A Novel Gaussian Kernel Function for Minimax Probability Machine

In recent years there is a growing interest around minimax probability machine (MPM) whose performance depends on its kernel function. Considering that the Euclidean distance has a natural generalization in form of the Minkovskypsilas distance, we replace the Euclidean distance in the Gaussian kernel with a more generalized Minkovskypsilas distance. This paper presents an empirical study for MPM prediction on Minkovskypsilas norm. The performance of this method is evaluated with the prediction of network traffic data for MPEG4, at the same timescale. Experimental results demonstrate that the best prediction accuracy is provided by kernels with Minkovskypsilas distance and the MPM using Gaussian kernels with Minkovskypsilas distance can achieve better prediction accuracy than the Euclidean distance.