Abstract:

To optimize the metaparameters of ε insensitive support vector regression (ε-SVR), the metaparameters optimization problem is divided into two sub-problems named as kernel parameter optimization and structure parameters (including insensitive parameter and regularization parameter) optimization according to the coupling degrees among them, and corresponding optimization methods are proposed. First, a new kernel  alignment coefficient is proposed for the former. Second, a method based on expectation training error is proposed for the latter. Finally, to estimate accurately the expectation error of the ε-SVR, a method to evaluate and adjust expectation error according to the distribution characteristics of the real training errors is proposed. Simulation results show that the proposed method is nearly as accurate as the cross validation method, and much more rapid.