Weight value initialization for improving training speed in thebackpropagation network

摘要

A method for initialization of the weight values of multilayernfeedforward neural networks is proposed to improve the learning speed ofna network. The proposed method suggests the minimum bound of the weightsnbased on dynamics of decision boundaries, which is derived from thengeneralized delta rule. Computer simulation in several neural networknmodels showed that the proper selection of the initial weight valuesnimproves the learning ability and contributed to fast convergence