快速综合学习粒子群优化算法  

Fast comprehensive learning particle swarm optimization

作　　者：杨帆[1,2] 乌景秀 范子武[1,2] 李子祥 朱沈涛[1,2] YANG Fan;WU Jingxiu;FAN Ziwu;LI Zixiang;ZHU Shentao(Nanjing Hydraulic Research Institute,Nanjing 210029,Jiangsu,China;Key Laboratory of Water Governance of the Taihu Lake Basin,Ministry of Water Resources,Nanjing 210029,Jiangsu,China)

机构地区：[1]南京水利科学研究院,江苏南京210029 [2]水利部太湖流域水治理重点实验室,江苏南京210029

出　　处：《水利水电技术(中英文)》2025年第2期30-44,共15页Water Resources and Hydropower Engineering

基　　金：国家重点研发计划项目(2022YFC320260303);广西科技重大专项项目(桂科AA23062053—2);江苏省水利科技项目(2022049,2023008);中央级公益性科研院所基本科研业务费专项项目(Y124002)。

摘　　要：【目的】粒子群优化算法在反问题求解、函数优化、数据挖掘、机器学习等研究领域广泛应用,但在求解复杂多峰问题时仍存在过早收敛的问题。为了提升粒子群算法在处理复杂多峰问题求解速度和精度,提出了快速综合学习粒子群优化算法(Fast Comprehensive Learning Particle Swarm Optimization,FCLPSO)。【方法】FCLPSO算法引入粒子学习概率、个体影响概率、群体影响概率三个属性,表征每个粒子个体“与生俱来”的不同学习能力,同时新增强化学习、粒子重生等策略,提升算法收敛速度以及监测并跳出“伪收敛”状态。选用14个标准测试函数以及6种常用粒子群变体算法开展FCLPSO算法性能分析。【结果】结果显示:在收敛性方面,FCLPSO算法平均排名为1.86,排名第一次数为7次、排名第二的次数为2次、排名最后次数为0,最终综合排名第一;在鲁棒性方面,FCLPSO算法成功率排名第一,平均值为94.3%,14个测试函数中最低成功率为73.3%;达到阈值所需适应度评价次数最少,平均值40817,较其他算法评价次数少一半。【结论】结果表明:FCLPSO算法在收敛精度、收敛速度和鲁棒性方面排名综合第一,对复杂多峰问题求解更具优势,可为工程应用中复杂优化问题求解提供重要手段。[Objective]The particle swarm optimization algorithm is widely used in research fields such as inverse problem solving,function optimization,data mining,and machine learning,but it still faces the problem of premature convergence when solving complex multimodal problems.In order to improve the speed and accuracy of traditional particle swarm optimization in handling complex multimodal problems,this paper proposes the Fast Comprehensive Learning Particle Swarm Optimization algorithm(FCLPSO).[Methods]The FCLPSO algorithm introduces three attributes:learning probability curve,presonal probability,and group influence probability,to characterize the different learning abilities of each individual particle.At the same time,strategies such as reinforcement learning and particle rebirth are added to improve the convergence speed of the algorithm and monitor and jump out of the"pseudo convergence"state.14 standard benchmark test functions and 6 commonly used particle swarm optimization variant algorithms were selected for performance analysis of the FCLPSO algorithm.[Results]The result showed that in terms of convergence,the average ranking of the FCLPSO algorithm was 1.86,with 7 times ranking first,2 times ranking second,and 0 times ranking last and the overall ranking was first;In terms of robustness,the FCLPSO algorithm ranks first with an average success rate of 94.3%,and the lowest success rate among the 14 test functions is 73.3%;The minimum number of fitness evaluations required to reach the threshold is 40817,which is half the number of evaluations compared to other algorithms.[Conclusion]The result indicate that the FCLPSO algorithm ranks first in terms of convergence accuracy,convergence speed,and robustness,and has more advantages in solving complex multimodal problems.It can provide an important means for solving complex optimization problems in engineering applications.

关 键 词：粒子群优化算法 强化学习 粒子属性 粒子重生 过早收敛 影响因素 人工智能 全局搜索 

分 类 号：TV301.6[水利工程—水工结构工程]