Research On Clustering Algorithm For Automated Learning And Adaptation Based On ECoS

With the wide application of artificial neural networks, data mining, intelligent information processing methods, associated with the improved algorithm also continue to emerge, to varying degrees, angle to improve the performance of the algorithm. However, with the rapid growth of the scale of the data, people want to take advantage of adaptive thinking freed users from the tedious process of data processing, and improving work efficiency.Proposed by Kasabov, ECoS (Evolving Connectionist Systems) effectively integrates the idea of adaptive self-learning, rapid learning of a large number of existing data, and incremental learning of new data. Therefore, the paper based on ECoS system research will apply the idea of self-adaptive self-learning on classical data mining clustering algorithm and artificial neural network algorithm to solve the algorithm with some defects. The main research as follows:(1) Analyzing and studying the classical clustering algorithm-k-means algorithm. Improving algorithm is proposed based on the silhouette coefficients. It adaptively selects the outstanding samples to determine the initial cluster centers. By calling the traditional k-means clustering algorithm and according to silhouette coefficient and distance from the center, improving algorithm adaptively selects the excellent samples and gets average of its excellent value as the initial cluster centers. With experiments on UCI datasets to verify its effectiveness, the improving algorithm has some advantages on time than other optimization algorithms.(2) Analyzing the ECoS systems, choosing one thought concise, fast processing algorithms, ECM, focus on it. ECM can adaptively determine the number of clusters and cluster centers on the basis of the existing cluster, and the new data can be processed directly. Compared with traditional clustering methods, the ECM algorithm in dealing with large data, using of existing knowledge, shortening clustering time, has a significant advantage. This paper carried out simulation experiments of ECM incremental learning algorithm, and visually analyzed their incremental processing of training.(3) Using improved k-means algorithm based on the silhouette coefficients to complete radial basis function neural network in the center of the selected basis functions,.as a result of improving the basis function center reliability.And using one-dimensional analog data to verifying the effectiveness of algorithm fitting function,using multidimensional UCI data sets to verifying classification algorithm predictive validity.(4) Applying ECM algorithm to complete the self-organizational selection to achieve a self-incremental learning RBF. And combining ECM with RBF in Matlab, using the GUI to simulate the incremental learning process.(5) Analyzing the possibilities of using ECM algorithm for outlier detection. And providing new ideas for further research of the efficient data preprocessing.