Abstract:

Solving a large-scale convex quadratic programming problem is the core of the support vector machine. An equivalence complementarity problem can be obtained based on an amended model of the surpport vector machine(SVM), therefore we propose a non-monotonic trust region algorithm for solving the complementarity problem based on the Fischer-Burmeister complementarity function. The new algorithm need not compute any Hesse or the inverse matrix, thus reducing the amount of computational work. Global convergence of the algorithm is proved without any assumptions. Numerical experiments show that the running speed of the algorithm is faster than that of the LSVM algorithm and the descent algorithm when solving largescale nonlinear classification problems and thus it provides a feasible method for solving SVM.

Key words: support vector machine；trust-region method；complementarity function；nonmonotonic strategies