A Hybrid Approach-based Recurrent Compensatory Neural Fuzzy Network

A recurrent compensatory neural fuzzy network based on a hybrid approach (HARCNFN) integrating a modified clustering method and the gradient descent method is proposed for the identification of dynamic nonlinear systems. With the recurrent nodes introduced in the second layer of the network, the recurrent compensatory neural fuzzy network (RCNFN) has the ability of dynamic mapping. The identification of the proposed network is composed of two phases: structure identification and parameter identification. In the first phase, a modified relational grade clustering (RGC) method is proposed to construct the initial fuzzy model of the RCNFN with five layers. In the second phase, the gradient descent method is used to tune the parameters of the network to obtain a more precise fuzzy model. Finally, the HARCNFN is applied for the identification of a dynamic nonlinear system. The simulations show that it is superior to the compensatory neural fuzzy network (CNFN) in modelling accuracy and convergence speed.