2001 年 121 巻 6 号 p. 1026-1034
This paper presents a new type recurrent neural network and its learning algorithm for nonlinear dynamics named “Velocity-Error Backpropagation (VEBP).”
In VEBP, learning is performed by 2 steps: (a) the velocity vector field of reference trajectories is approxi-mated by feedforward neural network with bi-connection layers by backpropagating velocity errors directly.
(b) recurrent neural network is constructed by adding integrators and output feedback loops to the trained feedforward neural network.
VEBP has some advantages with conventional learning method for recurrent neural networks named “back-propagation through time (BPTT).” Effectiveness of the presented recurrent neural network and its learning algorithm is demonstrated by simulation results for some examples of nonlinear dynamics.
J-STAGEがリニューアルされました! https://www.jstage.jst.go.jp/browse/-char/ja/