Abstract: The method of kernel function has been widely applied in machine learning field，such as artificial neural network and support vector machine，for avoiding dimensional disaster and improving performance of learning machine.But the selection of kernel function and construction of new kernel are the most important problems，which have a closed relationship with the performance of classification.But the research work in this field is not enough.This paper uses SVM as an example to evaluate the performance of frequently-used kernel functions through observing and computing the kernel matrix.Based on this，authors use the selected kernel functions to get a new mixed kernel function.Experiential data prove that the performance of SVM is improved by the mixed kernel function.If we select the weighted values properly，the correct rate even is 100%.This can not only give us a method to get a new learning machine，but also give a reference for selecting kernel function.

Key words: support vector machine, neural network, kernel function, kernel matrix, mixed kernel function

摘要： 核函数方法广泛应用于人工神经网络和支持向量机等机器学习领域，该方法的采用有效地避免了特征空间中的维数灾难的问题，改善了学习机的分类性能。但是核函数的选择及新的核函数构造一直机器学习领域的核心问题，直接关系到学习机性能的好坏。然而，这个方向的研究成果不多。以支持向量机为例，通过对核矩阵一些特性的计算和研究，从理论上对常用的核函数性能进行了预测。在此基础上，通过实验仿真证实了通过优选后的核函数所组成的混合核函数对分类性能的改善。在加权系数选择合适的情况下，学习机的识别率甚至可以达到100%。所以，不但构造出了性能优异的学习机，而且为核函数的选择提供了参考。

关键词: 支持向量机, 神经网络, 核函数, 核矩阵, 混合核函数