應用類神經網路於實驗計劃中遺失資料分析之研究

Abstract

在進行實驗收集資料時，可能會因為收集過程時間太長或其他不可抗拒的因素，導致所收集的資料有遺失的情形產生，而利用傳統的田口方法或實驗設計方法，並不能分析這些含有遺失資料的數據。所以便有許多學者發展出可以分析遺失資料的方法，如利用平均值代替遺失資料、直接忽略不看或是應用迴歸、MLE等方法。然而，利用平均值或是直接忽略不計，有時並不合理，其他像是迴歸或MLE等統計方法則需要繁複的計算或需具備統計知識，造成實務上使用的不便。所以本研究希望利用倒傳遞類神經網路，發展出一個有效且容易執行的分析遺失資料的方法。本研究提出了兩個方法，並利用三個案例來比較本研究所提出的方法和傳統方法的分析結果，藉此說明本研究所提方法的可行性。

The missing data can occur in an experiment owing to some unpredictable factors in the experimental process. Traditional Taguchi method and Design of Experiment cannot effectively analyze missing data, thereby many methods for analyzing missing data have been developed. For example, by using variable mean instead of missing data, directly ignoring missing data, using regression analysis and MLE, etc.. However, using variable mean or ignoring missing data is not reasonable. Other statistic methods, such as regression analysis and MLE, are computationally complicated and often difficult to explain to practitioners. In this thesis, two neural network-based procedures are developed, which are simpler than the traditional approaches and can deal with missing data efficiently. In addition, three numerical examples from previous literature are presented to compare the proposed procedures with the conventional approaches. The results reveal that proposed procedures are effective.