Fuzzy membership function design: An adaptive neuro-fuzzy inference system (ANFIS) based approach

Abstract

In recent times, neuro-fuzzy approaches have been widely used for solving real-world complex problems. Adaptive neuro-fuzzy inference system (ANFIS) is one of the prominent models in this field of soft computing. This model combines both the fuzzy and neural approaches to represent non-linear and ill-defined problems more preciously. However, membership function design is an important factor of fuzzy-based system design. The number and shape of membership functions affect the performance as well as the computational cost of the fuzzy system. The main objective of this research is to find the suitable type of membership function for fuzzy model using ANFIS approach. In this experiment, several types of major membership functions- triangular, trapezoidal, generalized bell, and gaussian are analyzed with an effective dataset. Backpropagation and hybrid optimization techniques are used for training purposes with different numbers of epochs. In addition to that, RMSE is measured for both linear and constant type membership functions to evaluate the performance. Our experimental result shows that the trapezoidal membership function outperforms other types of membership function in all aspects for designing a fuzzy-based model which can be used in real-life applications.