基于协方差矩阵调整的多目标多任务优化算法

Multi-Objective Multi-Tasking Optimization Algorithm Based on Adjustment of Covariance Matrix

多任务进化(EMT)是进化计算领域的一个新兴研究方向,区别于传统的单任务搜索算法,EMT通过在任务间传递有用知识,对多个任务同时实施进化搜索,以提升多个任务的收敛性能。目前,大多数进化算法只考虑了知识迁移而忽略了任务间的联系。提出一种多目标多任务优化算法,结合迁移学习的思想,采用任务间种群的协方差矩阵差异表示任务间种群分布特征差异,使用任务间种群均值的距离表示任务间种群的分布距离,并通过任务间种群的分布特征差异和分布距离表示任务间的相似度。对于某个目标任务,将其最相似任务中的解集实施K最近邻分类,以筛选出对目标任务有价值的解,并使其迁移到目标任务中。实验结果表明,与EMTSD、MaTEA、MO-MFEA-II等多目标多任务优化算法相比,所提算法具有较佳的收敛性能,平均运行效率约提高了66.62%。

Evolutionary Multi-Tasking(EMT)is an emerging research direction in the field of evolutionary computations.Differing from a traditional single-task search algorithm,EMT implements an evolutionary search for multiple tasks simultaneously by transferring useful knowledge between tasks,thereby improving the convergence of such tasks.At present,most evolutionary algorithms focus on knowledge transferring and ignore the connections between tasks.Therefore,by combining the ideas of transfer learning,a Multi-Objective Multi-Tasking Optimization(MTO)algorithm is proposed.First,the difference in the covariance matrix between tasks represents the difference in population distribution characteristics between tasks,and the distance between the means of different tasks is used to represent the distribution distance of the population between tasks.Thus,the difference in population distribution characteristics and the population distribution distance between tasks are used to express the similarity between tasks.For the target task,K-Nearest Neighbor(KNN)classification of the solutions to the most similar task is conducted,so as to screen out valuble solutions to target task and then transfer them to the target task.The experimental results show that,compared with EMTSD,MaTEA,MO-MFEA-II,and other MTO algorithms,the proposed algorithm achieves a better convergence,and the average running efficiency is increased by approximately 66.62%.