Abstract
A two-stage algorithm is proposed for the learning phase of support vector machines (SVM). The algorithm is a combination of the Sequential Minimal Optimization (SMO) and the projected quasi Newton method. Use of the quasi Newton method in the neighborhood of optimal solutions results in a substantial improvement upon SMO in the number of iterations, and hence in numerical accuracy of the solution. Computational results on the UCI Adult and Web data set show that the two-stage algorithm performs comparably with SMO in usual parameter settings, but outperforms SMO for large C values and small tolerances.
