Neural networks for process identification

The application of neural networks to the development of dynamic models is considered. In particular, the authors present a common layered structure used for backward error propagation that is modified by the addition of direct linear connections between the input and output layers. For problems which have a significant linear component, such as those posed by process identification, this neural network structure offers significant promise. The neural network can be initialized in a meaningful fashion using the linear formation. Compared to standard neural network structures, the network can learn faster, can extrapolate better, and can be used to provide information on the extent of nonlinearities of the problem and on the learning algorithm itself