Study On Several Issues Of Support Vector Machine

In this dissertation, several aspects about SVM(support vector machine) havebeen studied substantially as follows: In order to improve the speed of training SVM, through analyzing the mainreason that the training speed of SVM is slow, we put forward a pre-extractingSVs algorithm, an iteration algorithm, an incremental learning algorithm and asamples discarding scheme of the training of SVM. Under guaranteeing thecorrectness of training, the training speed of SVM is largely improved and the innermemory requirement of training SVM is greatly decreased with all these schemes.Many results of experiments and simulations show that although there are part ofsamples are discarded in our method, the useful information of classification is notlost. When there are new samples, the new training of SVM inherits the previoustraining result, and learns on the basis of previous classifier. Therefore an efficientapproach for the on-line learning of SVM is provided. To solve the problems and defections of existing methods of SVM multiclassclassification, a new method of SVM multiclass classification based on binary treeis presented. Several simulations demonstrate that compared with the existingmethods, the number of SVMs need to be trained is less by using the new method,the speed of training and decision is fast and the region that can not be classifieddoes not exist again. A new fuzzy SVM based on density is presented , which overcomes thedisadvantage that the traditional SVM is so sensitive to noises or outliers in thetraining samples set, and so SVM's performance of classifying is greatly effectedby noises and outliers. A fuzzy parameter of sample density is introduced into SVMto diminish the effect of outliers and noise. Several simulations demonstrate thatthis approach has obtained a better effect on diminishing the effect of noises andoutliers than the method of class-center vectors and the method of class-centerdistance published in the literatures. This approach greatly improves thegeneralization ability of SVM classifying and extends its application area. A new function approximation and regression approach which is based onthree-SVM is presented. At the same time of keeping performance of goodinterpolation of SVM, the new approach overcomes the bad performance inextrapolation of SVM function approximation and regression. The superiority ofthe new approach is demonstrated by experiment. A on-line identification method of complex nonlinear black-box system basedon SVM is presented. And it is used to the modelling and prediction of the tinylength change based on magnetostriction. Simulation analysis indicates thatcompared with the approach of RBF neural network, the present method has thefeatures of high learning speed, good generalization and better prediction precision.