新穎區間第二型模糊類神經系統之設計與應用

A Novel Interval Type-2 Fuzzy Neural System: Design and Its Applications

張奉宇(Feng-Yu Chang)

指導教授 ： 李慶鴻

元智大學/電機通訊學院/電機工程學系/碩士(2011年)

https://doi.org/10.6838/YZU.2011.00314

摘要

本論文提出兩個新穎區間第二型模糊類神經系統，分別為IT2RFNS-A (Interval type-2 recurrent fuzzy neural system with asymmetric membership function)以及IT2TFNS (interval type-2 TSK fuzzy neural system) ，兩系統均使用區間第二型三角非對稱模糊歸屬函數以增強其性能及降低計算量；透過設計之穩定同步擾動隨機近似演算法(stable simultaneous perturbation stochastic approximation algorithm)訓練網路，並應用於非線性系統。
在IT2RFNS-A中，我們加上內嵌的遞迴層使其具備動態的特性。然而，IT2RFNS-A的降階運算缺乏對於不確定性的量測。因此，在本論文提出的IT2TFNS使用不確定性邊界(uncertainty bound)設計降階運算層，進而降低計算複雜度。而做為訓練類神經系統的穩定同步擾動隨機近似演算法是利用目標函數值去組合梯度資訊的一種演算法，如此可省去計算梯度的困擾。同時，本文使用李亞普諾夫穩定性理論設計參數更新法則，除了保證系統之穩定以及效率外，亦提供最佳學習步伐之選擇。此外，我們也改善穩定同步擾動隨機近似演算法無法完成即時控制之缺點，透過順滑平面(sliding surface)機制設計一適應性即時控制器。本文所提出之方法將應用於非線性系統鑑別與即時控制驗證系統之可行性及效能。最後，我們設計一基於IT2TFNS之字元辨識系統及其數位信號處理(DSP)硬體實現。

關鍵字

區間第二型模糊類神經系統 ； 非對稱歸屬函數 ； 不確定性邊界 ； 同步擾動隨機近似演算法 ； 李亞普諾夫定理 ； 即時控制 ； 字元辨識

並列摘要

In this thesis, we propose two novel interval type-2 fuzzy neural systems: interval type-2 recurrent fuzzy neural system with asymmetric membership function (IT2RFNS-A) and interval type-2 TSK fuzzy neural system (IT2TFNS), and their training scheme via stable simultaneous perturbation stochastic approximation (SPSA) algorithm. Herein, we adopt interval type-2 triangular asymmetric fuzzy membership functions (IT2 triangular AFMFs) for the proposed two interval type-2 fuzzy neural systems to improve the performance and efficiency.
In IT2RFNS-A system, an embedded feedback layer is attached to extend the abilities to include dynamic problems. In addition, the type-reduction operation of IT2RFNS-A cannot provide the measurement of uncertainty. Therefore, the IT2TFNS is introduced to reduce computational cost which utilizes of the uncertainty bounds for the type-reduction operation. For training these two novel interval type-2 fuzzy neural systems, we propose a stable SPSA algorithm that only the measurements of objective function are needed to form the gradient information. Meanwhile, we employ the Lyapunov stability analysis to derive a time-variant optimal learning step length for guaranteeing the stability of the system and ensuring the efficient training. In addition, we also develop an on-line control method by using the proposed stable SPSA algorithm. Finally, we propose a fuzzy logic-based optical character recognition system using IT2TFNS and its digital-signal-processing (DSP) hardware implementation. Several simulations including nonlinear system identification and on-line control, and experimental result are done to illustrate the feasibility and the effectiveness of the proposed method.

並列關鍵字

uncertainty bounds ； simultaneous perturbation stochastic approximation algorithm ； Lyapunov theorem ； on-line control ； optical character recognition ； interval type-2 fuzzy neural system ； asymmetric fuzzy membership function

參考文獻

[33] P. Z. Lin and T. T. Lee, “Robust self-organizing fuzzy-neural control using asymmetric Gaussian membership functions,” International Journal of Fuzzy Systems, vol. 9, no. 2, pp. 77-86, 2007.

[66] 胡子偉，結合區間第二型模糊非對稱歸屬函數及遞迴類神經網路系統之研究與應用，元智大學電機工程學系，碩士論文，2008。

[67] 潘泓毅，非對稱歸屬函數之第二型模糊類神經網路分析及其應用，元智大學電機工程學系，碩士論文，2006。

[1] J. F. Baldwin and S. B. Karake, “Asymmetric triangular fuzzy sets for classification models,” Lecture Notes in Artificial Intelligence, vol. 2773, pp. 364-370, 2003.

[3] O. Castillo, N. Cazarez, and P. Melin, “Design of stable type-2 fuzzy Logic controller based on a fuzzy Lyapunov approach,” IEEE International Conference on Fuzzy Systems, pp. 2331-2336, 2006.