The Research Of The Fuzzy Rules Extraction Based On Genetic Algorithms

With the development of the science and technology, the people' s stand of living being improved prominently, the piled data having doubled and redoubled, which make the information needed to be deal with increase rapidly. Converting the vast data into the information and extracting the useful knowledge from the data, turning the knowledge into the strategy of directing the jobs, production, and doing in business and selling.In the fuzzy systems, the knowledge of human being has to be repressed by fuzzy rules. Only in this way can the knowledge of described fuzzy rules be used for the human being. Fuzzy rules can be given by the experience of experts in real application, which have difficulty in extracting the rules of not impersonality enough and specialist' s experience not being acquired so on. Therefore, it is a speaking volumes job that how to extract fuzzy rules in compactness and efficiency. The classical methods of extracting fuzzy rules is determined by experience, which is complex and more difficult.We study Fuzzy System, Genetic Algorithm, Neural Network and the integration of these approaches. To extract fuzzy rules, we combine Genetic Algorithm and Neural Network method to optimize and learn fuzzy partitions and parameter of membership function when generating membership function. So we switched from a completely subjective problem to a highly objective form. Then generate initial fuzzy rules based on these steps. Thus fuzzy rules which we extract can easily be understood to human being.In designing GA, we propose adaptive fitness evaluating functions, adaptive crossover probability and adaptive mutation probability. In order to extract fuzzy rules that are understandable to human beings from data using Genetic Algorithm methods, the interpretability of fuzzy systems are added to the fitness function of Genetic Algorithms. We proved the convergence of GA which we proposed adaptive fitness functions, adaptive crossover probability and adaptive mutation probability. It shows that the methods we proposed are reasonable.