A comparative study of neural vs. conventional methods for modeling and prediction

Neural network and conventional methods for system modeling and prediction are discussed in a unified way. Both linear and nonlinear examples are used to show that by using a black-box approach, the methods are equivalent with the exception of the parametrization process. A comparison of neural network methods with an extended Kalman filter for the case of a nonlinear system demonstrates that neural methods require very few a priori assumptions about the underlying model structure. It is shown, using a flexible space structure example, that neural networks can more readily handle the problem of underparametrization than conventional techniques. In all cases, the neural implementations provide results that are at least as accurate as the conventional methods, where the figure of merit is the variance of the output error signal