The Method Of Parameter Selection On Kernel Function Of SVM

Classification is the fundamental issue in the research of machine-learning, which is tremendously used in the practice. The theory of support vectors machine basing on Statistic-learning draws various attentions because of its outstanding performances in dealing with the cases of small sample, nonlinear and high-dimensional pattern recognition.In this thesis, we briefly introduce the developing progresses of the theory of support vectors machine, analysis the kernel functions, which include basic properties of such functions. In particular, the reasons of Gauss kernel's well-broad used are introduced.To cope with the present situation that the select of Gauss kernel's parameter largely depending on experiences, together with cluster, a new approach, called cluster- minimum distance, is obtained, which improves the method named the minimum distance method. In the new method, we focus on the data itself rather than any experiences. The experiments exhibit that our new method exerts better performance than the minimum distance and largely reduces the cost of calculations.Finally, the summary of the whole work is given, and put forward some further researches.