調控式的部份最小平方法之研究

Abstract

本論文的目的在於建構一種分析法則，在未經處理的原始資料去除不必要的隱藏訊息。此新的學習法則稱之調控式的部份最小平方法，是合併部份最小平方法和規律法的優點，即使在雜訊的資料下，可避免過度配適的現象，得到較好的估算結果。 在模擬數據分析部份，調控式部份最小平方法用來分析三種不同的波型，並以均方根誤差做為判定的標準說明調控式部份最小平方法可得到較好的結果;實際的測量數據分析部份，利用實際的聲音檔案以及血糖濃度的光譜資料來驗證所提出的調控式部份最小平方法的確具備去除雜訊能力。

The main purpose of this thesis is to develop a method of analyzing and reducing the unseen or noisy information from the source data without preprocessing. Here presents a novel learning algorithm—partial regularized least squares (PRLS). It combines the advantages of both the partial least squares (PLS) and regularization technique to provide an efficient procedure to avoid the circumstance of overfitting and attain better results when calibrating under noisy data. In the simulated experiments, PRLS is applied to analyze the three different kinds of simulated waves. According to estimated standard of root mean square error, proving that PRLS has better performance than PLS. In real calibrated experiments, demonstrating PRLS certainly has the ability of noise reduction.