Research And Application On Data Clustering Using Cellular Automata Based On Cell Clustering

Research on data clustering has a long history. Over the past years, its importanceand cross field feature with other research areas have been proved. However, there arefew research achievements on the application of cellular automata on data clustering.This paper studies the application of the combination of cellular automata and antclustering algorithm on data clustering. Through introducing the concept of cellclustering, we discuss the improvement on the previous algorithms.The paper first studies a clustering method based on linear cellular automata.Through discussions on boundary condition and neighborhood radius, the originalalgorithm is refined and improved, and the convergence speed is greatly increased. Next,based on the discussion of the divide point of different clusters, we point out thedrawbacks of the original algorithm, and through introducing the concept of cellclustering, the above problem is solved. Then, based on the observation of experimentresults, a cluster combination algorithm is proposed to decrease the number of clusterswith little influence on clustering accuracy. After that, the influence of differentparameters on clustering results is discussed through experiments on both synthetic andUCI data sets. Moreover, the algorithm is applied to Radar Emitter Recognition.Through the clustering of different types of the same radar class and different types ofdifferent radar classes, the effectiveness of the algorithm is further verified, and theinfluence of the algorithm parameter on similarity measurement is shown.Next, the concept of cell clustering is further introduced to a two dimensionalcellular automata clustering algorithm called ASM (Ant Sleeping Model). The cellcluster calculation formula and moving strategy as well as the specific implementationare provided. Experiment results show that the introducing of cell clustering to ASMcan solve the problem of class label decision as well as reduce iteration number at thesame time. Meanwhile, according to experiment results, we find that the convergence ofthe algorithm can be further speeded up through reducing the grid size to a certainextend.