EFS Approaches for Regression and Classification

Abstract

This chapter deals with an analytical description and comparison of several well-known evolving fuzzy systems approaches. Even though most of these methods were originally designed for regression problems, some of these were also extended to classification tasks. The list of methods we are dealing with includes

• the FLEXFIS family: FLEXFIS = FLEXible Fuzzy Inference Systems for regression and FLEXFIS-Class = FLEXible Fuzzy Inference Systems for Classification applied on two different classifier architectures plus spin-off eVQ-Class as rule-based classifier in the cluster space (Section 3.1)

• the eTS family: eTS = evolving Takagi-Sugeno fuzzy systems and eClass = evolving fuzzy Classifiers coming in four different variants (Section 3.2)

• ePL = evolving Participatory Learning (Section 3.3)

• DENFIS = Dynamic Evolving Neuro-Fuzzy Inference Systems (Section 3.4)

• SOFNN = Self-Organizing Fuzzy Neural Network (Section 3.5)

• SAFIS = Sequential Adaptive Fuzzy Inference System (Section 3.6)

• other approaches such as SONFIN, GD-FNN, ENFRN, SEIT2FNN and EFP (Section 3.7)

Most of these approaches rely on some methods demonstrated in Chapter 2 for calculating statistical measures and (non-)linear parameters in incremental manner. However, each one of these exploits apply different mechanisms for rule evolution and incremental learning of antecedent parameters and also combine these differently with the learning scheme of the consequent parts. In this sense, we will give a detailed description of the specific characteristics of the various methods.