Abstract:
Smoothing support vector machine (SSVM) can be solved by Newton algorithm and other fast algorithms. The classical smooth functions include the integral of sigmoid function, polynomial function, interpolation function, splines function and so on. From theoretic and numerical experiment, the paper compares and studies the accuracy of popular smooth functions approximating plus function and the convergence speed of the favorite algorithm for SSVM including Newton-Armijo algorithm, BFGS-Armijo algorithm and Newton-PCG algorithm.It is shown that the more the smooth function approximates plus function, the more accurate the solution is, while the train time is heavily increased. Newton-PCG algorithm is the fastest one, and Newton-Armijo algorithm is the slowest one.
