| Artificial Neural Network (ANN), also known as connectionist learning and parallel distributed processing, is finding its applications in diverse fields: many branches of engineering, health sciences, cognitive science, archaeology, finance, etc. This research tries to make some efforts to emphasize "design" methodology in ANN, and to explore the structures by which ANN can solve difficult problems by identifying proper ANN architecture. Two classes of ANN--multi-layer neural networks and recurrent networks--are investigated in the context of control of systems and estimation of unknown parameters. The multi-layer neural networks converge to optimal solutions by satisfying mathematical formulations associated with the Hamilton approach and the dynamic programming approach. A benchmark aerospace application is used for illustration. A variant of the Hopfield network, called the Modified Hopfield Neural Network (MHNN), is proposed to show the design approach to the determination of weights in recurrent networks. It is shown how the equilibrium point of this network helps with inversion operations arising in optimal gain determination. Control of dynamic systems using recurrent neural networks are presented. The robustness of the recurrent networks to parameter variation is considered in the context of weights. Analyses are carried out in the frequency domain and the time domain. |
