機器學習是近年來興起的一門熱門的研究領域,其主要的目的是從使用者及輸入資料等去獲得知識,以幫助使用者解決更多的問題,並減少錯誤,提高解決問題的效率。本論文的研究中將提出新的機器學習架構去處理含有雜訊與離異點的資料學習,並取得好預測結果。本論文我們提出三種方法並結合數位式華爾恕神經網路來分析含雜訊與離異點樣本資料,排除離異點,獲得其強健性。首先是使用最小平方裁減支持向量機回歸(LTS-SVMR) 做樣本資料的選取,在裁減前先使用SVMR評估樣本資料得到好的初始函數,再利用最小平方裁減來排除樣本中的離異點,後再結合華爾恕神經網路的數位式學習,達成較易取得合適的初值與數位近似。第二個方法是使用ε-SVMR來評估樣本資料的選取,後再結合華爾恕神經網路的數位式學習,達成較易取得合適的初值與數位近似。第三個方法是使用智慧型數位式神經網路在學習過程中加入非對稱與對稱影響函數來改善樣本中含非對稱的離異點的排除,獲得最佳的近似結果。最後,將我們提出的方法應用於系統生物的建模上。
In recent years, machine learning is one of the popular research fields. Its main purpose is for using the user’s and input data to acquire knowledge. In machine learning is not only to solve more problems, but also it can reduce errors and improve to efficiency of solving the problem. In this thesis, we proposed three new machine learning architectures to extend the learning process and achieve good predictions for noise and outlier data. That is, we proposed three methods; namely, support vector machine regression (SVMR) and least trimmed squares (LTS) and asymmetric influence function that combined with digital Walsh neural networks and can satisfy robustness and suitable to deal with noise and outliers data. The first architecture is to use the least trimmed squares support vector machine regression (LTS-SVMR) under the use of SVMR for the initial function in LTS to against the sampling data with noise and outliers, and then combine with digital Walsh neural networks to achieve digital approximations. The second method is to use ε-SVMR as the initial stage to choose suitable sample data under noise and outliers, and then combine with digital Walsh neural networks to achieve digital approximations. The third method is to use the intelligent digital neural networks that were added asymmetric influence function in the learning process to exclude the data samples containing asymmetric outliers, and to obtain the best approximation. Finally, we applied these propose methods on modeling of biological systems.