Abstract:

SVM (support vector machine) classification performance is mainly influenced by the SVM model selection (including the choice of the kernel function and parameters selected). It is not better to determine the SVM model parameters by the existing methods of SVM model parameter selection. Therefore a SVM parameter selection algorithm is presented based on the Fisher criterion. The selection algorithm makes full use of the samples of linear separability in the classes in the feature space, and combines with the gradient descent algorithm for parameter optimization. It is realized by Matlab. The experimental results show that this parameter selection algorithm not only improves the training performance of SVM, but also greatly reduces the training time through the simulation.

Key words: kernel function; SVM; Fisher criterion; gradient descent algorithm