A Multi-objective Genetic-based Method for Design Fuzzy Classification Systems

An approach based on multi-objective genetic algorithms is proposed to construct interpretable and precision fuzzy classification system from data. First, a multi-objective genetic algorithm is used to accomplish feature selection and dynamic grid partition with three objectives: maximization of precision, minimization of the number of features, and minimization of the number of fuzzy rules. The parameters of the membership functions are determined by neighboring overlap method, so the obtained initial fuzzy classification system is highly interpretable. Second, a compact fuzzy classification system is obtained using a genetic algorithm with three objectives: maximization of accuracy, minimization of the number of features, and minimization of the average length of fuzzy rules. Third, a constrained genetic algorithm is used to optimize the compact fuzzy classification system to improve its precision, while preserves its interpretability. The proposed approach is applied to the Iris and Wine benchmark classification problems, and the results show its validity.