Intelligent control based on fuzzy logic and neural network theory
Chuen-Chien Lee, Lotfi A. Zadeh · NASA Technical Reports Server (NASA) · 1990
A promising direction in the conception and design of intelligent systems involves the use of fuzzy logic and neural network theory to enhance such systems' capability to learn from experience and adapt to changes in an environment of uncertainty and imprecision. This thesis is an investigation of fuzzy logic based systems and cognitive neural models, and explores an intelligent system by integrating these multi-disciplinary techniques. During the past several years, many applications of fuzzy logic based systems for the control of industrial processes have successfully demonstrated that such a fuzzy logic based system could stimulate and even surpass the decision-making ability of a skilled human operator. In this thesis a survey of a fuzzy logic based system is presented; a general methodology for constructing such a system and assessing its performance is described; and problems which need further research are pointed out. Given the limitations of present-day neural networks, it is plausible to shift the emphasis placed on the macroscopic capabilities from the network level to the neuronal level. The higher level of intelligence could then be built on top of this. In this connection, the thesis introduces cognitive neural models that are consistent with animal learning theory and provide a basis for understanding and explaining Pavlovian conditioning and instrumental conditioning, respectively. In particular, one model captures the predictive nature of Pavlovian conditioning, which is essential to the theory of adaptive/learning systems. The other model reflects the associative nature of instrumental conditioning, which stores in memory the temporal aspects of behavior in the context of content-addressable memory systems. In relation to the design of intelligent systems, the thesis describes a self-learning rule-based system which integrates a fuzzy logic based controller with a cognitive neural model stimulating learning behavior and memory association. In effect, the proposed intelligent rule-based system learns from experience and modifies its rule base for better control strategy. Computer simulation results show that the learning capability of our system represents an improvement over previous approaches. Furthermore, the proposed system is relatively insensitive to variations in the parameters of the system environment. In addition, the system could be primed with pre-trained control knowledge which minimizes rapid changes during adaptation.