Curvature-based sparse rule base generation for fuzzy rule interpolation

Tan, Yao (2020) Curvature-based sparse rule base generation for fuzzy rule interpolation. Doctoral thesis, Northumbria University.

[img]
Preview
Text (Doctoral Thesis)
Tan.Yao_phd.pdf - Submitted Version

Download (6MB) | Preview

Abstract

Fuzzy logic has been successfully widely utilised in many real-world applications. The most common application of fuzzy logic is the rule-based fuzzy inference system, which is composed of mainly two parts including an inference engine and a fuzzy rule base. Conventional fuzzy inference systems always require a rule base that fully covers the entire problem domain (i.e., a dense rule base). Fuzzy rule interpolation (FRI) makes inference possible with sparse rule bases which may not cover some parts of the problem domain (i.e., a sparse rule base). In addition to extending the applicability of fuzzy inference systems, fuzzy interpolation can also be used to reduce system complexity for over-complex fuzzy inference systems. There are typically two methods to generate fuzzy rule bases, i.e., the knowledge driven and data-driven approaches. Almost all of these approaches only target dense rule bases for conventional fuzzy inference systems. The knowledge-driven methods may be negatively affected by the limited availability of expert knowledge and expert knowledge may be subjective, whilst redundancy often exists in fuzzy rule-based models that are acquired from numerical data. Note that various rule base reduction approaches have been proposed, but they are all based on certain similarity measures and are likely to cause performance deterioration along with the size reduction.
This project, for the first time, innovatively applies curvature values to distinguish important features and instances in a dataset, to support the construction of a neat and concise sparse rule base for fuzzy rule interpolation. In addition to working in a three-dimensional problem space, the work also extends the natural three-dimensional curvature calculation to problems with high dimensions, which greatly broadens the applicability of the proposed approach. As a result, the proposed approach alleviates the ‘curse of dimensionality’ and helps to reduce the computational cost for fuzzy inference systems. The proposed approach has been validated and evaluated by three real-world applications. The experimental results demonstrate that the proposed approach is able to generate sparse rule bases with less rules but resulting in better performance, which confirms the power of the proposed system. In addition to fuzzy rule interpolation, the proposed curvature-based approach can also be readily used as a general feature selection tool to work with other machine learning approaches, such as classifiers.

Item Type: Thesis (Doctoral)
Uncontrolled Keywords: Rule base reduction, feature selection, instance selection, curvature values, high-dimensional
Subjects: G400 Computer Science
G600 Software Engineering
Department: Faculties > Engineering and Environment > Computer and Information Sciences
University Services > Graduate School > Doctor of Philosophy
Depositing User: John Coen
Date Deposited: 11 Mar 2020 09:58
Last Modified: 11 Mar 2020 10:01
URI: http://nrl.northumbria.ac.uk/id/eprint/42439

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year

View more statistics