ИССЛЕДОВАНИЕ МЕТОДОВ И АЛГОРИТМОВ РЕДУКЦИИ НЕЧЕТКИХ ПРАВИЛ В МОДЕЛЯХ ОЦЕНКИ ДИСКРЕТНОГО СОСТОЯНИЯ ОБЪЕКТОВ

Abstract

The article describes the stages and results of the study of methods and genetic algorithms used to reduce fuzzy rules in models for assessing the discrete state of objects. Such models are fuzzy knowledge bases and fuzzy logical inference algorithms. A study was conducted of the influence of the parameter values of genetic algorithms on the results of reduction of the initial and intermediate knowledge bases (on the reduction of the number of rules in the knowledge bases and on the values of classification quality metrics). The following stages of research were implemented: selection of data sets from publicly available sources, formation of research knowledge bases, study of the influence of genetic algorithm parameter values on the results of knowledge base reduction, analysis of research results and development of recommendations for choosing optimal values of genetic algorithm parameters to achieve the best reduction results. Data for analysis was obtained from the public source UCI Machine Learning Repository. The selected sets are: Iris, Banknote Authentication and User Knowledge Modeling. To create research knowledge bases, the software package “Neuro-fuzzy system for generating fuzzy models for assessing the discrete state of objects” was used. Based on the analysis of the Iris data set, 3 research knowledge bases were formed, including 269, 339 and 542 fuzzy rules. For the Banknote Authentication data set, corresponding knowledge bases with 335, 643 and 1038 fuzzy rules were generated. For the User Knowledge Modeling set - with 877, 2094 and 2797 rules. For each of the knowledge bases, the values of the classification quality metrics Accuracy, Precision, Recall and F1-Score were calculated on the training and test samples. The reduction of knowledge bases was carried out in two successive stages, at each of which the corresponding genetic algorithm was used. Based on the results of the research, recommendations were developed for choosing the optimal values of the parameters of each of the genetic algorithms. The reduction of research knowledge bases based on the developed methods and algorithms made it possible to significantly reduce the number of their fuzzy rules (by an average of 85%) and at the same time slightly increase the generalizing ability of fuzzy models (by an average of 1.5%). In general, this made it possible to reduce the complexity of fuzzy models and their constituent knowledge bases, increase the accuracy of estimating the discrete state of objects and improve the interpretability of the generated solutions.