Class Incremental Fuzzy Genetics-based Machine Learning Method

Abstract

Fuzzy Genetics-based Machine Learning (GBML) is a well-known method to design fuzzy classifiers which have high explainability. Generally, fuzzy GBML is originally designed as a batch learning. However, when focusing on IoT or data streaming, new information is generated continuously and it can be assumed that untrained classes are added after the classifier is designed based on the data at hand. In this study, we extend fuzzy GBML to learn untrained classes continuously. The proposed method enables continuous learning by discarding useless rules for the classification to generate new rules for untrained class patterns. Then, the trained patterns are reduced through a clustering algorithm in order to mitigate the increase in the computational cost. Experimental results show our method can efficiently learn untrained classes in terms of computational cost.