A framework for data-driven enhancement of the TSK fuzzy neural network

Abstract

Takagi-Sugeno-Kang fuzzy neural networks (TSKFNNs) are powerful machine learning tools for modelling regression problems; in particular when modelling vague or imprecise information. However, the effective application of such models is often reliant on expert knowledge of the domain area that may not always be available, may contain human bias/error, and is challenging humans to articulate mathematically. Therefore novel, data-driven approaches to TSKFNN construction and operation are required to address this limitation. This research proposes a series of data-driven improvements to the state-of-the-art TSKFNN in an incremental way including the fuzzification process, the rule base generation, and rule aggregation. Ensemble learning has inspired a novel data- driven rule aggregation method that enables increased performance and generalizability compared with the existing state-of-the-art TSKFNNs. Additionally, the existing TSKFNNs suffer from the curse of dimensionality with rule bases growing exponentially with an increasing number of input features. Inspired by random forest models and pruning, a data-driven rule reduction method is proposed and evaluated relative to the state-of-the-art TSKFNN. Finally, an efficient and data-driven initialization algorithm to initialize the fuzzy membership functions is developed. The proposed models are evaluated relative to a wide range of datasets and show improved accuracy compared to existing models whilst minimizing the need for human-defined parameters with a TSKFNN. The findings of this research provide the foundation for future studies into data-driven fuzzy neural networks and the opportunity to model large datasets without the concern of the curse of dimensionality.

Thesis is embargoed until 31 July 2027.

Date of Award	Jul 2024
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	Queen's University & China Scholarship Council
Supervisor	Desmond Greer (Supervisor) & Richard Gault (Supervisor)