Adaptive training schema in Mamdani-type neuro-fuzzy models for data-analysis in dynamic system forecasting

This paper investigates the possibility of a pseudo-online adaptive training schema for Mamdani-type neuro-fuzzy models that have robust linguistic interpretability. As such verbatim models are incapable of complex constructs available to Takagi-Sugeno-type neuro-fuzzy models, a heuristic approach is developed to allow the rule bases to adapt accordingly to fundamental shifts in the characteristics of time-varying dynamic systems for the purpose of forecasting. Experimental results showed that the proposed model is capable of adapting its rule base over time, and uses a relatively fewer number of rules for generalization in dynamic systems.