Basedon statistical learning theory, Support Vector Machine (SVM) is a novel type of learning machine, andit contains polynomial, neural network andrad ial basis function (RBF) as special cases. In the RBF case, the Gaussian kernel is commonly used, while the spread parameter, in the Gaussian kernel is essential to generalization performance of SVMs. In this aper, determination of σ is studiedbasedon discussions of the in8uence ofσon generalization performance. For classification problems, the optimalσcan be computedon the basis of Fisher discrimination. Andfor regression problems, basedon scale space theory, we demonstrate the existence of a certain range ofσ, within which the generalization performance is stable. An appropriateσwithin the range can be achieved via dynamic evaluation. In addition, the lower boundof iterating step size ofσis given. Simulation results show the e:ectiveness of the presentedmethod.