An evolving Mamdani-Takagi-Sugeno based neural-fuzzy inference system with improved interpretability-accuracy

This paper presents a novel neural-fuzzy network architecture named the evolving Mamdani-Takagi-Sugeno neural fuzzy inference system (eMTSFIS) that addresses two deficiencies faced by neural-fuzzy systems. Firstly, the dynamic nature of real-world problems demands that neural-fuzzy systems be able to adapt their parameters and evolve their rule-bases to address the time-varying characteristics of their operating environments. Secondly, in practice, having good fuzzy rule-base interpretability and high modeling accuracy are contradictory requirements and one usually prevails over the other based on the modeling objective and fuzzy rule structure employed. The proposed eMTSFIS model is able to achieve life-long learning as it evolves and adapts its knowledge to the dynamics of the underlying environment. This effectively addresses the stability-plasticity dilemma. Also, the proposed eMTSFIS model combines Mamdani and T-S fuzzy modeling approaches, coupled with a localized parameter learning approach, to achieve both improved interpretability and accuracy. Experimental results from two benchmark applications demonstrate the learning robustness and modeling versatility of the proposed eMTSFIS model. The results are encouraging.