Study Of Genetic-based Neuro-Fuzzy System

Fuzzy logic, neural network, and genetic algorithm are nowadays three key technologies in the field of artificial intelligence. How to integrate these three technologies into one system and make them complementary and mutually beneficial and coordinately work in this system is even a current hot research subject and a very promising issue for study.Firstly, fuzzy inference system, neural network, and genetic algorithm that are three intelligent technologies are systematically introduced in this paper. Secondly, based on the analysis of strong and weak points of fuzzy system and neural network, several ways for their combination are discussed. And then, this paper focuses on the study of the structure and reasoning mechanism of Adaptive-Network-based Fuzzy Inference System (ANFIS), which is based on Sugeno fuzzy model and established by the way of equal-structured combination, and analyzes its learning method in detail. In view of the inherent disadvantage, high possibility of reaching local extremum, of the gradient-descent-based learning method, this paper presents a new hybrid learning method which combines the advantages of Genetic Algorithm (GA), Back Propagation (BP) algorithm and Least-Square Estimate (LSE). In this method, a global approximately optimal solution of all ANFIS parameters is obtained using genetic algorithm and then premise parameters and consequent parameters are fine tined respectively using BP algorithm and Least Squares Estimate. Therefore, this method not only improves the convergence of BP algorithm, which leads to greatly reducing the dependency of expert knowledge and enhancing the intelligent level of the system, but also improves the searching efficiency of genetic algorithm and strengthens the learning ability of ANFIS. Thispaper also presents a concrete scheme of simultaneously implementing the structure and parameters optimization of ANFIS using this new method. At last, in this paper the effectiveness of this method is verified by two simulation examples, and the simulation results show that it has better learning effects of ANFIS parameters than the former method.