Clustering-based genetic offspring generation using DBSCAN with correlation distance Clustering-based genetic offspring

• Published in: Journal of membrane computing Vol. 7; no. 2; pp. 184 - 202
• Main Authors: Li, Lianghao, Mi, Haibo, Li, Zhen, Fu, Jun
• Format: Journal Article
• Language: English
• Published: Singapore Springer Nature Singapore 01.06.2025
• Subjects: Artificial Intelligence; Computation by Abstract Devices; Computational Biology/Bioinformatics; Computer Science; Optimization; Research Paper; Theory of Computation; Intricate pareto set; Multiobjective optimization; Offspring generation; Mating restriction; Clustering algorithm
• ISSN: 2523-8906; 2523-8914
• DOI: 10.1007/s41965-024-00174-9

Differ from traditional benchmarks, real-world multiobjective optimization problems (MOPs) often have complex variable interactions, leading to intricate Pareto sets (PSs) with rotated or nonlinear forms. Simulated binary crossover (SBX), a common genetic operator for MOPs, would suffer performance declines when dealing with such complex PSs. To address this, a clustering-based mating restriction strategy (CRSBX) was designed for MOPs with complex PSs. It works by dividing the parent population into clusters that are approximately linearly distributed and then applying rotation-based SBX (RSBX) in each cluster to handle linear but rotated PSs by incorporating rotational properties into SBX. The kernel of this method lies in the utilization of clustering techniques to partition the current population into linear segments, thereby facilitating the fitting of the linear local distribution of the PS with a complex shape. However, CRSBX is subject to a limitation that the spectral clustering it employs necessitates a manually predefined parameter for determining the number of clusters. The impact of cluster numbers on algorithm performance is significant, yet it is challenging to determine in real-world issues. To overcome this challenge, we have made three improvements to CRSBX. Firstly, we employed DBSCAN as the clustering method, allowing the algorithm to operate without the need to preset the number of clusters. Secondly, we integrate the decision space with the objective space for clustering, rather than clustering solely within the decision space. Lastly, we utilized correlation distance as the criterion for distance measurement in clustering. Through these modifications, we achieve a more rational clustering of the population without the prerequisite of predefining the number of clusters, and concurrently, diminish the sensitivity to parameter settings. According to the experimental results on three test suites, the proposed algorithm is superior to existing representative evolutionary algorithms for MOPs with complex PSs.