Research On Fuzzy Neural Network Design Based On Modern Optimization Technique

As the merging of fuzzy system and neural network, fuzzy neural network (FNN)compensates for the shortcoming of lack of learning capabilities of the pure fuzzy system, besides, it makes the neural network with "black box" attribute transparent and interpretable. Because of capitalizing on the strengths of fuzzy system and neural network, during the past couple of years, FNN has emerged as one of the most active and fruitful way in the fields of intelligent control, prediction of time series, biomedicine, modeling, data mining and so on.The main research of FNN is to approximate a process of fuzzy inference through the structure of neural network, namely, it is to build a standard neural network, which is used to extract fuzzy rules with the learning capability of neural network so as to accomplish various tasks. So the problems we are often confronted in FNN modeling is the optimal design of the structure and parameters of FNN, that is how to appropriately decide the number of fuzzy rales and precisely define the parameters of each rule so that it can effectively implements the fuzzy input, fuzzy reasoning, the propagation in network and the interpretation of final results.Begin with the methods such as Genetic Algorithm and Simulated Annealing in 1980s, modern optimization technique mainly used for NP-hard problems in optimization problems. For that these methods have no use for the continuity and differentiability of the objective functions, they have strong adaptability for the uncertain data and the traits of agility, strong intuitiveness and randomness, they have been put into applications broadly in a short time so as to become active approach for solving optimization problems.Tabu Search (TS) algorithm is a meta-heuristic algorithm. TS can avoid circuit searching by using the flexible memory mechanism and respective tabu criteria. Also according to the aspiration criteria, TS can assoil some good solution status which is tabu, in doing so it can ensure the diversification search and obtain the global optimum.Besides, another new modern optimization technique named Particle Swarm Optimization (PSO) is proposed by Kennedy and Evberhart in 1995. As a stochastic optimization techniquebased on swarm intelligence, PSO is initially used for continuous space.Because TS and PSO are the two typical modern optimization techniques based on individual intelligence and swarm intelligence respectively, and the problem of optimal design for FNN can be transformed into a combinatorial optimization problem. This paper has mainly finished the following three research jobs based on the analysis of this two algorithms:(1) Proposed a hybrid TS algorithm based on fuzzy neural network, which was investigated for automatic generation of fuzzy rules by optimizing the structure and parameters of FNN simultaneously.(2) Aiming at the usual method for deciding the initial structure and parameters of FNN depends on expert knowledge, proposed a automatic design algorithm of FNN classifier based on TS, which is used for classification problem in data mining.(3) For that the local search ability and accuracy of PSO isn't good, we proposed a hybrid PSO based on adaptive local search, which add the accurate local search operator to the strong global search ability of PSO. The hybrid algorithm is used as the leaning algorithm of FNN parameters, and further its performance is validated on many famous benchmark functions.From the results of the researches given in this paper, combining the FNN and modern optimization technique can achieve excellent performance. The design for FNN based on TS can be used for modeling when the system is complex and is uneasy to summarize expert knowledge. Hybrid PSO based on adaptive local search have strong global search ability while improving the convergence speed and accuracy so that it can be used in various fields.