基于组群优化的聚类算法研究

发布时间：2018-12-25 19:55

【摘要】：聚类算法是非监督学习算法领域中的一个经典算法,在对数据集的分布情况没有先验知识了解的前提下,利用相似性度量将数据集进行分类。通过聚类,发现数据样本间的关联关系。然而,在算法性能方面,传统聚类算法的表现相对较差。因此,利用群智能优化的不同参数和启发式函数规则来提高聚类效果和鲁棒性越来越受到研究者的关注。近年新兴的组群搜索优化算法,具有参数少,操作简单,不易陷入局部最优等优势,同时也面临收敛速度较慢等问题。本文提出改进组群搜索优化算法,将其思想应用于求解聚类算法,提出四种基于组群搜索优化思想的聚类算法,将传统聚类问题作为优化问题来解决。主要创新点如下:第一,提出基于发现者适应度范围排序策略的DRGSO算法。引入排序策略,避免陷入局部最优,克服了传统组群搜索优化算法(Group Search Optimizer Algorithm,GSO)中组群算子搜索方向单一,资源信息利用不足的问题,丰富寻优启发式搜索的资源信息,提高发现者种群的全局搜索能力。在11组国际标准测试函数上的仿真实验结果表明:DRGSO算法性能明显提高,收敛速度较快。第二,提出基于差分演化策略的组群搜索优化算法(Differential Ranking-based Group Search Optimizer,DRGSO)。引入四种差分变异算子模式来设计组群算子,加快算法收敛速度,克服了 GSO算法存在的收敛速度较慢,精度不高等问题。与GSO相比,改进的组群搜索操作使发现者搜索结构更为优化,能确保算法搜索路径多样性,提高算法收敛精度。第三,提出四种基于组群优化的聚类分析方法,具体包括:GSO聚类算法、基于均值GSO聚类算法、DRGSO聚类算法和基于均值DRGSO聚类算法。首先,利用GSO优化算法的种群结构来编码簇中心位置,优化聚类算法的簇分配过程,在GSO迭代过程中完成簇移动,简化聚类分析方法的复杂度,提高聚类效果。其次,提出根据发现者种群适应度范围进行局部均值策略,避免陷入局部最优,完善个体间信息资源的共享模式。最后,利用差分变异算子模式,让簇移动更具多样性,提高聚类算法的全局搜索能力。在国际标准数据集Iris、Wine上的实验结果表明,组群搜索优化聚类算法的聚类效果更明显、具有较好稳定性和鲁棒性。
[Abstract]:Clustering algorithm is a classical algorithm in the field of unsupervised learning algorithm. In the absence of prior knowledge of the distribution of data sets, the data sets are classified by similarity measurement. By clustering, the correlation between data samples is found. However, the performance of traditional clustering algorithm is relatively poor. Therefore, using different parameters and heuristic function rules of swarm intelligence optimization to improve clustering effect and robustness has attracted more and more attention. In recent years, the new cluster search optimization algorithm has the advantages of few parameters, simple operation, not easy to fall into local optimum, and also faces some problems such as slow convergence speed and so on. In this paper, an improved cluster search optimization algorithm is proposed, and its idea is applied to the clustering algorithm. Four clustering algorithms based on cluster search optimization are proposed. The traditional clustering problem is solved as an optimization problem. The main innovations are as follows: first, the DRGSO algorithm based on the range ordering strategy of discoverer fitness is proposed. The sorting strategy is introduced to avoid falling into local optimum, which overcomes the problem of single search direction of group operator and insufficient utilization of resource information in traditional group search optimization algorithm (Group Search Optimizer Algorithm,GSO), and enriches the resource information of heuristic search. Improve the global search ability of the discoverer population. The simulation results on 11 international standard test functions show that the performance of DRGSO algorithm is obviously improved and the convergence rate is faster. Secondly, a group search optimization algorithm (Differential Ranking-based Group Search Optimizer,DRGSO) based on differential evolution strategy is proposed. Four kinds of differential mutation operator patterns are introduced to design group operators to accelerate the convergence speed of the algorithm and overcome the problems of slow convergence rate and low precision of GSO algorithm. Compared with GSO, the improved group search operation can optimize the search structure of the discoverer, ensure the diversity of the search path and improve the convergence accuracy of the algorithm. Thirdly, four clustering methods based on cluster optimization are proposed, including: GSO clustering algorithm, GSO clustering algorithm based on mean value, DRGSO clustering algorithm and DRGSO clustering algorithm based on mean value. Firstly, the cluster center position is encoded by the population structure of GSO optimization algorithm, and the cluster assignment process is optimized. The cluster movement is completed in the GSO iteration process, which simplifies the complexity of the clustering analysis method and improves the clustering effect. Secondly, according to the range of population fitness of discoverers, a local mean strategy is proposed to avoid falling into local optimum and perfect the sharing mode of information resources among individuals. Finally, the differential mutation operator pattern is used to make the cluster move more diverse and improve the global searching ability of the clustering algorithm. The experimental results on the international standard data set Iris,Wine show that the clustering effect of cluster search optimization clustering algorithm is more obvious, and the clustering algorithm has better stability and robustness.
【学位授予单位】：天津科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18;TP311.13

【参考文献】