Boosting Fuzzy Rules for Nonlinear System Identification Through Unscented Kalman Filter

This paper presents a new hybrid methodology for learning Sugeno-type fuzzy models via subtractive clustering, Adaptive Boosting Regression (AdaBoostR) and Unscented Kalman Filter (UKF). The generated fuzzy models are used for modeling nonlinear benchmark processes. In the proposed procedure, first one fuzzy rule is generated by subtractive clustering algorithm from available data of a given nonlinear process. Then this fuzzy rule is considered as a base model and AdaBoostR is employed in order to combine some of the weak learners (i.e. rules). Parameters of a rule are coded as the state vector in UKF and then UKF is used for fine tuning of these parameters. Moreover, as the second proposed method, Linear Kalman Filter (LKF) is utilized for adjusting only the output membership functions parameters (first order sugeno's parameters) of base models (i.e. rules). Three case studies are considered for illustrating the applicability of our proposed boosting methods. Results apparently show the obtained fuzzy models are superior to Adaptive Neuro-Fuzzy Inference System (ANFIS) in terms of both modeling accuracy and computational requirements. Also the comparison results confirm that the obtained fuzzy models are well comparable with those of achieved by one of the powerful and recently developed fuzzy identification methods.