IMPROVING THE TRAINING PERFORMANCES OF MULTILAYER NEURAL NETWORKS BY SAS-BASED OPTIMAL ESTIMATION OF INITIAL WEIGHTS