Neural Networks for the Intelligent Control of Dynamical Systems

A control system is one in which some physical quantities are maintained more or less accurately around prescribed values. One of the fundamental contributions of control theory to science is the concept of feedback whose essence is the use of the error between the desired output and the actual output to control the system. Even very advanced control systems essentially use the same concept, but in such systems precisely what information is to be collected and how it is to be used in the control context is generally not evident. Examples of extremely efficient control systems abound in the biological world. The ease with which such systems perform the tasks of processing sensory information and interacting with uncertain environments has, for a long time, inspired engineers to attempt the design of artificial systems with similar capabilities. Most practical systems have multiple inputs and multiple outputs and the coupling that exists between them makes their control more difficult than that of single variable systems. External disturbances and parameter variations add to this difficulty and the problem of control becomes truly formidable when the system is nonlinear. Our objective in this paper is to describe a methodology for the fast, accurate, and stable control of nonlinear multivariable dynamical systems using neural networks.