Self evolving Takagi-Sugeno-Kang fuzzy neural network.
Nguyen Ngoc Nam
Date of Issue2012
School of Computer Engineering
Centre for Computational Intelligence
Fuzzy neural networks is a popular combination in soft computing that unites the human-like reasoning style of fuzzy systems with the connectionist structure and learning ability of neural networks. There are two types of fuzzy neural networks, namely the Mamdani model, which is focused on interpretability, and the Takagi-Sugeno-Kang (TSK) model, which is focused on accuracy. The main advantage of the TSK-model over the Mamdani-model is its ability to achieve superior system modeling accuracy. TSK fuzzy neural networks are widely preferred over their Mamdani counterparts in dynamic and complex real-life problems that require high precision. This Thesis is mainly focused on addressing the existing problems of TSK fuzzy neural networks. Existing TSK models proposed in the literature can be broadly classified into three classes. Class I TSK models are essentially fuzzy systems that are unable to learn in an incremental manner. Class II TSK networks, on the other hand, are able to learn in incremental manner, but are generally constrained to time-invariant environments. In practice, many real-life problems are time-variant, in which the characteristics of the underlying data-generating processes might change over time. Class III TSK networks are referred to as evolving fuzzy systems. They adopt incremental learning approaches and attempt to solve time-variant problems. However, many evolving systems still encounter three critical issues; namely: 1) Their fuzzy rule base can only grow, 2) They do not consider the interpretability of the knowledge bases and 3) They cannot give accurate solutions when solving complex time-variant data sets that exhibit drift and shift behaviors. In this Thesis, a generic self-evolving Takagi–Sugeno–Kang fuzzy framework (GSETSK) is proposed to overcome the above-listed deficiencies of existing TSK networks with the following contributions: A novel fuzzy clustering algorithm known as Multidimensional-Scaling Growing Clustering (MSGC) is proposed to empower GSETSK with the incremental learning ability. MSGC also employs a novel merging approach to ensure a compact and interpretable knowledge base in the GSETSK framework. MSGC is inspired by human cognitive process models and it can work in fast-changing time-variant environments. To keep an up-to-date fuzzy rule base when dealing with time-variant problems, a novel ‘gradual’-forgetting-based rule pruning approach is proposed to unlearn outdated data by deleting obsolete rules. It adopts the Hebbian learning mechanism behind the long-term potentiation phenomenon in the brain. It can detect the drift and shift behaviors in time-variant problems and give accurate solutions for such problems.