Some Modified And Designed SVM Algorithms

Conventional support vector machine (SVM) solves classification problem by a constrained quadratic programming problem, and has good generalization ability. Moreover, it has been used in some fields such as bioinformatics, face recognition, handwriting recognition, fringerprint recognition, database learning, identification validation and so forth. This thesis based on statistical theory lays emphasis on some researches on construction of kernels, on fast training SVM algorithms and on cropping techniques of the training set of SVM.This thesis mainly completes tasks as follows:1 Design a mixture of kernel (Kop) that can balance between the generalization ability and the learning ablility based on SVM. Compared with conventional RBF kernel, the number of support vectors of SVM based on Kop is smaller, and with better classification performance.2 Propose a new classification algorithm based on support vector ----ACNN-SVM to overcome the flaw that NN-SVM can not guarantee it obtains the least needed antetype sample points.3 Propose a novel algorithm in this thesis: WMPSVM aiming to overcome the singular problems appearing in GEPSVM. It does not need to solve specific two hyperplanes but GEPSVM does. Moreover, it can correct the misclassification problem with GEPSVM at some cases, and avoid singular problems appearing in GEPSVM。4 PPSVM is proposed on the basis of within-class scatter. This algorithm guarantees good classification performance and learning speed,and then a new PSVM algorithm called FPSVM is constructed by reformulating the optimalization problem in PSVM. Compared with PSVM, PPSVM has better learning speed. This is so because it can avoid the calculation of some matrices and the multiplication of some matrices and other ones.