Scalable Batch Bayesian Optimization Via Subspace Acquisition Functions

A new batch Bayesian optimization method that scales to large batch sizes by sampling from axis-aligned subspaces has been posted to the arXiv preprint server. The approach, proposed by Dawei Zhan, sidesteps a known degradation in efficiency that afflicts many existing batch methods as batch size grows [1][2]. Bayesian optimization is a machine learning technique rooted in statistical and mathematical optimization methods [3]. Extending it to evaluate multiple candidates in parallel can exploit modern parallel computing hardware, but most current batch approaches lose efficiency when the batch size increases [1][2]. The new work introduces a different strategy: rather than optimizing over the full problem space for every point in a batch, it draws a batch of axis-aligned subspaces and selects one candidate from each subspace using standard acquisition functions [1][2]. Numerical experiments reported in the paper indicate that the subspace method converges significantly faster than sequential Bayesian optimization and performs competitively against ten existing batch algorithms [1][2]. The implementation has been made publicly available on GitHub [2]. The manuscript was submitted to arXiv on 25 November 2024 and revised twice, with the latest version posted on 17 June 2026 [1]. arXiv, which began operating in 1991, is an open-access repository that hosts preprints across physics, mathematics, computer science, and related fields, and now receives roughly 24,000 new articles per month [8]. The paper appears under the machine learning category (cs.LG) and is accessible through standard arXiv abstract pages, which also offer community-built tools such as Bibliographic Explorer and CORE Recommender via the arXivLabs framework [1][6][7].