應用類神經網路於實驗計劃中受限資料分析之研究

Abstract

在進行實驗的時候，可能會因為設備或其他人為因素影響的緣故，導致部份實驗數據 無法取得，形成所謂的受限資料，當實驗中有受限資料的時候就無法以一般分析程序分析 。傳統上分析受限資料的方法大都利用高深的統計技巧，而且計算程序相當繁複，使得這 些既有的方法不論是在理解或是實際操作上都具有相當的困難度。此外，由於實驗資料並 不完全，傳統上有的方法將受限資料以受限點代替進行分析，這是不合理的。本研究利用 倒傳遞類神經網路，針對重覆性實驗的單一受限資料發展有效且容易執行的受限資料分析 方法。本研究提出兩個分析方法，它們都比傳統的受限資料分析方法（如最大概似估計法 與精密累積分析法）簡單得多而且更有效。本研究並利用三個案例，藉由比較本研究所提 方法與傳統方法分析結果，顯示本研究所提方法的有效性及可行性。

Censored data refers to incomplete data while performing experiments owing to equipment or factitious causes. In practice, although difficult statistical techniques are employed to analyze censored data, computational processes are extremely complex. In addition, incomplete data may inhibit the ability of some methods to analyze censored data instead of a censored point. In this thesis, we develop efficient and esaily implemented procedures based on the back-propagation neural network to analyze singly censored data in replicated experiments. Also proposed herein are two procedures, both of which are simpler and more efficient than conventional approaches such as maximum liklihood estimation (MLE) and minute accumulating analysis (MAA). In addition, three numerical examples are analyzed. Comparing the results of proposed procedures with those of conventional approaches demonstrates the reliability of the proposed procedures.