Multi-Valued Neuro-Fuzzy Classifier for Extracting Rules from Real-World Data

2017 Fiscal Year Final Research Report

• Project/Area Number: 15K00333
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Soft computing
• Research Institution: University of Fukui
• Principal Investigator: Murase Kazuyuki 福井大学, 学術研究院工学系部門, 教授 (40174289)
• Research Collaborator: HATA Ryusuke 福井大学, 大学院工学研究科 ; M.A.H. Akhand Kuhlna University of Engineering and Technology, Professor ; Pintu Chandra Shill Kuhlna University of Engineering and Technology, Professor ; Md. Monirul Islam Bangladesh University of Engineering and Technology, Professor
• Project Period (FY): 2015-04-01 – 2018-03-31
• Keywords: ニューロファジィ / 多元数 / クラス分類 / 関数近似 / 深層学習 / オートエンコーダー / 畳み込みニューラルネットワーク

Outline of Final Research Achievements

When computer tells you what the object is in a photograph, it is essential to know how and why the computer determined it. In the Neuro-Fuzzy, the decision is made with a combination of rules. In Neuro-Fuzzy learning algorithm, in contrast, the rules are generated from a training set of the given data. With the conventional methods, however, a large number of rules are generated and it is hard to understand. In this study, by using complex numbers and quaternions, new algorithms were proposed to reduce the total number of parameters to be determined during training. In real-world benchmark data, the multi-valued versions of Neuro-Fuzzy with reduced number of parameters converged much faster exhibiting equivalent or better accuracy than the real-valued counterpart. Understanding rules will be thus easier for given problems.

Free Research Field

ソフトコンピューティング