A Locally Recurrent Fuzzy Neural Network With Support Vector Regression for Dynamic-System Modeling

This paper proposes a new recurrent model, known as the locally recurrent fuzzy neural network with support vector regression (LRFNN-SVR), that handles problems with temporal properties. Structurally, an LRFNN-SVR is a five-layered recurrent network. The recurrent structure in an LRFNN-SVR comes from locally feeding the firing strength of each fuzzy rule back to itself. The consequent layer in an LRFNN-SVR is a Takagi-Sugeno-Kang (T-S-K)-type consequent, which is a linear function of current states, regardless of system input and output delays. For the structure learning, a one-pass clustering algorithm clusters the input-training data and determines the number of network nodes in hidden layers. For the parameter learning, an iterative linear SVR algorithm is proposed to tune free parameters in the rule consequent part and feedback loops. The motivation for using SVR for parameter learning is to improve the LRFNN-SVR generalization ability. This paper demonstrates LRFNN-SVR capabilities by conducting simulations in dynamic system prediction and identification problems with noiseless and noisy data. In addition, this paper compares simulation results from the LRFNN-SVR with other recurrent fuzzy models.