Search Process Analysis of Multiobjective Evolutionary Algorithms using Convergence-Diversity Diagram
Search Process Analysis of Multiobjective Evolutionary Algorithms using Convergence-Diversity Diagram
复制标题
DOI：
10.1109/scisisis55246.2022.10001961
复制
发表时间：
2022-11
期刊：
影响因子：
--
通讯作者：
Takato Kinoshita;Naoki Masuyama;Yusuke Nojima
中科院分区：
文献类型：
--
作者：
Takato Kinoshita;Naoki Masuyama;Yusuke Nojima
现实世界中的许多任务都是多目标优化问题（MOP）。基于群体的方法有望解决 MOP。特别是，多目标进化算法（MOEA）很受欢迎，并且在过去二十年中得到了积极的研究。然而，由于直接显示和比较多维解集并不容易，因此很难使用散点图等直接可视化技术来分析 MOEA 的搜索过程。本文通过扩展作者之前的工作，即收敛-多样性图，提出了一种在收敛性和多样性方面比较多个搜索过程的分析方法。通过计算实验，所提出的方法揭示了三个代表性 MOEA 和六个测试问题的特征和相似性。此外，本文还讨论了算法设计、DTLZ 测试套件中的偏差以及基于实验结果的可视化改进。
Many tasks in the real world are multi-objective optimization problems (MOPs). Population-based approaches are promising for solving MOPs. In particular, multi-objective evolutionary algorithms (MOEAs) are popular and have been actively studied over the last two decades. However, since it is not easy to directly display and compare multi-dimensional solution sets, it is difficult to analyze the search process of MOEAs using direct visualization techniques such as scatter plots. This paper proposes an analytical method to compare multiple search processes in terms of convergence and diversity by extending the authors’ previous work, i.e., Convergence-Diversity Diagram. Through computational experiments, the proposed method reveals characteristics and similarities in three representative MOEAs and six test problems. In addition, this paper provides discussions on algorithm design, biases in the DTLZ test suite, and the improvement of visualization based on experimental results.