arXiv：cs.LG· L. A. Zhukov, E. V. Shaburova, D. V. Antonets·· 3 小时前
tidyHEBO：提升异方差进化贝叶斯优化的鲁棒性
tidyHEBO: Improving the Robustness of Heteroskedastic Evolutionary Bayesian Optimization
AI 导读
tidyHEBO 是一种联合学习输出变换与高斯过程参数的贝叶斯优化模型，在原始目标尺度上评估采集函数，并在进化搜索中保留有潜力的新候选。它在 Olympus 基准上综合排名第一，典型优化质量、最差尾部表现和运行间波动三项平均排名均最佳，在 Needle-in-a-Haystack 问题上也取得与专用优化器相当的结果。
正文
Abstract:Bayesian optimization is widely used for expensive black-box optimization problems. While methodological advances often focus on individual components, practical performance depends on how these components work together. We present tidyHEBO, a Bayesian optimization model. tidyHEBO jointly learns output warping and Gaussian-process parameters, evaluates acquisition functions on the original objective scale, and preserves promising novel candidates throughout evolutionary search. On the Olympus benchmark, tidyHEBO ranked first among the evaluated methods, achieving the best average ranks for typical optimization quality, worst-tail outcomes, and run-to-run variability. It also obtained results comparable to specialized optimizers on Needle-in-a-Haystack problems. On the Bayesmark benchmark, tidyHEBO also demonstrated strong results in hyperparameter tuning. Across these benchmarks, tidyHEBO combines strong optimization results with robustness across diverse black-box problems.
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | arXiv:2607.10669 [cs.LG] |
| (or arXiv:2607.10669v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2607.10669 arXiv-issued DOI via DataCite |
Submission history
From: Lev Zhukov [view email]
[v1]
Sun, 12 Jul 2026 09:27:57 UTC (1,698 KB)
[v2]
Wed, 23 Sep 2026 17:55:52 UTC (1,152 KB)
[v3]
Wed, 14 Oct 2026 06:17:33 UTC (1,239 KB)
来源：arXiv：cs.LG · arxiv.org