Scholarlink

<p>Bayesian optimization is a widely used algorithm for solving expensive black-box optimization problems. However, its performance significantly degrades in high-dimensional problems, since optimizing the acquisition function becomes increasingly difficult as dimensionality grows. In this work, we propose the Adaptive Dropout (AdaDropout) approach to address this challenge. The core idea is to adaptively reduce the dimensionality of the acquisition function along the iterations. In the event that a newly sampled point does not improve on the current best, AdaDropout reduces the number of selected dimensions by one. Conversely, if an improvement is observed, the number of currently selected dimensions is maintained. By gradually reducing the dimensionality of the acquisition function, the search transitions from global exploration to local exploitation. This makes the acquisition function increasingly easier to optimize. Numerical experiments demonstrate that AdaDropout effectively tackles high-dimensional challenges and achieves superior performance compared with the standard Bayesian optimization baseline and seven state-of-the-art high-dimensional Bayesian optimization algorithms. This work provides a simple yet efficient solution for high-dimensional expensive optimization.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An adaptive dropout approach for high-dimensional bayesian optimization

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Jundi Huang,
  • Dawei Zhan

摘要

Bayesian optimization is a widely used algorithm for solving expensive black-box optimization problems. However, its performance significantly degrades in high-dimensional problems, since optimizing the acquisition function becomes increasingly difficult as dimensionality grows. In this work, we propose the Adaptive Dropout (AdaDropout) approach to address this challenge. The core idea is to adaptively reduce the dimensionality of the acquisition function along the iterations. In the event that a newly sampled point does not improve on the current best, AdaDropout reduces the number of selected dimensions by one. Conversely, if an improvement is observed, the number of currently selected dimensions is maintained. By gradually reducing the dimensionality of the acquisition function, the search transitions from global exploration to local exploitation. This makes the acquisition function increasingly easier to optimize. Numerical experiments demonstrate that AdaDropout effectively tackles high-dimensional challenges and achieves superior performance compared with the standard Bayesian optimization baseline and seven state-of-the-art high-dimensional Bayesian optimization algorithms. This work provides a simple yet efficient solution for high-dimensional expensive optimization.