限制型利基柏拉圖基因演算法及其於製程參數最佳化之應用

Abstract

特徵選取的目的在於將不重要的特徵消除，保留有用的特徵，使分析過程更為迅速，分析結果更為可靠。利基柏拉圖基因演算法是一種針對多目標最佳化問題所發展出來的工具，可找到多個目標之間的折衷最佳解，以供分析人員選擇，而這些折衷最佳解所形成之理想曲線，稱為柏拉圖最佳邊際。當此方法使用於分類問題之特徵選取上時，會發生在挑選過程中，基因演算法以減少使用特徵數為優先的狀況。此將導致最後所找到之柏拉圖最佳邊際發生特徵數刪減過多，且分類準確率沒有明顯提升的狀況，此問題在原始的總特徵數越多時，會越加明顯。由於在實際問題中，提升分類準確率，常常會比大幅刪減使用特徵數來得重要，本研究提出限制型利基柏拉圖基因演算法，透過在基因演算法的挑選過程中，增加一準確率限制的手段，調整基因演算法之搜尋方向，以避免特徵的過度刪減，並確保分類準確率的提升。 本研究使用了三筆數值資料，比較原始的利基柏拉圖基因演算法與限制型利基柏拉圖基因演算法之差別。結果顯示限制型利基柏拉圖基因演算法的確能找到分類準確率較高的特徵組合，且不會發生特徵數刪減過多的問題。本研究亦將此方法應用於一DVD光碟片之製程問題上，找出製程中的重要參數；接著利用這些製程中的重要參數，建立製程關係模型，再針對此模型進行製程參數最佳化，達到提升品質、降低不良率的目的。

Feature selection is an important step to deal with classification problem. The essential effect of feature selection is to improve the accuracy and speed of classification systems. Niched Pareto Genetic Algorithm (NPGA) is a powerful method to solve optimization problems with multiple objectives. It can yield a diverse population of solutions among the multiple, conflicting objectives. As NPGA is applied to feature selection in the classification problem, it will show a preference to reduce the feature used, but not to improve the classification accuracy. This will cause the final Pareto optimal frontier perform feature subsets with many features erasing and less improvement of classification rate. In order to improve the accuracy of classification, it is necessary to avoid the over trimming of the features. In this study, we propose a constraint based niched Pareto genetic algorithm to adjust the thread of search. Three numerical examples are employed to demonstrate the difference between original and constraint based niched Pareto genetic algorithm in this study. We discuss the constraint can avoid genetic algorithm from erasing too many features and improving the classification accuracy. The proposed method is also applied to find key factors in a Digital Video Disk manufacturing system. The key factors will be used to help modeling and optimizing the manufacturing process parameters.