Answers are grounded in this paper and cite their sources.
This paper demonstrates a novel differentiable approach to optimize acquisition functions under cardinality constraints, enhancing exploration efficiency.
Bayesian Optimization (BO) is a widely used framework for efficiently optimizing expensive black-box functions. In many real-world applications, such as industrial product design, it is often necessary to select exactly k variables out of d, which imposes strict cardinality constraints. Although these constraints are of considerable practical importance, they have received limited attention in the BO literature. The resulting search space becomes combinatorially complex, making standard BO methods less effective. In this work, we formalize the problem setting of BO under cardinality constraints and propose a novel framework that enables gradient-based optimization of acquisition functions in a differentiable manner. Specifically, we combine the Gumbel-Top-k sampling strategy with the Straight-Through Estimator (STE) to enforce exact k-sparsity during optimization while maintaining end-to-end differentiability. We also investigate the limitations of the widely used Matérn kernel and empirically show that it tends to concentrate on fixed subsets under cardinality constraints. To address this issue, we introduce an alternative kernel formulation that facilitates more effective exploration in sparse subset selection problems. Experiments on synthetic datasets demonstrate that our method outperforms conventional baselines in terms of both optimization quality and constraint satisfaction. This work provides a scalable and differentiable solution to cardinality-constrained optimization in high-dimensional settings, where combinatorial complexity has traditionally posed substantial challenges.
No takes yet. Share an insight, caveat, or question.
Fukushima et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: