一種以PSO為基礎之動態階層式模糊類神經網路-使用在未知結構之連續函數之學習

A PSO-Based Dynamic Hierarchical Fuzzy Neural Network for Learning Continuous Function with Unknown Structure

李樹昌(Shu-Chang Li)

指導教授 ： 王偉彥

國立臺北科技大學/機電學院/自動化科技研究所/碩士(2008年)

https://doi.org/10.6841/NTUT.2008.00496

摘要

大量調整參數的問題限制了模糊類神經網路(Fuzzy neural network, FNN)的適用性的它是一個”curse of dimensionality”，特別在未知的連續函數裡，我們使用一個方法處理這個問題，這個問題是建構在一個動態階層式模糊類神經網路(Dynamic Hierarchical Fuzzy Neural Network, HFNN)。在這篇論文中，我們討論一種兩階段的最佳化演算法，此架構聰明的去建構一個動態階層式類神經網路，基礎建立在合併式類神經網路(Merged-FNN)在未知架構的連續函數；第一階段，我們使用基因演算法(Genetic Algorithm, GA)，GA是一種普遍的演算法，這裡使用 GA在隨機族群(GA with Randomly Grouping, GA_RG)去建構HFNN。第二階段，使用簡化型基因演算法(reduced-form genetic algorithm, RGA)，RAG在HFNN透過GA_RG得到最佳化的架構，但在第二階段的RGA最佳化需要很長的運算時間，因此我們使用粒子群聚最佳化(Particle Swarm Optimization, PSO)來處理，PSO與GA很相似的演算法，這篇論文使用PSO後能減少運算時間，並利用PSO之特性，將GA建立在PSO的基礎上，將此合併運算，以達到減少運算時間之效果，在真實世界應用，這些方法這邊被使用在近似於台灣股票市場及人為操作的化學裝置上。

關鍵字

模糊類神經網路 ； 基因演算法 ； 粒子群集演算法 ； 階層式函數

並列摘要

A serious adjustable parameter problem limiting the applicability of the fuzzy neural networks is the “curse of dimensionality”, especially for unknown structure continuous functions. A way to deal with this problem is to construct a dynamic hierarchical fuzzy neural network. In this paper, we propose a two-stage optimization algorithm to intelligently construct the dynamic hierarchical fuzzy neural network (HFNN) based on the merged-FNN for unknown structure continuous functions. At the first stage, we use a genetic algorithm (GA) which is a popular optimization algorithm. In this paper, we use GA with Randomly Grouping (GA_RG) to construct the HFNN. At the second stage, a reduced-form genetic algorithm (RGA) optimizes the HFNN constructed by GA_RG. But in the second stage, the RGA optimizes operation demands a longer evolution. Accordingly, Particle Swarm Optimization (PSO) can addresses this problem. PSO is similar to a genetic algorithm, but it can reduce operation time, combining GA with PSO, operation time can be significantly reduced, for a real-world application. The presented method is used to approximate the Taiwanese stock market and the human operation at a chemical plant.

並列關鍵字

fuzzy neural network ； GA ； PSO ； and hierarchical function

參考文獻

Chinese part:

English part:

[8] O. Huwendiek, and W. Brockmann, “Function approximation with decomposed fuzzy systems,” Fuzzy Sets and Systems, vol. 101, 1999, pp. 273-283.

[9] Chi-Hsu Wang, Wei-Yen Wang, Tsu-Tian Lee, and Pao-Shum Tseng, "Fuzzy B-Spline Membership Function(BMF) and Its Applications in Fuzzy-Neural Control", IEEE Transactions On System and, Man, and Cybernetics, Vol. 25, No. 5, 1995, pp. 841-851.

[10] Wei-Yen Wang, and Yi-Hsum Li, “Evolutionary Learning of BMF Fuzzy-Neural Networks Using a Reduced-Form Genetic Algorithm”, IEEE ransactions On System and, Man, and Cybernetics-Part B: Cybernetics, Vol. 33, No. 6, 2003, pp.966-976.