Intelligent closed-loop control using dynamic recurrent neural network and real-time adaptive critic.

摘要

Researchers in the field of intelligent control are introducing new concepts and techniques for control. Given a control problem, researchers working in the field of intelligent control typically use an approach to control that is motivated by the forms of representation and decision-making in human/animal/biological systems, and often heuristically construct what turns out to be a nonlinear, perhaps adaptive controller. While simulations results are typically used to "verify" the approach and successful implementations have been achieved (e.g., via fuzzy, expert, and neural network control), it is often the case that no nonlinear stability analysis is performed to verify the behavior of the closed-loop system.; Neural network have been proven to be very efficient for the control of nonlinear dynamical systems. Neural networks make use of nonlinearity, learning ability, parallel processing ability, and function approximation for application to advanced adaptive control. Major neural network topics include supervised learning control, inverse control, neural adaptive control, back-propagation of utility, and adaptive critics. However, neuromorphic control is effective only for a specific task in a specific environment, since the neuromorphic controllers have no meta-knowledge or data base.; In recent years, fuzzy logic has emerged as an important tool to control a system whose model is not known or ill-defined. The fuzzy logic nonlinear universal function approximation property, shared with feedforward neural networks, is often utilized. Fuzzy logic is very powerful because the fuzzy inference engine can be derived from either numerical data or linguistic knowledge. However, most fuzzy logic controllers fail to provide rigorous stability analysis.; In this Ph.D. dissertation, a rigorous mathematical formulation is presented to guarantee system performance under approximation-based control. The control of systems with unknown dynamics is accomplished using neural networks and fuzzy logic systems. Novel on-line learning algorithms are developed based on Lyapunov theory, and tracking performance and robustness properties are rigorously proven with performance verified through numerical examples.