Neural network fits to neutron induced reactions using weighted least-mean-squares

A new error function based on weighted least-mean-square analysis is suggested for use in a back-propagation algorithm in neural networks. The standard back-propagation algorithm has been revised to minimize the sum of the weighted residual error squared. This method has been applied to obtain the cross-section for 14.6 MeV neutron induced (n, p) and (n, α) reactions in the mass region 40 < A < 240. In comparison to the normal back-propagation algorithm based on simple least mean square error estimation, the proposed method has better estimation power particularly in a situation when the targeted outputs have large variations in experimental errors.