The Research And Application Of Fast Learning Algorithm Of Multi-output Support Vector Regression With Data Dependent Kernel

SVM (support vector machine) has become a hot topic in recent years, because of its good generalization in solving the problems of which are nonlinear and small samples. Learning algorithm is an important part of the SVM, and the fast learning of algorithm is of great significance to the theory research and application of SVM.In order to improve the learning efficiency of the MSVR algorithm, and to solve the problem of slow convergence rate and low accuracy of prediction of the model parameters in the optimization process, this paper presents a sort of modified BFGS algorithm of second-order convergence rate, and adapts the linear search method to gain the fit step factor to make the MSVR convergence. It uses linear search technology to ensure the convergence by determining the step size, with the consideration of the declining volume and the global convergence of the iterative process model. Based on the analysis of the geometric structure of kernel function and the selection of kernel parameters, a simple multiple kernels learning algorithm is presented to improve the efficiency of the algorithm as the basic kernel function and applied to the prediction of wind speed and wind direction. The main research work as follows:1. The paper introduced the principle of SVM and its fast learning algorithm based on the statistical learning theory, and derived the solving equation of the regression model. The single SVR was extent to the MSVR, with the data dependent kernel function, by the correlation of the MSVR sample data. By analyzing of the disadvantages of traditional orisginal space methods, this paper proposed a general solution of the two programming problem to solve the regression model. An access controlled scheme supporting efficient revocation.2. To solve the problem of regression model, this paper proposed a modified BFGS algorithm of rank correction rule as the updated rule to construct an approximate matrix of Hessian matrix, and obtained an approximate positive definite matrix of Hessian matrix. Combined with the inexact linear search technique, the step factor in the model iteration process was gained, to make the algorithm have a faster convergence rate.3. By analyzing the spatial geometric structure of the kernel function and the samples complexity model of the model, this paper used a multiple kernels learning algorithm as the basic kernel function of the data dependent kernel. The model achieved better mapping performance by using a linear combination of the elementary kernel of which enables the model use the multi-feature space of the corresponding sample. As a result,the speed of iteration algorithm was sped up by optimizing the model parameter. Finally, the MSVR applied to the prediction of wind speed and wind direction.