摘 要
自適應共振理論2 (ART2) 屬於一種非監督式學習的類神經網路，它具備了分群的穩定性與可塑性之特性，可以在不需預先設定群數大小，而能夠快速學習到新的群集特性。但ART2分群品質的好壞，決定於此網路架構下的相關內部參數設定。在過去文獻中通常是以試誤法，反覆地實驗來找出最佳的內部參數。如此，造成實務上秏費太多時間，而且亦無法滿足線上即時地動態搜尋最佳內部參數。有鑑於上述的問題，本研究提出新ART2演算法與其參數尋優機制，適用於半導體或TFT-LCD製程機台的資料特性，藉以改善生產資料的分群效益，而且可應用至先進虛擬量測系統裡的量測資料品質評估指標模組，精進此模組線上即時偵測異常量測資料的準確度。
新ART2演算法方面，在相似度比對過程中，增加了歐氏距離的檢查層，改善原ART2演算法僅比對樣本與群組間對應之向量角度相似度的不足。而在ART2參數尋優機制方面，利用連檢定及簡單移動平均法等理論確認資料特性的種類，依據樣本資料特性變化群組，自動搜尋出該資料集的最佳ART2內部參數設定。此外，為評估新ART2演算法分群的效益，本研究亦利用側影係數與平均誤差平方等方法判別分群結果的品質。最後，將新ART2演算法應用於量測資料品質評估指標的流程，以實際某TFT-LCD廠的PS-Height製程進行驗證，經由各個實驗案例得知此架構皆有良好的效果。

Abstract
Adaptive Resonance Theory 2 (ART2), an unsupervised neural network that solves the common stability-and-plasticity dilemma found in other clustering technique, which updates its model fast to the new data without specifying the number of clusters. However, the network’s clustering result is greatly influenced by the setting of the parameters (e.g. alert parameter, ρ). Most existing researches use trial-and-error method, which is time consuming and may not be feasible to achieve the approximately optimum combination of parameters or dynamically adjust the parameters to the real-time situation. To solve the aforementioned problems, in this research, New ART2 algorithm and its Automatic-Selection Scheme for optimum parameters are introduced. The New ART2 algorithm is capable in achieving better clustering validity, especially to dataset of semiconductor or TFT-LCD industries. Also, it can be applied to the Advance Virtual Metrology System’s metrology data quality evaluation (DQIy) scheme, which improves the effectiveness in detecting metrology data abnormality.
Apart from the cosine similarity measure of the classical ART2 algorithm’s, the New ART2 algorithm add Euclidean Distance Check for double checking the similarity between input vector and patterns. Accompany with New ART2 algorithm is the Automatic Selection Scheme for ART2 Optimum Parameters, which firstly utilizes methods such as Run Test and Simply Weighted Moving Average Approaches to predefine the patterns according to the variation and shifting of the process data, and then automatically search for the optimum combination of parameters. In addition, Silhouette Coefficient and Mean Square Error are applied to evaluate the clustering validity of the New ART2 algorithm. Finally, the New ART2 algorithm is employed to DQIy, and is evaluated with real PS-Height data from TFT-LCD industries. Experiment results show that better performance of DQIy is achieved with the new ART algorithm proposed.

參考文獻
[1] Y.-J. Chang, Y. Kang, C.-L. Hsu, C.-T. Chang, and T. Y. Chan, “Virtual Metrology Technique for Semiconductor Manufacturing,” in Proc. 2006 International Joint Conference on Neural Networks (IJCNN’06), pp. 5289-5293, July 2006.
[2] M.-H. Hung, T.-H. Lin, F.-T. Cheng, and R.-C. Lin, “A Novel Virtual Metrology Scheme for Predicting CVD Thickness in Semiconductor Manufacturing,” IEEE/ASME Transactions on Mechatronics, vol. 12, no. 3, June 2007.
[3] Y.-T. Huang, H.-C. Huang, F.-T. Cheng, and T.-S. Liao, and F.-C. Chang, “Automatic Virtual Metrology System Design and Implementation,” in Proc. IEEE International Conference on Automation Science and Engineering, Washington, D.C., U.S.A., August 2008.
[4] G. A. Carpenter, and S. Grossberg, “ART 2: Self-Organization of Sable Category Recognition Codes for Analog Input Patterns,” Applied Optics, vol.26, no.12, pp.4919-4930, Dec 1987.
[5] J. R. Whiteley, J. F. Davis, A. Mehrotra, and S. C. Ahalt, “Observations and Problems Applying ART2 for Dynamic Sensor Pattern Interpretation,” IEEE Transaction on System Man And Cybernetics - Part A: System And Humans, vol. 26, no. 4, pp.423-437, July 1996.
[6] L. Li, B. Zhang, and Y. Che, “The Improved Algorithm of ART2 in Data Mining,” in Proc. 2009 First International Workshop on Database Technology and Applications, pp.177-180, 2009.
[7] M. Halkidi, , Y. Batistakis, and M. Vazirgiannis, “Clustering validity checking methods: part II,” ACM SIGMOD Record, vol. 31 n.3, Sept, 2002.
[8] D. L. Davies, and D. W. Bouldin, “A cluster separation measure,” IEEE Transaction Pattern Analog Machine Intelligence, vol. 1, pp.224–227, 1979.
[9] P.J. Rousseeuw, “Silhouettes: a graphical aid to the interpretation and validation of cluster analysis,” Journal of Computational and Applied Mathematics, vol. 20, pp. 53-65, 1987.
[10] Cliffs, “Fundamentals of Neural Networks: Architectures, Algorithms, And Applications,” Chapter 5, 1994.
[11] 陳鵬帆「以自適應共振理論網路為基礎建構彩色濾光片為觀瑕疵辨識系統實作」，國立成功大學工業工程與資訊管理學系碩士學位論文，2005。
[12] 陳順宇「應用統計學」，二版，華泰書局，2005。
[13] 張正賢「統計品質管理」，華泰書局，1995。
[14] V.Natalija, and C.C.Howard, “Vector Quantization of Images Using Modified Adaptive Resonance Algorithm for Hierarchical Clustering,” IEEE Transactions on Neural Networks, vol. 12, no. 5, Sept, 2001.
[15] M.Chen, A.A.Ghorbani, and V.C.Bhavsar, “Incremental Communication for Adaptive Resonance Theory Networks,” IEEE Transactions on Neural Networks, vol. 16, no. 1, January 2005.
[16] R. Xu, and D. Wunsch II, “Survey of Clustering Algorithms,” IEEE Transactions on Neural Networks, vol. 16, no. 1, May 2005.
[17] G. A. Carpenter, and S. Grossberg, “A Massively Parallel Architecture for A Self-Organizing Neural Pattern Recognition Machine,” Computer Vision Graphics Image Processing, vol.13, Issue 7, pp.37-54, July 1987.
[18] T. W. Anderson, “Asymptotic Theory for Principal Component Analysis,” Ann. Math. Statist, vol. 34, pp. 122–148, 1963.
[19] I. H. Witten, and E. Frank. “Data Mining: Practical Machine Learning Tools,” San Francisco, CA: Morgan Kaufman, 2005.
[20] P. –N Tan, M. Steinbach, and V. Kumar, “Introduction to Data Mining,” Boston, MA: Pearson Education, Inc., 2006.
[21] Bradley,“Distribution-Free Statistical Tests,” Chapter 12, 1968.
[22] 簡禎富「決策分析與管理」，雙葉書廊，2005。
[23] 陳順宇「多變亮分析」，四版，華泰書局，2005。