Continuous-time neural networks for system identification and control

A comprehensive technique that takes a total advantage of neural networks, in devising a fast learning neurocontroller is developed under this research. This is a concurrent process which uses a continuous time neuroplant and neurocontroller; and provides an effective design and algorithm for nonlinear plant control.;Over the last several years, there have been a number of researchers who have used feedforward and feedback neural networks for system identification and control to solve both linear and nonlinear problems. A majority of these techniques have used discrete algorithms, and have delivered successful neurocontrollers for specific applications. However, very little attention has been paid to continuous time neural network algorithms for the use of dynamic pattern recognition. A serious effort is made here in introducing a continuous time neuroidentification and neurocontroller design and architecture for the nonlinear plant control. The continuous time algorithm developed here is directly applicable to nonlinear plant, and eliminates any need for discrete plant models. The application of this algorithm along with the use of a rolling wave data window (RWDW) technique, allows for an easy on-line learning process. The RWDW method, as proposed here, has proven to be an effective means to handle and manage real-time plant data or dynamic patterns. In fact, this plays a critical role in selecting and organizing the dynamic patterns required by the neurocontroller. The neurocontroller design, once integrated with the RWDW and the on-line training process, results in an efficient, fast learning architecture. This work further introduces a process for achieving a standardized neural network architecture that is applicable to both regulating and tracking types of problems. Special attention is given to the overall neural network complexity in terms of the number of layers and neurons. In fact, it is shown that a simpler neural network configuration can produce accurate results without utilizing complex adaptive control techniques in parallel mode. This work further capitalizes on the fundamentals of neural network methods and applies them to achieve realistic and, at the same time, introduces a vibrant design of the continuous time neurocontroller.