On a generalised backpropagation algorithm based on optimal control theory

A novel learning mechanism for the multilayered neural network is formulated as the optimal trajectory along which the state and weight vector of each layer should evolve. This approach leads to a rigorous proof of the backpropagation algorithm, points out several limitations of the generalized delta rule, and presents a way of overcoming them. A simple network is examined as the model for solving a nonlinear system identification problem. Simulated results reveal that the asymptotic accuracy and the convergence rate of the proposed algorithm are superior to those of the standard algorithm