Robust estimation of support vector regression via residual bootstrap adoption

Abstract

As current system designs grow increasingly complex and expensive to analyze, the need for design optimization has also grown. In this study, a more stable approximation model is proposed via the application of a bootstrap to support vector regression (SVR). SVR expresses the nonlinearity of the system relatively well. However, using SVR does not always guarantee accurate results because it is sensitive to the input parameters. To overcome this drawback, we apply a bootstrap to SVR, using the residual from SVR as the bootstrap. The performance of the proposed method is evaluated via application to numerical examples and a real problem. We observed that the proposed method not only produced valuable results but also noticeably eliminated the negative effects of input parameters.