The Research On Multiple Objective Evolutionary Algorithm For Interpretable Fuzzy System Construction

In this thesis, we discuss the interpretability of fuzzy systems and propose an agent-based evolutionary approach to extract fuzzy rules considering both the accuracy and interpretability. And two types of problems are applied based on the agent-based approach: the nonlinear system approximation problem and the classification problem. Experimental results show that multiple fuzzy system solutions can be obtained with better interpretability and a higher or comparative accuracy compared with other methods known in the literature.Chapter 1 gives a comprehensive description about the current research work. The research background, research objectives and my main contributions are given in this chapter.Chapter 2 gives some fundamental concepts about fuzzy systems. Then the interpretability issues are discussed in detail. We propose six items related to the interpretability of fuzzy systems. They are the completeness and distinguishability. consistency, compactness, utility. crossness, and coverage.In Chapter 3, we discuss some details related to the binary code genetic algorithm and real code genetic algorithm, both of which are the basis of the hierarchical genetic algorithm. In the end of this chapter, we introduce the hierarchical genetic algorithm which has the ability of topology optimization and has been demonstrated to be applied successfully to the optimization of fuzzy systems.In Chapter 4, we give a brief overview of the multi-objective evolutionary algorithm. Some fundamental concepts are introduced and the development of MOEA is depirted in the chronological order. A very promising type of the second generation MOEA named NSGA-II is described in detail, which is applied in our proposed agent-bused framework to study the trade-off between the accuracy and interpretability of fuzzy systems.Chapter 5 describes the proposed agent-based evolutionary approach in detail, including the inner behaviors of fuzzy sets agents and the interaction mechanism among them. The agent-based approach is proposed to build fuzzy systems considering both the accuracy and interpretability. For the interpretability issues, not only the number of rules can be optimized, but also the number of fuzzy sets and their distributions can be optimized. And our current agent-based approach is applied to two types of problems: nonlinear system approximation problems and classification problems.In Chapter 6. we use five benchmark problems to show the experimental results of our proposed agent-based approach. The 2nd order nonlinear plant. Lorenz system and Mackey-Glass time series are the examples for nonlinear system approximation problems, while Iris Data and Wine Data are taken as examples for classification problems. The experimental results are analyzed through the comparison with other methods in the literature.Finally. Chapter 7 concludes this thesis, and some future research topics are also proposed.