Please use this identifier to cite or link to this item:
Title: Data-based neuro-fuzzy systems
Authors: Tung, Sau Wai
Keywords: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
Issue Date: 2012
Source: Tung, S. W. (2012). Data-based neuro-fuzzy systems. Doctoral thesis, Nanyang Technological University, Singapore.
Abstract: Neuro-fuzzy systems are hybrid systems that possess the functionalities of the two individual systems, where a fuzzy IF-THEN model can be extracted from given numerical data using neural learning approaches. Depending on the numerical data presented, there are generally two approaches to the design of a neuro-fuzzy system; namely, design via batch learning, and design via sequential learning. Currently, the main problems dogging existing neuro-fuzzy systems of the former approach are: (1a) an inconsistent rulebase; (1b) the need for prior knowledge such as the number of clusters to be computed; (1c) heuristically designed knowledge acquisition methodologies; and (1d) the stability-plasticity trade-off of the system. Such weaknesses are directly related to the techniques employed for the fuzzy partitioning and fuzzy rulebase identification in the systems. On the other hand, existing neuro-fuzzy systems of the latter approach are commonly plagued by the following problems: (2a) susceptibility to noisy data; (2b) an ever-expanding knowledge structure; and (2c) a lack of interpretability of the system. Such drawbacks are the consequences of online learning in an uncertain and time-varying environment. This thesis consists of 2 sections. In the first section, the thesis addresses issues faced by batch learning neuro-fuzzy systems using a novel hybrid intelligent Self-adaptive Fuzzy Inference Network (SaFIN); while in the second section, the thesis focuses on improving the robustness of online learning in an uncertain and time-varying environment using a novel evolving Type-2 neural Fuzzy Inference System (eT2FIS). The proposed models address the problems with the following contributions: (A) A novel fuzzy partitioning technique named Categorical Learning Induced Partitioning (CLIP) is employed in SaFIN to obtain an initial fuzzy partitioning of the input and output spaces such that it is tailored to the numerical data presented. Inspired from the behavioral category learning process demonstrated in humans, the one-pass CLIP is able to incorporate new clusters in each input-output dimension when the existing clusters are not able to give a satisfactory representation of the incoming data. This not only omits the need for prior knowledge regarding the number of clusters needed for each input-output dimension, it also allows SaFIN a flexibility to incorporate new knowledge with old knowledge in the system. In addition, a self-automated rule formation mechanism proposed within SaFIN ensures that it obtains a consistent resultant rulebase. The SaFIN model is subsequently evaluated using a real-life application involving the modeling of highway traffic flow density. (B) Most existing online sequential learning neuro-fuzzy systems are vulnerable under noisy learning environment because they are Type-1 models. On the other hand, eT2FIS synergizes an online learning system with a Type-2 fuzzy model, which is better known for its noise resistance ability. In addition, a three-steps online learning paradigm is proposed in eT2FIS to directly curb the ever-expanding knowledge structure and improve the interpretability of the system: A new rule is created only when a newly arrived data is novel; and an obsolete rule is deleted when it is no long relevant to the learning environment; while highly over-lapping clusters in the input-output spaces are merged to reduce the computational complexity and improve the overall interpretability of the system. Subsequently, the eT2FIS model is used in the modeling of a real-life application involving the tracking of stock price movement.
DOI: 10.32657/10356/48902
Fulltext Permission: open
Fulltext Availability: With Fulltext
Appears in Collections:SCSE Theses

Files in This Item:
File Description SizeFormat 
SCEG0700008E.pdf3.32 MBAdobe PDFThumbnail

Page view(s) 20

Updated on Aug 2, 2021

Download(s) 20

Updated on Aug 2, 2021

Google ScholarTM




Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.