Um paradigma baseado em algoritmos genéticos para o aprendizado de regras Fuzzy

Abstract

The construction of the knowledge base of fuzzy systems has been beneficited intensively from automatic methods that extract the necessary knowledge from data sets which represent examples of the problem. The evolutionary computation, especially genetic algorithms, has been the focus of a great number of researches that deal with the problem of automatic generation of knowledge bases as search and optimization processes using di erent approaches. This work presents a methodology to learn fuzzy rule bases from examples by means of Genetic Algorithms using the Pittsburgh approach. The methodology is composed of 2 stages. The first one is the genetic learning of rule base and the other one is the genetic optimization of the rule base previously obtained in order to exclude redundant and unnecessary rules. The first stage uses a Self Adaptive Genetic Algorithm, that changes dynamically the crossover and mutation rates ensuring genetic diversity and avoiding the premature convergence. The membership functions are defined previously by the fuzzy clustering algorithm FC-Means and remain fixed during all learning process. The application domain is multidimensional pattern classification, where the attributes and, sometimes, the class are fuzzy, so they are represented by linguistic values. The proposed methodology performance is evaluated by computational simulations on some real-world pattern classification problems. The tests focused the accuracy of generated fuzzy rules in di erent situations. The dynamic change of algorithm parameters showed that better results can be obtained and the use of don t care conditions allowed to generate a small number of comprehensible and compact rules.