Comparing the performance of support vector machines to regression with structural risk minimisation

The structural risk minimisation (SRM) principle based on the statistical learning theory of Vapnik aims to prevent the phenomenon of overfitting by balancing the complexity of models with their fit to the data. This principle has been embodied in support vector machines, a widely acclaimed generic approach to machine learning. This paper investigates the performance of the SRM principle in its application to standard least-squares regression and compares it with its integration with support vector machines.