EKF Based Multiple Parameter Tuning System for a L2-SVM Classifier

Performance of support vector machines (SVM) is sensitive to the setting of kernel and regularization parameters. Hence, parameter selection becomes an important challenge that the SVM users need to face. In this paper, it is shown that the multiple parameter tuning for a 2-norm SVM (L₂-SVM) classifier could be viewed as an identification problem of a nonlinear dynamic system, which could be solved using the extended Kalman filter (EKF), because of the reachable smooth nonlinearity of the L₂-SVM system. We describe the proposed method and compare it with the commonly used gradient descent (GD) approach using the Wisconsin Diagnosis Breast Cancer (WDBC) data from the UCI benchmark repository. We demonstrate that the EKF approach can be an effective tool for the multiple SVM parameter tuning.