Recurrent neuro-fuzzy hybrid-learning approach to accurate system modeling

A recurrent neuro-fuzzy approach with RO-LSE hybrid learning algorithm to the problem of system modeling is proposed in the paper. The proposed recurrent neuro-fuzzy system possesses six layers of neural network to perform the fuzzy inference. The recurrent structure is formed using lagged membership-grade signals as internal feedbacks to the layer of membership functions of fuzzy sets, and it is expected having great potential to trace the temporal change of signals. Fuzzy sets with time-varying kernels have excellent property, with that the input-output mapping of the neuro-fuzzy system is no longer fixed but time varying. In this study, a new parameter learning approach is proposed for NFS with good learning convergence, in which the hybrid RO-LSE learning algorithm is utilized for the update of parameters. The well-known random optimization (RO) method is used to update the parameters of the premise parts of the proposed system, and the method of least square estimation (LSE) to update those of the consequent parts. The hybrid algorithm is found useful, and it has shown fast convergence of parameter learning for the proposed system. Three examples are used to demonstrate the brilliancy of the proposed approach. Excellent performance of the proposed approach in modeling accuracy and learning convergence is observed.