A hyperparameters selection technique for support vector regression models

Support vector regression models are powerful surrogates used in various fields of engineering. Due to the quality of their predictions and their efficiency, those models are considered as a suitable tool for surrogate evaluation. Despite their advantages, support vector regression models require an accurate selection of the configuration parameters in order to achieve good generalization performance. To overcome this limitation, a new hyperparameter selection method is developed. This method takes into account the training error to identify the optimal parameters set using evolutionary optimization schemes. Moreover, building on state-of-the-art techniques, an alternative analytically-assisted genetic algorithm is proposed in order to enhance the accuracy and robustness of the optimization scheme. The configuration is elaborated from a new search strategy in the design space. The results verify that the proposed technique improve the prediction accuracy and its robustness. Several test cases are used to demonstrate the capabilities of the method and its application potential to real engineering problems. The results prove that a surrogate model coupled with this adaptive configuration technique provides a useful prediction model suitable for various types of numerical experiments.