Adaptive neuro-fuzzy inference system for modelling and control

A new approach for an adaptive neuro-fuzzy inference system for modeling and control is proposed. This approach uses a general regression neural network with a different learning capability from the classical clustering algorithm normally used by this specific network. The antecedent parameters of the regression network are obtained through an iterative grid partition process instead of the usual gradient descent algorithm or the classical grid partition method in the literature of neural network modeling. The membership functions used in the antecedent part are asymmetric and with varying shapes (triangles, gaussian, trapezoidal, etc) which is less common in the fuzzy modeling literature. The consequent parameters are obtained using the least squares estimates algorithm. In the simulation, the adaptive neuro-fuzzy inference system architecture is used to model a nonlinear function and to control the motion of a helicopter in the hover flight mode with promising results.