基于子空间对齐和反向学习的进化多任务优化算法

Evolutionary Multi-tasking Optimization Algorithm Based on Subspace Alignment and Opposition Learning

与传统的单任务进化算法不同,进化多任务优化(EMT)利用任务之间的信息共享,同时对多个任务实施进化搜索,进而提升多个任务的收敛性能,但这些任务可能拥有不同的特征。为实现任务之间信息的有效传递,提出了一种基于子空间对齐和反向学习的进化多任务优化算法(EMT-SOL)。该算法首先通过历史支配关系选择合适的迁移个体;然后,通过迁移学习中的子空间对齐学习策略,建立任务之间的映射关系,利用映射关系减小迁移个体与目标任务种群个体之间的差异;同时,利用目标任务的个体对映射后的迁移个体进行反向学习,提高目标任务种群的多样性;最后,通过9个标准测试函数对该算法进行测试,并与6个现有算法对比分析。研究结果表明,本文所提出的算法在收敛性能和个体正迁移比例方面均优于对比算法。

Evolutionary multi-task optimization(EMT)uses information sharing between tasks.Unlike traditional single-task evolutionary algorithms,EMT employs information sharing between tasks to perform evolutionary search on multiple tasks at the same time,thus improving the convergence performance of multiple tasks.However,these tasks may have different characteristics.In order to realize the effective transmission of information between tasks,an evolutionary multi-task optimization algorithm based on subspace alignment and reverse learning(EMT-SOL)is proposed.The algorithm selects the appropriate migration individuals through the historical dominance relationship.Through the subspace alignment learning strategy in transfer learning,the mapping relationship between tasks is established,and the mapping relationship is used to reduce the difference between the transfer individual and the target task population individual.Besides,the individual of the target task is used to reverse-learn the mapped migrated individuals to improve the diversity of the target task population.The algorithm is tested by 9 standard test functions and compared with 6 existing algorithms.The experimental results show that the proposed algorithm is superior to the comparison algorithm in terms of convergence performance and individual positive transfer ratio.