Kernel selection is the key both in support vector machine（SVM） research and application.Existing methods mostly select a kernel from basic kernels based on data.However

recent applications have shown that using multiple kernels not only increase the flexibility of SVM but also enhance the interpretability of the decision function and improve the performance.Here

an accurate and efficient multiple kernel learning method

which is based on the upper bounds of the generalization error of SVM

is presented.The form of the solution of the optimization problem is provided first.Then

with an efficient iterative algorithm

the solution is computed.Finally

a learning bound of our proposed method is presented based on the Rademacher complexity.The bound has only a logarithmic dependency on the total number of basic kernels

which indicates that the bound is valid for very large number of basic kernels.Experiments show that our method gives better accuracy results than both SVM with the uniform combination of basis kernels and other state-of-art kernel learning approaches.