On the optimization aspects of parameterized neurocontrol (PNC) design

Abstract

Neural-networks are fast gaining acceptance as a new technology for manufacturing control. The applications of the technology are diverse and can be classified into two broad categories: neural-networks as process models and as controllers. Much has been written on the first topic, and in this article we focus on the second. We review several ways of developing neural-network controllers, concluding with the concept of parameterized neurocontrollers (PCN's). The PNC concept allows a neural-network controller to be developed once by extensive off-line optimization and then used over a range of professes and performance criteria without any further application-specific retraining. The concept is illustrated with a simple example. Throughout this article, we view neural-network control design as a nonlinear optimization problem and discuss related aspects such as the choice of cost function, the type of optimization algorithm, and the intimate connection between the two.