Algorithm Research On Support Vector Machine And Its Application To Intelligence Transport System

The Traditional Statistics is a gradual theory which the amount of specimen is tending to infinite, while Statistical Learning Theory (SLT) is a theory at the condition of small specimen amount. Since the specimen amount is always limited in reality, SLT has its theoretical advantage. Based on SLT and Structure Risk Minimum (SRM) principle, Support Vector Machine (SVM) is not only simple in structure, but also has good generalization ability. So SVM has gradually become one of hotspot issues in the field of machine learning.With the rapid development of communication, information science & electronic engineering and computer technology, Intelligence Transport System (ITS) has been attached extensive importance increasingly, thus put forward a greater demand on technologies in many fields, such as recognition of vehicle-type and vehicle-plate, prediction of traffic flow etc. The purpose of this article is to, based on the SVM, conduct research on basic methods of model identification and regressive analysis, and their applications to ITS.This article includes the following main tasks and existing achievements:1. Based on the Proximal SVM, this thesis drew upon the idea of incremental learning algorithm and the solution of imbalance in training specimen amount at different sorts, later defined a new kind of weighted coefficient matrix, put forth a multi-class classification algorithm of balanced and increment. Experimental results have showed that this algorithm has fairly high stability and classification accuracy.2. By conducting research on several multi-class classification algorithms, summarizing their merits and demerits and combining with technology of Huffman Tree, this dissertation brought forward a multi-class classification algorithm which is based on Huffman Decision Tree. Later on conducted simulative experiments on many opened databases, have proved that algorithm is not only with relatively high classification accuracy, but also with very high training efficiency.3. Firstly, we conducted relatively systematic research on theory and realization of Support Vector Regression, then analyzed the merits and demerits of v-SVR, and further put forth the v-ESVR algorithm. Finally, conducted the experiment, the results of which indicate this algorithm is not only with similar features to v-SVR, but also its dual problem is very simple, so can be used to design incremental learning algorithm...