Research And Application On Incremental Learning Algorithm Based On Support Vector Machine

Support vector machine based on the statistical learning theory is a machine learning method to search for the optimal classification of the hyper plane,It had demonstrated the advantages in solving small sample,nonlinear,high dimensional data problems.However,the classical support vector machine training algorithm can not plunk for incremental learning.Therefore,it is significant to study the incremental learning about the support vector machine.In order to improve the performance of the classical incremental learning algorithm of support vector machine in training time and accuracy,the thesis divide this learning process into two steps,namely an initial training phase and incremental learning stage,Focusing on on the support vector machine method based on the improved initial training sample set and the selection of non support vector set in incremental learning.From geometric knowledge of Support Vector Machine principle,the support vector sets obtained at the initial stage of the training are included in the corresponding border vector set.the improved k nearest neighbor method is used to select the set of boundary vectors as the initial training set,the initial point is not selected randomly.But by finding the very points located in the circle whose radius is half of the distance between the sample center points and whose center is the middle point of the line connecting two sample center points.Compared to the k-nearest algorithm,this method appropriately reduces the initial training time.During the incremental learning phase,a valid non-support vector set is selected by using the Karush-Kuhn-Tuc Ker condition usually.A new method is more effective than the KKT condition is applied when the sample set is too large to meet the need.The new method based on the central density overcome the difficulty that the KKT condition will become more complex when the incremental sample set is gradually increase.The thesis formed a new support vector learning method on the basis of the boundary vectors combined with the initial training phase and incremental learning stage.The training time and classification accuracy of three Support Vector Machine incremental learning algorithms are compared by experiments,and verified the advantages of the improved algorithm of improving time and accuracy.Finally,the new method and classical method are both applied to the Breast-Tissue sample set to make an incremental learning experiments of recognition prediction.Simulation results show that the improved method is more effective concerning on improving the rate of recognition prediction than the classical method.