Solution to China's GDP Prediction Problem by BP Neural Network

摘要

Because the choice and important of learning rate η, the higher of η and the faster convergence it will be, but it may cause instability or function vibration if η is too high; if η is lower, although it may avoid instability, the speed of function convergence will reduce. In order to solve the contradiction, we introduce a variable of △W_(ij)(n_0) = η∑_(m-0)~n_0α~(n_0-1)δ_i(m)y_j(m) = -η∑_(m-0)~n_0α~(n_0-1)∑_m {partial deriv}F(m)/{partial deriv}W_(ij)(m) and if the ∑_m {partial deriv}F(m)/{partial deriv}W_(ij)(m) this time is the same as that of the previous time, the weighted summation value will increase and it results in the regulation speed of right value W at the stable regulation; and if the ∑_m {partial deriv}F(m)/{partial deriv}W_(ij)(m) this time is contrary to that of the previous time, it indicates that a certain vibration and now the result of summation will make the value of △W_(ij)(n_0) decrease to play a role in stability and increase the speed of function convergence.