Individual And Ensemble Pattern Classification Models Using Enhanced Fuzzy Min-Max Neural Networks

Pattern classification is one of the major components for the design and development of a computerized pattern recognition system. Focused on computational intelligence models, this thesis describes in-depth investigations on two possible directions to design robust and flexible pattern classificati...

全面介紹

Saved in:
書目詳細資料
主要作者: F. M., Mohammed
格式: Thesis
語言:English
出版: 2014
主題:
在線閱讀:http://eprints.usm.my/46164/1/Mohammed%20F.%20M.24.pdf
標簽: 添加標簽
沒有標簽, 成為第一個標記此記錄!
實物特徵
總結:Pattern classification is one of the major components for the design and development of a computerized pattern recognition system. Focused on computational intelligence models, this thesis describes in-depth investigations on two possible directions to design robust and flexible pattern classification models with high performance. Firstly is by enhancing the learning algorithm of a neural-fuzzy network; and secondly by devising an ensemble model to combine the predictions from multiple neural-fuzzy networks using an agent-based framework. Owing to a number of salient features which include the ability of learning incrementally and establishing nonlinear decision boundary with hyperboxes, the Fuzzy Min-Max (FMM) network is selected as the backbone for designing useful and usable pattern classification models in this research. Two enhanced FMM variants, i.e. EFMM and EFMM2, are proposed to address a number of limitations in the original FMM learning algorithm. In EFMM, three heuristic rules are introduced to improve the hyperbox expansion, overlap test, and contraction processes. The network complexity and noise tolerance issues are undertaken in EFMM2. In addition, an agent-based framework is capitalized as a robust ensemble model to house multiple EFMM-based networks. A useful trust measurement method known as Certified Belief in Strength (CBS) is developed and incorporated into the ensemble model for exploiting the predictive performances of different EFMM-based networks.