An Approach to Support Vector Regression with Genetic Algorithms

Support Vector Machines (SVM) are learning methods useful for solving supervised learning problems such as classification (SVC) and regression (SVR). SVM´s are based on the Statistical Learning Theory and the minimization of the Structural Risk [1], an enhancement over neural networks such as Multi-Layer Perceptrons. However, the major drawback is the high computational cost of the constrained Quadratic Problem (QP) combined with the selection of the kernel parameters they involve. Here we discuss varepsilon -SVRVGA, a detailed implementation of SVR that uses the non-traditional Vasconcelos Genetic Algorithm (VGA) [2] as tool for solving the associated QP along with the tuning of the kernel parameters. This work does not explore the automatic tuning of the regularization parameter C associated to the VC dimension [1] of the SVM what is considered an open research area. The varepsilon -SVRVGA fitting capability was tested with onedimensional Time Series (TS) data by reconstructing their n-dimensional state space [3] and adding Gaussian noise. Results show that varepsilon -SVRGVA is able to model successfully the TS in spite of a noisy environment as well as the self-selection of kernel parameters.