Neural networks as optimal nonlinear controllers - Concepts and flight control applications

Abstract

Current neural network controllers are designed relative to specific plant models and performance criteria with the intent of completely replacing conventional control designs. While some progress is being made towards this goal, we believe that a meaningful, and perhaps more achievable, intermediate goal is to augment conventional linear control designs with the nonlinear capabilities of neural networks to improve overall system performance. To this end we introduce the concept of parametrized neurocontrollers (PNCs), neurocontrollers with inputs that are used t o adjust control system performance and to provide information about the plant dynamics. PNCs are optimized in simulation over spaces of plant models and performance criteria; no application-specific training is needed. This structure circumvents one of the main drawbacks of traditional neurocontrollers; the need for individual training for different plants and retraining for any plant variations or changes in control objectives. A particular instance of the PNC concept, the NeuroPID controller where the external parameters are the well-known PID control gains, is presented. An application of the PNC concept t o flight control design is described.