基于改进自监督学习群体智能(ISLCI)的高性能聚类算法  

Improved self supervised learning collection intelligence based high performance data clustering approach

作　　者：曾令伟[1] 伍振兴[1] 杜文才[2] 

机构地区：[1]琼州学院电子信息工程学院,海南三亚572022 [2]海南大学信息科学技术学院,海南海口570228

出　　处：《重庆邮电大学学报（自然科学版）》2016年第1期131-137,共7页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：2014年海南省高等学校科学研究项目(HNKY2014-65)~~

摘　　要：针对现有数据聚类算法(如K-means)易陷入局部最优和聚类质量不佳的问题,提出一种结合改进自监督学习群体智能(improved self supervised learning collection intelligence,ISLCI)和K均值(K-means)的高性能聚类算法。已有的自监督学习群体智能演化方案具有计算效率和聚类质量高的优点,但当应用于数据聚类时,收敛速度较慢且极易陷入局部最优。为ISLCI加入突变操作,增加其样本多样性来降低早熟的概率,提高最优解的求解质量;计算每个样本的行为方程,获得其行为结果;通过轮盘赌方案来选择群体智能学习的对象和群体中其他样本学习目标对象的属性来提高自己。同时,利用K-means操作提高其收敛速度,提高算法计算效率。对比试验结果表明,本算法具有收敛速度快、聚类质量高、不易陷入局部最优的特点。For the problems that traditional data clustering approaches easily converge to local optima and the quality of the solution is not good,a high performance clustering approach which combines the improved self supervised learning collection intelligence and K-means is proposed. The existing self supervised learning approach has the advantage of computation efficiency and quality of clustering,but has the problem of low speed of convergence and trapping in local optimal easily.Firstly,a mutation mechanism is added to ISLCI that aims to reduce the probability of optima and the quality of optimal solution is improved; Secondly,the action function of each candidate is computed. Lastly,the object of the collection intelligence learning is selected by roulette approach,and the others in the population learn from the object to improve themselves. The converge speed is speeded up with K-means approach and the computation efficiency is improved. The compared experiment result demonstrated that the proposed approach has the characteristic of converge quickly,good quality of clustering solution and low probability to fall to local optima.

关 键 词：自监督学习群体智能 数据聚类 突变操作 簇内距离 函数评价次数 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]