A Clustering Individual Changes-driven Transfer Learning and Interpolation Strategy for Dynamic Multiobjective Optimization
提出一种新方法，通过聚类个体变化特征来设计子种群预测机制，并用最近邻插值改善帕累托前沿分布，同时引入双源域适应提升问题适应性，在基准测试和实际任务中表现优异。
Clustering-based prediction strategies have shown promising performance in dynamic multiobjective optimization. Most of them cluster a population by independent static positions of individuals and adopt a same prediction mechanism for subpopulations. However, this limits the preferred search tasks of clusters and tends to cause homogeneity of individuals. Therefore, this paper proposes a clustering individual changes-driven transfer learning and interpolation strategy method. First, a subpopulation prediction strategy based on clustering of individual features divides populations into different clusters and designs different prediction mechanisms based on the preferences of clusters, aiming to enhance the prediction accuracy of subpopulations and guide individuals to evolve in promising directions. Second, a nearest neighbor interpolation strategy inserts new objective points to sparse regions of the current Pareto optimal front according to neighboring point information, improving the geometric distribution. Third, a dual-source domain adaptation mechanism is presented to further increase the adaptability of solving problems with various types by integrating knowledge learned on two most useful environments. The comprehensive experimental results demonstrate the outstanding performance of the proposed method over six state-of-the-art competitors on the benchmark suite and good engineering applicability on a real-world task.