Please use this identifier to cite or link to this item:
|Title:||Sparse dynamic reorganising fuzzy neural networks||Authors:||Zhou, Jair Weigui||Keywords:||DRNTU::Engineering::Computer science and engineering||Issue Date:||2017||Source:||Zhou, J. W. (2017). Sparse dynamic reorganising fuzzy neural networks. Doctoral thesis, Nanyang Technological University, Singapore.||Abstract:||Fuzzy logic systems can broadly be grouped into two main types; namely: linguistic fuzzy systems (Mamdani) and precise fuzzy systems (Takagi-Sugeno-Kang). Fuzzy neural systems have been extensively researched and utilized in recent years because of their learning capability and interpretability in data modelling. Fuzzy rules are formed in fuzzy neural systems using either the batch learning approach (offline) or the sequential/evolving learning approach (online). Evolving fuzzy neural systems normally incorporate rule pruning mechanisms which refine the knowledge base to better represent the current trend in the observation, and also reduce the system complexity. In real life, many factors may contribute to uncertainty in data and these include unexpected movement in the data trend or noise corrupted data, etc. An unexpected trend movement in the data can occur in many different problem domain such as in the financial field (stock trading), the medical field (blood glucose regulation), or in nature (flash floods, earthquakes, tsunami, etc), and certain levels of noise interference during the data collection is inevitable. Traditional fuzzy systems or offline models require a relatively complete rule base to provide good modelling of the data. Additionally, this approach is not designed to change after the training process. Hence, the offline model is constrained by the boundaries of the fixed rule base which limits the flexibility of handling data drift. On the other hand, the online models often build up the rule base from scratch based on the data presented. As the size of rule grows incrementally, the problems of exploding rule base will be encountered. Rule pruning mechanism which removes rules that are of less significance was proposed. However such pruning mechanism may removes crucial, yet currently inactive, fuzzy rules with overlapping fuzzy sets from the rule base. Some evolving models had claimed to be able to handle concept drift problem. When new data is presented to the model, it will be able to inference a result and create new rules if rule creating criteria is met. However, in the case of concept shift problem where new data is significantly different from the past, poor inference result may be produced. Sparse rule base may be formed commonly in two conditions. One, due to the lack of expert knowledge in a specific problem domain, or two, when the fuzzy neural network with rule pruning ability removes overlapping fuzzy rules unintentionally, as mentioned above. The limited flexibility of the offline model to handle data drift, and deficient modelling ability of online model to handle data shift have encouraged the exploration of incorporating fuzzy interpolation and extrapolation technique (FIE) to both models. Fuzzy interpolation technique is able to infer a result in a sparse or an incomplete fuzzy rule base. On the other hand, fuzzy extrapolation technique is able to infer a result when new data is found beyond the existing coverage of the rule base. Such technique enables the fixed rule base of an offline model to infer a result when new data is beyond the boundary of existing rule base, and also enable an online model to infer a result either in a sparse rule base or when new data is too far beyond the coverage of existing rule base. In this thesis, a new fuzzy interpolation and extrapolation technique extended to the Takagi-Sugeno-Kang fuzzy inference system will be introduced. The incorporation of the proposed FIE technique into two existing fuzzy neural system, representing both offline and online models, will be presented. The results shall show how the incorporation of the proposed FIE technique handles a sparse rule base or concept drift/shift and produces better result as compared to the original extended model.||URI:||http://hdl.handle.net/10356/72395||DOI:||10.32657/10356/72395||Fulltext Permission:||open||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Theses|
Updated on Nov 29, 2020
Updated on Nov 29, 2020
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.