自我群體組織合作式學習演算法於模糊類神經網路設計及其於印刷電路板校準檢測的應用

Abstract

近來將演化式演算法用來訓練模糊類神經網路的參數已經變成一個相當受歡迎的研究領域，主要是因為演化式演算法具有在解空間平行搜尋的能力，以及更容易收斂至最佳解的特性。然而，現有的學習演算法確有容易陷入局部最佳解，以及必須事前決定許多關鍵性的參數的問題。本論文提出了自我群體組織合作式學習演算法來改善上述所面對的問題。自我群體組織合作式學習演算法利用網路參數學習和自我群體組織技術來達成自動化設計模糊類神經網路的參數。在網路參數學習中，它被用來決定網路中適當的參數，包含了兩階段自組織方法以及資料探勘為基礎的群體選擇方法。在傳統群體共生進化中，模糊規則數量必需事前設定，不適當的參數會導致系統效能的低落。因此，兩階段自組織方法主要用來決定適當的模糊規則數量；而資料探勘為基礎的群體選擇方法主要用來解決傳統隨機選取群體的方法，利用資料探勘技術來找到表現較佳的群體，進而選取出適當的群體進行演化。而在自我群體組織方面，用來分析網路架構以及決定適當的網路架構，可以解決在傳統群體共生進化演算法中，群體數量需事前設定的問題，自我群體組織技術會在整個學習過程中，依照目前系統表現，即時的調整合適的群體數量。此外，透過自我群體組織共生學習演算法所訓練的網路可被應用在印刷電路板校準檢測的現實世界的實例。實驗結果顯示，本論文所提出的演算法比起其它的方法，具有更好的效果。

Recently, the evolutionary algorithms are using for training parameters of neural fuzzy network has become a popular research field because evolutionary algorithms simultaneously evaluate many points in the search space and are more likely to converge toward the global solution. However, current learning algorithms have problems which easily fall into local optimal and many critical parameters must be predefined carefully. In this dissertation, a self group organization cooperative learning algorithm is proposed to improve above problems. The proposed algorithm uses network parameter learning and self group organization technique to automatically design neural fuzzy networks. The network parameter learning is used to determine the suitable parameters of the neural fuzzy networks efficiently. It consists of two phase self organization and data mining group selection. In the traditional group-based symbiotic evolution, the number of the fuzzy rules must be predefined carefully because the inappropriate number of the fuzzy rules would cause the lower performance. Therefore, the two phase self organization was applied to determine the suitable number of the fuzzy rules automatically and data mining group selection was applied to determine the suitable group while the traditional group-based symbiotic evolution selected randomly. The self group organization technique is used to analyze the network structure and determine suitable structure of the network. It can deal with the inappropriate number of the group in the traditional group-based symbiotic evolution and timely adjust the suitable number of the group according to the system performance. Besides, the trained network can be applied to the real world application which was the printed circuit board alignment inspection. The simulation results are shown the better performance of the proposed algorithm than the other methods.