Adaptive Support Vector Regression

Abstract

A novel scheme for training support vector regression (SVR) with self-adaptive mechanism, called adaptive SVR (ASVR), is introduced herein to tune automatically user-defined free parameters, C and ε-tube, optimally in SVR. In the traditional support vector regression, two free parameters, C and ε -tube, are set in the default values, infinite and zero, respectively. However, this default setting is not optimal one for any SVR forecasting applications, and thus it may encounter some big residual errors leading to worst prediction accuracy. In order to best fit SVR model, adaptive support vector regression is applied to tuning free parameters C and ε-tube optimally. In such this way, the generalization capability can be enhanced in SVR model so as to improve prediction accuracy highly.