基于多偏好物理规划的代理辅助多目标优化方法  

Surrogate-assisted multi-objective optimization method based on multi-preference physical programming

作　　者：许焕卫 杨学睿 何晗瑾 魏文张 XU Huanwei;YANG Xuerui;HE Hanjin;WEI Wenzhang(School of Mechanical and Electrical Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China)

机构地区：[1]电子科技大学机械与电气工程学院,四川成都611731

出　　处：《计算机集成制造系统》2023年第8期2574-2584,共11页Computer Integrated Manufacturing Systems

基　　金：国家自然基金面上资助项目(51975106)。

摘　　要：针对目前普遍存在的目标数目较多,Pareto前沿离散程度较高的昂贵多目标优化问题,现有大部分算法无法利用较少函数评估得到优质Pareto前沿,因此提出一种基于多偏好物理规划的代理辅助多目标优化算法(M3pEGO)。该方法首先设置偏好矩阵,通过不同物理规划总偏好值将多目标优化问题转化为单目标优化问题,接着与Kriging代理模型相结合,利用高效的全局优化(EGO)算法实现自适应优化。最后通过9个经典测试函数,将此方法与ParEGO算法和多目标EGO算法进行对比。结果表明,所提算法在解决昂贵多目标优化,尤其是Pareto前沿离散程度较高的问题上优势明显,在有限次迭代后能够精确拟合到真实Pareto前沿,且能够得到收敛、均匀的非支配解集。Aiming at the expensive multi-objective optimization problem with a large number of objectives and a high degree of dispersion of the Pareto fronts,most of the existing algorithms cannot use less function evaluation to obtain a high-quality Pareto front.Therefore,a surrogate-assisted multi-objective optimization algorithm based on multi-preference physical programming named Multi-preference Physical Programming Efficient Global Optimization(M3pEGO)was proposed.In this method,first sets up a preference matrix was built,and the multi-objective optimization problems was transformed into a single-objective optimization problem.Then the algorithm was combined with the Kriging surrogate model using Efficient Global Optimization(EGO)algorithm to achieve adaptive optimization.This method was compared with the ParEGO algorithm and multi-objective EGO algorithm through nine classical test functions.The results showed that the proposed algorithm had obvious advantages in solving expensive multi-objective optimization,especially problems with high dispersion of Pareto fronts.It could accurately fit to the true Pareto frontier after a finite number of iterations,and could obtain convergent and uniform non-dominated solution sets.

关 键 词：昂贵多目标优化 Kriging代理模型 多偏好物理规划 EGO算法 自适应优化 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]