A generalized framework for fuzzy neural architecture

• Main Author: Tung, Whye Loon.
• Other Authors: Quek, Hiok Chai
• Format: Theses and Dissertations
• Published: 2008
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: http://hdl.handle.net/10356/2505
• Institution: Nanyang Technological University

This thesis consists of 2 sections. A neural fuzzy (neuro-fuzzy) system/network is the neural implementation of a fuzzy system and is characterized by its fuzzy sets, IF-THEN fuzzy rules and the node operations responsible for its reasoning and decision-making capabilities. Generally, there exist two classes of neural fuzzy systems. Class I neural fuzzy systems such as ANFIS and ARIC have a predefined network structure and perform only parameter-learning. By comparison Class II neural fuzzy systems, which are the focus of this work and consist of the POPFNN, SOFIN, HyFIS, DENFIS and the Falcon-ART networks, are able to perform structural learning by automatically crafting the fuzzy rules from the numerical training data prior to the tuning of the fuzzy set parameters. Currently, the main problems dogging existing neural fuzzy systems/networks are: (1) Susceptibility to noisy training data; (2) the Stability-Plasticity dilemma; (3) The need of prior knowledge for structural learning; (4) An inconsistent fuzzy rule-base and (5) Heuristically-defmed node operations that lead to a poor interpretation of the reasoning process. Such weaknesses are directly related to the techniques employed for the derivation of the fuzzy sets and IF-THEN fuzzy rules (structural learning), the tuning of the fuzzy sets (parameter-learning) and the choice of fuzzy reasoning/inference scheme used to define the node operations of the systems.