Complex Systems, Intelligent Modeling And Control

This dissertation explores the theoretical foundation and main approaches of intelligent modeling and control based on fuzzy logic and artificial neural networks. Taking advantage of the evolutionary algorithms and other intelligent computation methods, the dissertation tries to lay its emphasis on the development of self-organizing and self-learning algorithms for intelligent control of complex systems.Three parts are found in this dissertation. Part one briefly looks back on the history of artificial intelligence(AI), and compares the two overwhelming paradigms of AI research, namely, symbolism and connectionism. Inspired by the mechanism of natural intelligent systems, part one suggests a novel paradigm for intelligent control systems, which integrates most of the existing tools for intelligent control systematically. As conclusion of this part, the popular learning methods both for natural and artificial intelligent systems are surveyed.Structure self-organization and parameter learning of fuzzy logic control systems are first investigated in part two, which results in several new adaptive fuzzy modeling and control algorithms. After analyzing the characteristics of the Sugeno-type fuzzy system, part two proposes a specific neural network structure to realize the on-line adaptation of the fuzzy system. This adaptive fuzzy system exhibits distinctive merits when it is applied to modeling and control of complex systems. Another important contribution of part two of the dissertation is to broach a concept of fuzzy linearization for control of nonlinear systems. The related issues, such as the identification of the fuzzy sub-systems and the synthesis of the controller are fully described. Last but not the least, the stability problem of the fuzzy linearization system is discussed in part two.Part three concerns itself with the neural-network-based approach to intelligent control. On the basis of information criterion theory and genetic algorithms, the optimization of the structure and learning rate of neural networks is described. An adaptive neurocontrol scheme for a class of nonlinear systems is subsequently suggested. This work is completed with the analysis of stability and convergence of the neural controller. In order to overcome the limitations of the static feedforward neural networks, the well-known Hopfield network is adapted so that it is applicable for dynamic system modeling and control. The proposed dynamic neural network is then applied to robot modeling and model reference adaptive control.