Reconstruction Model Of Support Vector Machine And Research On SMO Algorithm

Support vector machine, which is a branch of machine learning, has been developed during recent years. It is known as a novel and unique machine learning method with very high efficiency in dealing with the classification and regression problems.In this thesis, SVM and its reformation model are researched. Firstly, we state the background of the machine learning and its significance, provide a summary of SVM and its development, mainly discuss the algorithms of SVM. SMO and its working set selection method are introduced in details. Secondly, we rebuild the model of SMO by using the reproducing property of Mercer kernel in a Reproducing Kernel Hilbert Space(RKHS). In order to promote the model, we introduce the error. The new model reduces the restriction of the solution so that it can be used in common situation. During the process of solving the model, direct method and SMO algorithms are compared. The direct method decomposes matrix directly, then yields analytic solution with the inverse matrix calculated. The SMO algorithm only uses two rows of the multidimension matrix to carry on the optimization per iteration, so the memory space is reduced. Because only a two-variable quadratic programming is solved per iteration, it is easy to get the analytic solution and reduces the algorithm running time. The calculated rows are recorded by cache technique, the computational cost of the kernel matrix is reduced and the efficiency of the algorithm is improved.The experiment shows that comparing with the direct methods, the SMO algorithm can increase the dimensions of the matrix and accelerate the operation speed. Adopting the technology of cache can save time in calculating 3large-scale matrix and raise the operation efficiency.