作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (7): 55-64. doi: 10.19678/j.issn.1000-3428.0066530

• 进化和群体智能算法与应用 • 上一篇    下一篇

自适应迁移的分解多目标多任务进化算法

蔡倩倩, 史旭华   

  1. 宁波大学 信息科学与工程学院, 浙江 宁波 315211
  • 收稿日期:2022-12-15 出版日期:2023-07-15 发布日期:2023-07-14
  • 作者简介:

    蔡倩倩(1998—),女,硕士研究生,主研方向为智能计算

    史旭华,教授、博士、博士生导师

  • 基金资助:
    国家自然科学基金(61773225)

Decomposition Multi-Objective Multi-Task Evolutionary Algorithm with Adaptive Transfer

Qianqian CAI, Xuhua SHI   

  1. Faculty of Engineering and Computer Science, Ningbo University, Ningbo 315211, Zhejiang, China
  • Received:2022-12-15 Online:2023-07-15 Published:2023-07-14

摘要:

多目标多任务进化优化是多目标优化的一个重要研究方向,通过跨任务共享有益信息以同时解决多个相关任务的优化问题。然而,现有多目标多任务进化优化研究存在相似任务匹配准确度低、缺少对知识迁移的动态控制等问题。为提高多目标多任务进化优化算法的优化效果,引入相似性动态指标和迁移概率动态调整机制,提出自适应迁移的分解多目标多任务进化算法。为了给目标任务子问题匹配关联度最高的迁移源,同时考虑种群的当前分布以及种群的进化方向2个指标,设计一种基于种群静态和动态特征相结合的迁移源匹配策略。为了合理地控制任务间的信息传递,提出基于种群进化状态的知识迁移概率自适应调整策略,在优化过程中根据优化任务的进化状态自适应地调整任务间的知识迁移概率,以满足优化任务在不同进化阶段对外部知识的需求。实验结果表明,相比MOEA/D、MO-MFEA、MO-MFEA-Ⅱ等算法,该算法具有较优的稳定性和收敛性,在常用的9组(18个独立任务)多目标多任务测试问题中有15个表现较优,优化率为83%。

关键词: 多目标多任务优化, 进化算法, 迁移优化, 分解策略, 自适应策略

Abstract:

Multi-objective multi-task evolutionary optimization is an important research direction for solving multiple related tasks simultaneously by sharing beneficial information across tasks. However, existing multi-objective multi-task evolutionary optimization studies have problems, such as low accuracy in matching similar tasks and a lack of dynamic control over knowledge transfer. To address these issues, a dynamic index of similarity and dynamic adjustment mechanism of transfer probability are introduced to propose an decomposition multi-objective multi-task evolutionary algorithm with adaptive transfer. The transfer source matching strategy is designed based on a combination of static and dynamic characteristics of the population, to match the transfer source with the highest degree of correlation to the target task subproblem, considering the current distribution and the evolution direction of the population. To reasonably control the information transmission between tasks, an adaptive knowledge transfer probability adjustment strategy based on the population evolution state of the optimization task is proposed, thereby satisfying the different needs of external knowledge in different evolution stages of the optimization task. The experimental results show that compared to MOEA/D, MO-MFEA, and MO-MFEA-Ⅱ, the proposed algorithm displays better stability and convergence. Among the commonly used nine groups(eighteen independent tasks) of multi-objective and multi-task test problems, fifteen performed better, with an optimization rate of 83%.

Key words: multi-objective multi-task optimization, evolutionary algorithm, transfer optimization, decomposition strategy, adaptive strategy