Ensembles of support vector machines for regression problems

Support vector machines (SVMs) tackle classification and regression problems by nonlinearly mapping input data into high-dimensional feature spaces, wherein a linear decision surface is designed. Even though the high potential of these techniques has been demonstrated, their applicability has been swamped by the necessity of the a priori choice of the kernel function to realize the nonlinear mapping, which sometimes turns to be a complex and non-effective process. In this paper, we advocate that the application of neural ensembles theory to SVMs should alleviate such performance bottlenecks, because different networks with distinct kernel functions such as polynomials or radial basis functions may be created and properly combined into the same neural structure. Ensembles of SVMs, thus, promote the automatic configuration and tuning of SVMs, and have their generalization capability assessed here by means of some function regression experiments.