Abstract: Kernel selection is an important issue in Support Vector Machine（SVM） modeling.SVM is a popular machine learning tool with good generalization ability，but its performance is often dependent on selected kernel function.For a given problem，it is difficult to choose an appropriate kernel function.This paper proposes an approach for kernel selection based on SVM ensemble.Some basic SVMs are constructed by adopting different kernel function or parameters and then the final prediction is obtained through aggregating the results of these basic SVMs.The proposed algorithm will integrate kernel selection with SVM learning.In so doing，not only the influence of kernel selection on a single SVM can be avoided，but also good generalization performance can be obtained.Simulation results on UCI benchmark datasets demonstrate the validity of the proposed approach.

Key words: Support Vector Machine（SVM）, ensemble learning, kernel selection, heterogeneity SVM, homogeneity SVM

摘要： 核选择问题是支持向量机（Support Vector Machine，SVM）建模中的一个关键问题，虽然支持向量机具有良好的泛化性能，但其性能受核函数的影响比较明显，而对于一个给定问题，选择合适的核函数及参数通常很困难。提出一种基于SVM集成的核选择方法，利用不同的核函数构造子SVM学习器，然后对子学习器的预测结果集成。提出的核选择方法将SVM集成学习与核选择同时进行，不仅避免了单个SVM的核选择对泛化能力的影响，而且可以获得良好的泛化能力。在UCI标准数据集上的结果说明了提出的方法的有效性。

关键词: 支持向量机, 集成学习, 核选择, 异质SVM, 同质SVM