Adaptive self organizing feature map neuro-fuzzy technique for dynamic system identification

Conventional fuzzy system design involves finding optimal numbers, shapes and fine tuning of membership functions in the input and output universes of discourse. One way to circumvent a trial-and-error approach is to use expert knowledge. When expert knowledge is not available or is insufficient, this drawback can be overcome by combining the advantages of neural networks which have well established supervised and unsupervised learning algorithms available, and fuzzy systems. In this paper an approximate simplified fuzzy reasoning method is introduced. This method can automatically determine the optimal number and locations of pseudo antecedent and consequent rules on input and output universes of discourse for a given data driven fuzzy system. A self-learning algorithm for this proposed neuro-fuzzy architecture is developed and its applications to several dynamic system identification problems are presented