A Novel SVR Parameter Selection Base on Bi-level Programming Problem

The selection of parameters plays an important role to the performance of support vector regression (SVR). In this paper, a novel parameter selection method for SVR is presented based on the bi-level programming problem. The proposed method does not need priori knowledge the value of the parameter epsilon. At the same time, the parameter epsilon can be calculated by the new SVR. And the number of the support vector will be controlled by the parameter C, even if the value of the parameter C is too big, the regression function still adapts to real function. And then, the complexity doesn't increase. Experimental results show that the better performance could be obtained by using the new SVR than the standard SVR.