Research On Kernels Method And Application

Since Prof. Vapnik proposed the Support Vector Machine (SVM) based on Statistical Learning Theory in 1995, kernel method based machine learning algorithm has been developed rapidly. It becomes one of the hot points in academic research now and has been widely used in image processing, biology information technology, text classification and intrusion detection, etc. So it is of great significance for both the development and improvement of kernel theory and its expansion of application.Though SVM has shown excellent performance in many fields, it is still in its early stage. Many issues in the theoretical research and practical applications are still to be solved, e.g., how to reduce the complexity of the SVM when trained on large-scale data, and how to construct new kernel functions for specific application fields.This paper mainly studies SVM classification and regression, including the simplification of SVM, as well as its application in agricultural pest forecast. The classification performance is improved and the application fields are expanded. The main results are as follows:(1) As for the low efficiency of the training and decision-making of the SVM on large-scale data, an SVM algorithm based on cooperative clustering is proposed, in which the number of support vectors is reduced effectively and the speed of classification is also improved. As for the multi-category classification, by further expansion of the cooperative clustering, a multi-category classification method is presented; RSVM and LS-SVM algorithm are improved.(2) The center of RBF is the key to the performance of RBF neural network classification problems. A method of selecting the center of RBF is illustrated, which is based on cooperative clustering. Comparison results with the k-means selection method show that the proposed method performs better.(3) For computing complexity is high due to support vectors are much when large sample of regression function is estimated, a method of regression SVM based on cooperative clustering is put forward. Experimental results show that test time needed by this method is significantly less than others, and the regression precision is higherthan others.(4) Based on a multiple mirror classification algorithm, and with respect to the fact that the selection of image points is quite complex, a new multiple mirror classification algorithm is proposed, in which the mirror pairs are replaced with cooperative clustering pairs and the training speed is upgraded effectively.(5) Agriculture information is an important symbol of modern agricultural techniques in our country; forecast of pest is an important link in agriculture information. Improving the level of prediction, we can decrease pest damage and increase economic benefit. Because the Support Vector Machine is a powerful tool of regression and classification, we have built a kernel method based agricultural pest forecasting mode. The proposed methods are applied in the prediction of agricultural pests and the results are quite good.