基于近似边界和聚类的昂贵多目标优化算法-《计算机技术与发展》

文章信息/Info

Title:: Expensive Multi-objective Optimization Algorithm Based on Approximate Boundary and Clustering

Author(s):: ZHANG Feng; 28th Research Institute of China Electronics Technology Group Corporation,Nanjing 210007,China

Keywords:: expensive many - objective optimization problem; nadir point; many - objective evolutionary algorithm; expensive multi -objective optimization algorithm; Gaussian process

摘要:: 不少工程优化问题涉及同时优化超过 2 个冲突的目标,并且目标函数的评估比较耗时,这类问题就属于昂贵的超多目标优化问题。由于目标空间过于庞大,并且只能使用较少的目标函数评估次数进行求解,这使得算法的收敛速度比较缓慢以及难以保持较好多样性。此外,许多算法往往忽略使用极值点的有效信息来加速算法收敛。为了解决上述问题,该文在一种新颖的超多目标进化算法的基础上,提出了一种基于近似边界和聚类的昂贵多目标优化算法。通过使用一组高斯过程近似目标函数来辅助算法进行评估,算法还使用极值点来加速收敛并优化出一个较好的候选种群,然后进一步提出使用一种评价指标来批量挑选出一些最有价值的候选解,借此使得算法能够保持较好的收敛性和多样性。最后通过与多个流行的求解昂贵超多目标优化问题算法进行对比实验,证明了算法的有效性。

Abstract:: Many engineering optimization problems involve optimizing more than two conflicting objectives at the same time,and the evaluation of the objective function is time - consuming. This type of problem is the expensive many - objective optimization problem.Because the objective space is too large,and only a small number of objective function evaluation times can be used for solving,thismakes the convergence speed of the algorithm slower and it is difficult to maintain good diversity. In addition,many algorithms oftenneglect to use the effective information of extreme points to accelerate the algorithm’s convergence. In order to solve the aboveproblems,we propose an expensive multi-objective optimization algorithm based on approximate boundaries and clustering on the basis ofa novel many-objective evolutionary algorithm. By using a set of Gaussian processes to approximate the objective function to assist thealgorithm for evaluation,the algorithm also uses extreme points to accelerate convergence and optimize a better candidate population.Then we further propose to use an evaluation indicator to select some of the most valuable candidate solutions in batches,so that thealgorithm can maintain better convergence and diversity. Finally,through comparative experiments with a number of popular algorithmsfor solving expensive many-objective optimization problems,the effectiveness of the algorithm is proved.