Research Of Swarm Intelligence Algorithm For Clustering Problems

Swarm intelligence algorism is an evolutionary computation technique,has extremely special relationship with artificial life,in particular with evolutionary strategy and genetic algorithm,and has become an attentive focus of numerous researchers at present.A single natural organism is not intelligent normally,however, the whole biotic community can represent the ability of handling complex problems, and swarm intelligence is the application of the collective behavior in artificial intelligence problems.The researchers make a series of new solution schemes to traditional problems through simulating social insects and the researches are swarm intelligence researches. The swarm intelligence research field mainly has two algorisms at present:ant colony algorithm and particle swarm optimization algorithm.The ant colony algorithm is the simulation of the process that ant population collects food and has been successfully applied in various discrete optimization problems.The particle swarm optimization algorithm was originated from the simulation of the biotic population,the particle swarm optimization algorithm simulated the process of bird population feeding initially,individuals can learn from experience mutually through information sharing to promote the evolution and development of the whole community,and the particle swarm optimization algorithm was developed into a good optimization tool then and has been applied in the fields such as multi-objective optimization,pattern recognition,signal processing,decision support,system control,pattern classification,fuzzy system control,neural network training and the like at presentClustering is to divide data objects into a plurality of groups,wherein each group is called as one class,and data in the class is called as objects.The purposes are to enable the characteristics between the objects in one class to be as similar as possible and to enable the characteristic differences between the objects in different classes to be as big as possible.The measurement of the characteristic differences is determined on the basis of the value of the data object to describe attribute,and the attribute is normally described through using the distance between objects.Clustering analysis is the process of dividing a data set into a plurality of classes, association rules can be found out through characteristic data of research objects,and thereby the clustering analysis is a powerful method for processing information.The clustering analysis is based on similarity,simplifies data through data modeling, thereby is widely applied in various fields such as data mining,pattern recognition, feature extraction,image segmentation,space remote sensing technique,signal compression and the like,and has obtained satisfactory effect.In an existing clustering algorithm,K-means clustering algorithm is simpler, thereby the method is normally used to select a distance measurement as the similarity measurement between the objects,then,a criterion function which evaluates the quality of clustering partition results is determined,and the best clustering partition result which enables the criterion function to obtain the extreme value is found out through an iterative method after an initial clustering center point is given.It is discovered in practical application process that the K-means clustering algorithm may be affected by an initial clustering center and is converged in suboptimal solution too early.In view of this,many scholars proposed a clustering method on the basis of genetic algorithm.The method adopts basic ideas of the K-means clustering algorithm, differently,the method uses the genetic algorithm to optimize cluster partition,the research discovered that the clustering method on the basis of the genetic algorithm does not need prior distribution knowledge of data which needs classifying and is less affected by initial solution selection to obtain the suboptimal solution.The paper uses the thought of using the genetic algorithm to solve the clustering problems for reference and realizes the algorithm of adopting the swarm intelligence theory to solve the clustering problems.The application of the genetic algorithm,the particle swarm optimization algorithm and the differential evolution algorithm in clustering segmentation is discussed through researching the genetic algorithm,the particle swarm optimization algorithm and the differential evolution algorithm,and the mixing of various algorithms is emphasized.Experiment results show that the use of the swarm intelligence algorism in solving the clustering problems is feasible and can achieve good effect.The main job of this article:First of all the paper makes a brief introduction of the subject sources,the background and significance of particle swarm optimization algorithm,as well as the development of the status at home and abroad.Then it introduces the clustering algorithm,have the simply analyze to existing clustering algorithm,give the mathematics description.At last give the detail research of the K-means clustering algorithm,describe the basically theory of the K-means clustering algorithm,algorithm flow and the excellence and disadvantage when solve the clustering problems.Nest the paper gives the algorithm based on DE for the division clustering problems.DE-algorithm is the tentative algorithm basically on the collective,it is easy to implement,no change of the parameter,the good capability.The DE-algorithm have three important parameters:F,weighting coefficient,the application of the new trial solution;C_R,chiasma probability,it decides the quantity of the trial solution can add to the exciting solution;evolution policy,it decides which evolution policy can be used to found the trial vector.The realization of DE-algorithm is limited by the F and CR,this article by dynamic change the F and C_R to improve the efficiency.On the solution of the clustering problems,it use the DELB- algorithm to solve closely to the best answer,the problem of the speed of converge will slow down.The result show that keep the speed of converge at same time,it can good to restrain the algorithm early converge.The algorithm based on PSO for the division clustering problems is proposed.In the swarm optimization algorithm of the paper,use the basically coding of clustering center,every particle represent the center point of the K cluster,then every particle can express z_i =(c_i1,C_i2,...,C_iK),the c_ij express the i particle of j cluster center coordinate vector,the clustering are buildup by many sort of solutions,the appraise of the sort solutions it the main importance on the application of optimize for the clustering.In the traditional K-means clustering algorithm inside the division clustering algorithm,can have the stochastic of the next generation cluster solutions, not easy to get into the local dinky solution.By the analyze of the theory and the answer of test,the mixed clustering algorithm can not only solve the problem traditional clustering algorithm depend on the initialized seriously and easy to get into the local dinky solution.,converge smoothly and no surge,also have the quickly speed of converge.Based on the algorithms above,the PSO/DE mixed algorithm for division clustering problems is proposed.Mainly introduce the PSO/DE algorithm which is mixed by the PSO-algorithm and DELB-algorithm,talk about the basically structure and compute theory of the mixed algorithm.The mixed algorithm mentioned on this paper use the community network structure model,can prevent the information through the network to protect the inside difference,at same time can improve the capability of algorithm,restrain the algorithm early converge to the suboptimal solution.The mixed algorithm has two choices update policy to the every particle:DE update policy and PSO update policy.The mixed choice policy enable the mixed algorithm inherit the advantages of the differential evolution algorithm and particle swarm optimization algorithm,experiment results show that this algorithm ave the best effect to have the best solution and speed of converge.