摘要
在现实中测试昂贵双层优化问题解通常需要耗费大量的资源,因此在优化此类问题时常用方法是依靠代理模型预先评估问题解的质量,以此来减少所需要进行的实际测试。这种基于代理模型的优化方法的核心步骤为插值,即以代理函数的评估质量为标准决定采用哪个解进行实际测试。研究人员发现对于此类问题常用的全局优化方法在昂贵问题上的表现低于预期。针对这个问题,本文提出了一种基于代理模型的强化应用的双层优化算法,实验表明它在低预算下与预测值插值优化算法比较具有竞争力,在高预算下保持有全局优化的特性,在不同的双层优化问题上能够实现稳定的优化表现。
Real world bilevel optimization problems are often expensive in nature,therefore,when searching for optimal solution,it is common to use a metamodel/surrogate to estimate the potential quality of different solutions and reduce the number of evaluations.This paradigm is often referred to as surrogate based optimization.Its key step is an infill process.A commonly used infill scheme is expected improvement maximization.It can balance exploitation and exploration,and its global convergence is mathematically sound.However,there have been increasing number of studies show that its asymptotic performance is not guaranteed in empirical applications.Its simpler alternative which only exploits the best solution from surrogate shows competitive performances.In this study,we propose a new approach to compensate the disadvantage of expected improvement scheme,so that when used in expensive bilevel optimization with limited budget,its performance holds against exploitation only approach.We verify the usefulness of the proposed algorithm on a set of benchmark test problems.Experimental results indicate that the proposed algorithm not only improve expected improvement scheme’s performance but also is competitive with the exploitation only scheme.
作者
林辉
LIN Hui(University of Chinese Academy of Sciences,Beijing,100049,China;Key Laboratory of Speech Acoustics&Content Understanding,Institute of Acoustic,Chinese Academy of Sciences,Beijing,100190,China)
出处
《网络新媒体技术》
2022年第4期16-25,共10页
Network New Media Technology
关键词
进化算法
双层优化
代理模型
全局优化
evolutionary algorithm
bi-level optimization
surrogate based optimization
global optimization
