Research On Swarm Clustering Algorithms

Cluster analysis is a typical unsupervised learning problem.Compared with su-pervised learning,cluster analysis is characterized by analyzing the similarity pattern between the research objects without sufficient prior knowledge,so as to partition a set of physical or abstract objects into several groups.The groups obtained from the clustering process are known as clusters.In order to divide the objects into the correct clusters,the algorithm needs to mine the similarities between the objects.Swarm intelligence is a commonly used optimization technique and also is one of the methods for solving clustering problems.The clustering method based on swarm intelligence often chooses a random search algorithm based on population as its frame-work,and encodes the individuals in the population as one or a set of cluster center vectors.With the iterative evolution of the population,individuals can simultaneously search different regions of the solution space,thereby finding a global optimal solution.Currently,most existing clustering algorithms based on swarm intelligence are striving to find a set of cluster centers.However,for centroid-based algorithms it is difficult to process clusters of arbitrary shapes.At the same time,the existing methods still have some deficiencies.For example,it is hard to determine the size of the population and how to initialize the population.This thesis uses the method based on swarm intelli-gence to solve the clustering problem.The thesis mainly includes the following two aspects.(1)A swarm clustering algorithm named SCA has been proposed.In SCA,each data point is encoded as an individual in the swarm.Therefore,the size of swarm is con-sistent with the size of the data set;the kernel density estimation is adopted as a fitness function in SCA to evaluate the density of particles;leader particle is selected for each particle to assist the particles to fly,thereby reducing the error of the algorithm;a dy-namic inertial weight adjustment strategy is used.The performance of SCA was tested on commonly used synthetic datasets and compared with other benchmark algorithms.The experimental results show that the SCA is a competitive algorithm.(2)A novel efficient swarm clustering algorithm named SCA2 has been proposed.Compared with SCA,SCA2 mainly made the following improvements:1)the radial basis function network is adopted as surrogate model to approximate the density distri-bution of the problem space,so the trained radial basis function network is used as the fitness function to reduce the time cost of the algorithm.Besides,a merging process is applied in SCA2,which merges multiple particles that meet the merge conditions during the flight into a single particle to fly,thereby further reducing the time cost of the algorithm;2)leader particle is expanded to K-leader list,which can give the par-ticles the opportunity to choose different leader particles,thereby reducing the initial misallocation and improving the performance of the algorithm;3)a simplified strategy is used to update the position of each particle,which can better respond to the distinct density differences and handle them separately.The experiments were carried out on commonly used synthetic data sets and real world data sets to evaluate the performance of SCA2,and compared it with other different types of clustering algorithms.The re-sults show that the SCA2 algorithm has achieved good results in most tested data sets and is a more advantageous algorithm.This thesis not only promotes the research of clustering methods based on swarm intelligence and evolutionary algorithms,but also has reference value for the in-depth research and further development of swarm intelligence models and algorithms.