臺灣博碩士論文加值系統

以模糊聚類分析法最佳化模糊系統及其應用

模糊系統中,模糊歸屬函數及模糊規則庫為決定模糊推論系統性能的兩大
主因,其設計之優劣對系統性能影響至深且巨.設計者往往要借助嘗試與錯
誤(Trial and Error)的經驗法則以試出較佳的規畫方式,不僅浪費時間與
成本,而且亦不知所調整出來的歸屬函數是否為最佳化的結果.近年來雖有
許多研究學者提出法則試圖解決上述問題,但都有若干缺點與不完備之處.

針對此一問題,乃於本文中提出以修正型模糊均值法(Modified Fuzzy C-
Means, MFCM)作為模糊系統中模糊歸屬函數的初步規畫,找出其最佳設置
數與相關參數,再以神經網路的倒傳遞學習法則微調模糊歸屬函數,以解決
規畫模糊系統所遭遇到的困難.本文並以sinc函數,gaussian函數等非線性
函數建模問題以及蝴蝶花分類問題,作為測試所提法則的最佳化能力.最後
再以汽車防撞系統中的煞車曲線及閃躲式防撞曲線作為模擬試驗,以印證
其對於解決實際問題的有效性.藉由與論域均分法(Equalized Universe
Method)和SCM(Subtrative Clustering Method)等不同方法的比較,可以
充份地說明MFCM法則的可靠性.本研究成果將有助於解決模糊歸屬函數規
畫時所面臨的瓶頸,並對於最佳化模糊系統整體性能有較佳的貢獻.

(關鍵字:模糊系統,模糊規則庫,模糊歸屬函數,模糊均值法,類神經網路,
倒傳遞學習,建模)

Optimization of Fuzzy System by Fuzzy Clustering
Analysis

ABSTRACT Fuzzy rule base and
fuzzy membership functions(MFs) are two major factors in
deciding the performance of fuzzy inference system. Therefore,
the design plays an important role for the performance stated
above. Trial and error was usually the way to solution, which
was not only costly and time-consuming but also promised no
optimized result. In recent years, many papers were presented
about this topic, but none of them has perfect answer.

To attack the above problems, we propose the Modified Fuzzy C-
Means Method(MFCM) for tuning the parameters of MFs. Then, we
fine-tune the MFs with backpropagation learning method.

MFCM will be examed for modeling with highly complicated
nonlinear functions, such as sinc function and gaussian
function, and pattern classification. Finally,there is a
simulation test of anti-collision driving system, including
first kind of trajectory, second kind of trajectory and evading
trajectory of anti-collision driving system, to prove MFCM is
suitable for the real world application. The results are quite
impressive compared with other approaches such as equalized
universe methods(EUM) and subtractive methods(SCM) and show the
efficacy of MFCM. Via the MFCM, the bottleneck to be overcomed
while designing MFs and the fuzzy system is optimized and has
better performance. (Key words: Fuzzy System, Fuzzy
Rule Base, FUzzy Membership Function, Fuzzy C-Means, Neural
Networks, Backpropagation, Modeling.)