Synthesis of intelligent hybrid systems for modeling and control

Abstract

The intelligent information processing performed in humans is now being mimicked in a new generation of adaptive machines as the state-of-the-art technology. Inspired by the functionality of brain nerve cells, artificial neural networks can learn to recognize complex patterns and functions, and based on the biological principle of "survival of the fittest", genetic algorithms are developed as powerful optimization and search techniques. Likewise, fuzzy logic imitates the mechanism of approximate reasoning performed in the human mind, and hence can reason with linguistic and imprecise information.

Although these intelligent techniques have produced promising results in some applications, certain complex problems cannot be solved using only a single technique. Each technique has particular computational features (e.g. ability to learn, explanation of decisions) that make it suitable for particular problems and not for others. These limitations have motivated the creation of intelligent hybrid systems where two or more techniques are combined. Although there is an increasing interest in the integration of fuzzy logic, neural networks, and genetic algorithms to build intelligent hybrid systems, no systematic synthesis framework has been developed so far. Therefore, the objective of this thesis is to construct an intelligent learning scheme that incorporates the merits and overcomes the limitations of the three paradigms. he applications considered for the proposed scheme are modeling and control.

The generic topology of the system used in this thesis has a transparent structure; its parameters, links, and signals and modules have their own physical interpretations. Moreover, the learning scheme uses task decomposition to identify the systems' parameters. The learning task is decomposed into three subtasks (phases). The first phase performs a coarse identification for the systems' numerical parameters using unsupervised learning (clustering) algorithms. The second phase finds the linguistic-association parameters (linguistic rules) using unsupervised as well as supervised learning algorithms. In the third phase, the numerical parameters are optimized and fine-tuned using supervised learning and search techniques. The performance of the scheme is assessed by testing it on two benchmark modeling applications. The results are compared to that of other intelligent modeling approaches to show the performance characteristics of the proposed scheme. The scheme is also assessed by applying it to nonlinear control problems. The synchronous machine voltage regulation and speed stabilization problems have been tackled using the proposed scheme. Several comparative studies are carried out to show the advantages of the proposed control approach over conventional approaches.