Kernelized based functions with Minkovsky's norm for SVM regression

Presents an empirical study for support vector machine (SVM) regression using Minkovsky's norm in a Gaussian kernel function. Due to the encouraging results with RBF kernels, more generalized forms based on some distance measure are suitable to be investigated. The Euclidean distance has a natural generalization in the form of the Minkovsky distance function. The results presented on the approximation of sincos functions as well as on a time series prediction function show that Gaussian kernels with Minkovsky's distance (/spl alpha/ = 3) and (/spl alpha/ = 6) evaluated on a 10-k cross validation basis present better generalization accuracy than RBF kernels (/spl alpha/ = 2).