The Identification Of Parameter And Structure Optimization Of A Kind Of Fuzzy Neural Network

The fuzzy neural network which combines with characteristics of both fuzzy systems and artificial neural network can handle the linguist information and possess the ability of self-learning. It has been widely applied in modeling and control of complex system. Current researches on fuzzy neural network mainly include determining the structure of network, establishing parameter identification method and optimizing the structure of network. In this paper, parameters identification and structure optimization of fuzzy neural network based non-polynomial consequent and polynomial consequent are respectively investigated. The main contents of this paper are shown as follow.Firstly, the paper presents a type of parameters identification algorithm for fuzzy wavelet neural network (FWNN) based on extreme learning machine (ELM). The consequent part of fuzzy neural network is constructed by wavelet function. The linear parameters of the network is identified through the ELM algorithm while the nonlinear parameters is updating by gradient decent algorithm. Besides, multi-class modeling strategy is proposed to optimize the network structure. By partitioning the samples and identifying the parameters in each local region, the approximation accuracy of fuzzy neural network can been improved. The simulations indicate that the proposed fuzzy wavelet neural network model can approximate the time series, nonlinear function and UCI data sets with high accuracy.Secondly, this article applies generalized Bernstein polynomial as the consequent part of inference rules, to construct fuzzy Bernstein neural network. The K-means clustering algorithm is use to identify the parameters of the membership function. The network connection weights are calculated by partial least square algorithm while principal component analyses are adopted in the process of calculation, which can improve the approximation accuracy. By applying fuzzy Bernstein neural network to nonlinear dynamical system modeling, simulations results demonstrate the effectiveness of the proposed model.Lastly, this paper investigates the interval-valued time series forecasting based on the fuzzy Bernstein neural network. Firstly, getting the mid-point and range of the interval value time series, then fuzzy bernstein neural network is used to forecast the values of these two time series model. Accordingly, the prediction of interval-valued time series can be realized. The experiments results show that fuzzy bernstein neural network can effectively deal with the modeling problem of the interval-valued time series.