Chinese Journal of Computers   Full Text
Title	Adaptive Construction for Kernel Function Based on the Feature Discriminability
Authors	REN Shuang-Qiao WEI Xi-Zhang LI Xiang ZHUANG Zhao-Wen
Address	(Institute of Spacial Electronic Information Technology, School of Electronic Science and Engineering, National University of Defense Technology, Changsha 410073)
Year	2008
Issue	No.5(803—809)
Abstract &   Background	Abstract This paper firstly gives the distinguishable condition for separating the features space by linear classification hyper surface. Then the paper puts forward a novel adaptive construction algorithm for the kernel function, which is based on the feature distinguishable condition and the property of kernel function. This new kernel functions model includes the polynomial model and the B-Spline model. As the experimental results shown, validated with the actually measuring data, the performance of the new adaptive kernel function, such as classification capability and generalized capability are improved obviously in contrast to the classical kernel function. Keywords SVM; feature discriminability; kernel function; polynomial; B-Spline Background Support vector machines (SVM) is a powerful machines learning method based on VC dimension theory and structural risk minimization principle, which are the important foundation of statistical learning theory (SLT). The kernel function will seriously effect the classification and generalized performance of SVM, so it is a urgent difficult problem in SVM to adaptively construct the kernel function. The general method includes two steps, firstly, selecting the kernel function model by experience, then giving the parameters value of the kernel function by some rules. In the few years, most kernel function construction algorithms are established by analyzing the experimental results in some special engineering fields. This paper puts forward a novel adaptive construction algorithm for the kernel function, which is based on the feature distinguishable condition and the property of kernel function. This kernel functions model includes the polynomial and the B-Spline model, and the model parameters can be estimated by an optimization problem. Consequently, the new method can conveniently select the optimum kernel function according to the data. This work is supported by the National Science Foundation of China, under grant No.60402032. The aim of this project is to solve the automatic recognition for radar observing objects. This paper will offer the kernel function adaptively construction method, which is an important portion to the automatic objection recognition using SVM.