Improving performances of Battiti-Shanno´s quasi-Newtonian algorithms for learning in feed-forward neural networks

The authors describe a new improved Quasi-Newtonian algorithm (named OSSV) for effective learning in MLP-networks. OSSV, which is a variant of Battiti-Shanno´s original OSS method, is able to speed-up the convergence process of the network, maintaining an O(N) complexity. Numerical results show that by OSSV the computational effort of the original OSS can be reduced by a factor increasing with the number of epochs