does not necessarily translate to better decision-making, as not all parameters may significantly affect the efficacy of the decisions made in the presence of uncertainty. In this work, we propose GoBOED (Goal-driven Bayesian Optimal Experimental Design) to directly optimize the experimental design to reduce the model uncertainty that critically affects the quality of the downstream decision-making task of interest. We establish a computationally tractable connection between BOED and robust optimal control based on an uncertain model through convex optimization. This new integrated framework for robust control under uncertainty enables efficient gradient computation through a decision layer in GoBOED. Leveraging amortized variation inference, we create a differentiable pipeline that can identify optimal experiments targeting decision value. Unlike traditional information-maximizing designs, GoBOED can provide flexibility in experimental selection, as the experiment with the lowest data acquisition cost may be prioritized when multiple experiments lead to equivalent decision quality despite their difference in reducing the parameter uncertainty. The application of GoBOED to real-world problems, such as epidemic management and pharmacokinetic control, demonstrates the efficacy of our proposed goal-driven experimental design approach.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 19243