Training a SVR(support vector regression)requires the solution of a very large QP(quadratic programming)optimization problem.Despite the fact that this type of problem is well understood,the existing training algorithms are very complex and slow.In order to solve these problems,this paper firstly introduces a new way to make SVR have the similar mathematic form as that of a support vector machine.Then a versatile iterative method,successive overrelaxation,is proposed.Experimental results show that this new method converges considerably faster than other methods that require the presence of a substantial amount of data in memory.The results give guidelines for the application of this method to large domains.