NSF Award
9014065
The state-of the-art control system design methodology for complex nonlinear systems is best described as a "black art". The traditional approach (gain scheduling) is an arduous, multi-step process: multiple linear controllers are employed in an ad hoc fashion requiring substantial manual adjustment. Although an extensive and well-developed theory exists for linear systems, a similarly effective theory for nonlinear control system design remains elusive. A further difficulty arises after the control system has been designed. Because the physical models used during the design process rarely accurately reflect the actual plant dynamics, extensive on-line manual tuning of the nominal control system is required to ultimately achieve satisfactory performance. Artificial neural networks (ANNs) offer unique capabilities that may be exploited to overcome the limitations of conventional nonlinear control system design methods; namely, the inability to accommodate spatial nonlinearities and adapt to time- varying dynamics. The potential benefits of practical neurocontrol systems apply to a wide assortment of complex nonlinear systems and processes; these include atmospheric and underwater vehicles, manufacturing processes, and robotic systems. For such applications, a neurocontrol system might be used advantageously to synthesize the required nonlinear feedback control law automatically, objectively, and near-optimally (given the appropriate reinforcement). To fully realize the potential benefits of this approach, new conceptual and empirical tools for designing and analyzing ANNs for feedback control will have to be developed. The main objective of this project is to develop the techniques, algorithms, and theory required to implement ANNs for the control of complex multi- input/multi-output systems having spatial nonlinearities and time- varying dynamics. An important underlying is to develop a sound theoretical framework that may be used to refine, formalize, and extend research results in this area.
