ACTION-BED: Task-Driven Bayesian Experimental Design with Singly Intractable Objectives

Traditional Bayesian Experimental Design (BED) typically aims to maximize the Expected Uncertainty Reduction (EUR) between a prior and a posterior distribution. This approach is often computationally prohibitive because it leads to "doubly intractable" objectives, requiring nested expectations and complex posterior approximations. This paper asks: can we reformulate BED to be directly task-driven, avoiding the need for explicit posterior estimation while remaining computationally efficient?

Drawing on first principles from Bayesian decision theory, the authors propose a paradigm shift: instead of measuring the value of data by how much it reduces uncertainty, they measure it by the performance of the downstream actions it enables. They define the Expected Future Loss (EFL) as the objective function. By introducing an action policy that maps observed data to downstream decisions, they transform the problem into a "singly intractable" objective. This allows for the joint optimization of both the design policy (how to collect data) and the action policy (how to use data) using standard stochastic gradient descent.

• Task-Driven Design: The ACTION-BED framework allows researchers to customize data acquisition to specific downstream tasks (e.g., point estimation or decision-making) rather than relying on generic uncertainty metrics like Shannon entropy.
• Computational Efficiency: By avoiding explicit posterior estimation or marginal likelihood calculations, the method is significantly simpler to implement and train using deep neural networks.
• Superior Performance: Empirically, the authors demonstrate that ACTION-BED often outperforms state-of-the-art methods in terms of information gain, even when the primary goal is a specific downstream loss rather than direct uncertainty reduction.