Designing nonlinear algorithms with kernel functions satisfying the Mercer condition

has become a novel nonlinear technique in the machine learning. By using kernel idea the kernel perceptron algorithm nonlinearly generalizes the linear perceptron algorithm.It can handle the linearly non-separable classification problems in the original input space and the linearly separable ones in the feature space.The linear pocket algorithm improves the perceptron algorithm and can deal with the linearly non-separable problems directly.In order to improve the linear pocket algorithm and kernel perceptron algorithm

in this paper the nonlinear pocket algorithm based on kernels (i.e.kernel pocket algorithm) is proposed

whose objective is to find a nonlinear discriminant function that can minimize the number of misclassified training samples.Its convergence is also proved.Its advantage is to implement a nonlinear classifier using a simply iterative procedure and kernel functions.The experiment results from some benchmark data sets show that the performance of our kernel technique is prior to that of the linear pocket algorithm and kernel perceptron algorithm.