| Neural networks have been successfully used to find solutions to various problems requiring the minimization of cost functions. Such problems often occur in control design for the generation of the control law and for system identification coupled with control techniques. For this work, neural-network-based controllers were designed by minimizing cost functions for weighted one-step-ahead control and model-reference control. Stability requirements were developed. When system models were not analytically available, a second network was used for output prediction and to determine the sensitivity of performance cost functions to control weights. The estimator networks were trained using observed data. Thus, even in the absence of analytical plant models, control can be achieved.; An independent on-line, discrete-time, Lyapunov-based system identification method was developed. The technique, based on a Taylor series expansion of a cost function, was shown to be effective for various estimator forms including parametric linear models, parametric nonlinear models, and neural-network-based models. The method employs several updating schemes and adjusts estimation weights based on a performance measure. This approach was used to generate the estimates needed for the state transformation and the control law for an adaptive backstepping approach. Stability bounds were developed based on estimator quality. The technique can be applied to a wider variety of plants since the system uncertainties can be more general than with traditional backstepping control. |
