Abstract: Using the sparseness of a Suport Vector Machine(SVM) solution, properties of the raidal basis function kernel and the inter-median parameters in training the SVM, a new algorithm for estimating the generalization performance is presented, which overcomes many disadvantages of the existing algorithm such as longer computation time and narrower application range without additional complex computing. Theoretic analysis and experiments prove that it is a universial method for estimating the generalization performance of an SVM and it can be applied to a wide range of problems of pattern recognition using SVM.

Key words: raidal basis function support vector machines, generalization performance estimating, cross-validation, leave-one-out, pattern recognition