Integration of magnified gradient function and weight evolution with deterministic perturbation into back-propagation

An approach for the improved performance of back-propagation (BP) learning systems, based on the integration of magnified gradient function and weight evolution with deterministic perturbation, was discussed. In this regard, the regression and character recognition problems were considered. The simulation results, in terms of the convergence rate and global convergence, showed that the integrated approach always outperformed other traditional methods.