Neural Controllers

Abstract

Conventional control techiques such as DIP control, robust control, and optimal control have traditionally been used to control linear and nonlinear systems. These techniques are based on the model of the system, and their performance is largely dependent on the accuracy of the model. Modern control systems are charactgerized by their increassing complexity, nonlinearity, intelligence, enhanced ability to adapt and learn, and roubstness to uncertainty. Often, the dynamics of these systems are time‐delayed, ill‐defined, extremely nonlinear, and very complex, which makes it difficult to obtain an accurate analytical model of the system. This reality added to the uncertainties present in the environment and measurement devices, makes it particularly challenging to employ conventional techniques to control such systems. The ability of neural networks to handle nonlinear and ill‐defined problems, perform in the presence of noise and uncertainty, learn and adapt to changes in system parameters, and approximate functions and identify nonlinear systems have made them a popular choice for controlling such complex systems. Extensive research has been carried out over the last two decades for using neural networks to control nonlinear systems, and several neural network architectures have been proposed. This progress has led to a widespread application of neutral controllers in several of fields, such as the aerospace, communications, automated manufacturing, robotics, medical diagnosis systems, electric power systems, and process industries.