A comparison of fast training algorithms over two real problems

This paper compares the speed of convergence of several popular training algorithms over two real problems. The algorithms compared are backpropagation, SuperSAB, Quickprop, conjugate gradient, RPROP and SASS. This work builds on previous studies, by making use of a benchmark collection Proben1, which is designed to improve the quality of training algorithm comparisons. Statistical significance tests are applied to the results so that conclusions are backed by statistical evidence. The results show that the relative performance of the algorithms depends on the problem, but that RPROP and SASS produce the best overall performance