ASSESSMENT OF FEEDFORWARD NEURO-CONTROLLERS USING DESCRIBING FUNCTIONS

Neural controllers (NCs) are typically analyzed only in the time domain using performance measures such as RMS tracking error. The analogy is made here between NCs and human operators in compensatory and pursuit control tasks. Frequency domain methods of analysis such as describing function estimation have been used to study human operator models, and here we present an analysis of a NC using similar techniques. The NC used in this analysis is an Adaptive Cluster Network in a feedforward architecture. The controlled element is a simplified model of an F-18 pitch rate system, and the control objective is to minimize tracking error. In a numerical experiment the NC is trained using a sum-of-sinewav e command signal and then the NC describing function is shown to closely match the predicted frequency response. An actuator failure is simulated via a reduction in the control effectiveness, and the NC is shown to reconfigure itself in response to this failure.