Increasing predictive accuracy of neuro-fuzzy using quartiles to initialize the membership function

Abstract

© 2020 ACM. Neuro-Fuzzy is one of the techniques commonly used in classification as it has been proven satisfactory in predicting results. The neuro-fuzzy system is divided into 3 layers: the first layer is the fuzzification process, the special hidden layer is bipolar transform, and the output layer is the learning process. This research proposes the alternatives to the initialization of membership function (MF) for the Neuro-fuzzy system that maintains the properties of direct calculation and rule-based for classification. In the fuzzification phase, define membership functions in accordance with the characteristics of the data using five summarized values such as min, 1st quartile (Q1), median, 3rd quartile (Q3) and max to determine the initial position of MF to convert input data to fuzzy sets or linguistics variable. The Gaussian is used as MF and transforms members into bipolar in a special hidden layer to be sent to the final part of the artificial neural network. The 12 data sets with various classes and features are used in the experiment and give satisfactory results.