arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2205.03015math.OC
收藏

受Lipschitz启发的HALRECT无导数全局优化算法

Lipschitz-inspired HALRECT Algorithm for Derivative-free Global Optimization

Vilnius University, Institute of Data Science and Digital Technologies(维尔纽斯大学,数据科学与数字技术研究所)

Linas Stripinis, Remigijus Paulavičius

AI总结:

针对Lipschitz常数未知的带箱约束全局优化问题,提出结合对分与多点采样的HALRECT算法及变体,实验证明其性能优于现有前沿DIRECT类算法,鲁棒性更强。

摘要:

本文研究Lipschitz常数未知的Lipschitz连续函数的带箱约束全局优化问题。受著名的DIRECT(分割矩形)算法启发,提出了一种新的HALRECT(对分矩形)算法。该确定性新方法将对分(二分法)与新型多点采样方案相结合,区别于现有大多数DIRECT类算法采用的三分法和中点采样方式。新的划分与采样方案可利用关于目标函数的更全面信息,同时提出四种不同的潜在最优超矩形选择策略,以有效挖掘目标函数信息。在DIRECTGOLib v1.1的96个带箱约束基准函数及其96个扰动版本上,对原始HALRECT算法及另外提出的HALRECT变体(共12种)与其他12种近年提出的DIRECT类算法进行了测试对比。大量实验结果表明,与前沿的DIRECT类全局优化方法相比,该算法性能极具前景。新的HALRECT方法在不同复杂度的问题上均具备高鲁棒性,覆盖从简单的单峰低维问题到复杂的多峰高维问题。

英文摘要:

This article considers a box-constrained global optimization problem for Lipschitz-continuous functions with an unknown Lipschitz constant. Motivated by the famous DIRECT (DIviding RECTangles), a new HALRECT (HALving RECTangles) algorithm is introduced. A new deterministic approach combines halving (bisection) with a new multi-point sampling scheme in contrast to trisection and midpoint sampling used in the most existing DIRECT-type algorithms. A new partitioning and sampling scheme utilizes more comprehensive information about the objective function. Four different strategies of selecting potentially optimal hyper-rectangles are introduced to exploit the information about the objective function effectively. The original HALRECT algorithm and other introduced HALRECT variations (twelve in total) are tested and compared with the other twelve recently introduced DIRECT-type algorithms on $96$ box-constrained benchmark functions from DIRECTGOLib v1.1, and 96 perturbed their versions. The extensive experimental results show a very promising performance compared to state-of-the-art DIRECT-type global optimization. New HALRECT approaches offers high robustness across problems of different degrees of complexity, varying from simple - uni-modal and low dimensional to complex - multi-modal and higher dimensionality.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑