Discrete-time neuro identification without robust modification

In general, neural networks cannot exactly represent nonlinear systems. A neuro identifier has to include robust modification in order to guarantee Lyapunov stability. An input-to-state stability approach is used to create robust training algorithms for discrete-time neural networks. It is concluded that the gradient descent law and a backpropagation-type algorithm used for the weight adjustments are stable in the sense of L-infinity and robust to any bounded uncertainties.