|Abstract: ||土壤液化是大地工程領域所關切的重要議題，如何正確的判斷場址液化與否乃是專家及學者所追求的目標。過去數十年所發展的土壤液化評估方法多以現地試驗資料所發展的簡易經驗法為主，且目前國內、外所採行之規範以及工程實務上亦多引用該法所發展之經驗公式。然而地震機制及土層特性具有高度不確定性，不易選擇適當的經驗公式來進行迴歸分析，因此各專家學者仍試圖找出較傳統經驗公式更為合理簡便且更加準確判斷土壤液化的分析模式。 隨著科技的進步，人工智慧(Artificial Intelligence, AI)已成為現今熱門科技研究及發展的電腦應用技術之一，簡單的說，它是利用電腦技術去模擬人類腦部思考運作，使電腦懂得自行分析及理解事物的一種科技應用技術。人工智慧的研究在70年代末期，專家系統(Expert System)發展時期達到高峰，之後由於技術上逐漸陷入瓶頸而陷入低潮。到了90年代中期，技術上逐漸有所突破，尤其是機器學習(Machine Learning)與模糊邏輯(Fuzzy Logic)資訊處理方法的發明，引起人們對人工智慧的廣泛注意與憧憬。因此AI又成為一門國際間競相投入的研究題材。 機器學習(Machine Learning)算是人工智慧領域最受關注的一支，目前應用範疇相當廣泛，而且已經成功的應用在文字、語音及影像辨識系統上。在大地工程領域方面，藉由能模擬人類思維與學習功能的機器學習理論，例如人工類神經網路(Artificial Neural-Networks, ANN)及支持向量機 (Support Vector Machine, SVM) 對於土壤液化這種高度非線性問題，均展現出相當優越的分析處理能力，許多學者研究亦證明此一技術對於處理土壤液化問題是一強而有力的工具，其優越性明顯高於傳統經驗評估法。因此本研究以機器學習理論為主要研究方法，針對目前應用較廣之ANN 及SVM等方法進行探索，期能建立快速及高準確率之視窗化液化評估模式，做為工程設計以及防災規劃之參考。 研究結果發現ANN及SVM所表現之液化分類正確率均優於傳統之簡易經驗法，其中又以SVM高達98.3 % 之分類正確率最為出色。SVM具有嚴格的統計理論基礎及處理高度非線性問題之能力，經證實可以應用在工程實務上，因此本研究發展一套以SVM 理論為基礎的液化評估程式LA-SVM (Liquefaction Assessment based on SVM)，它是建構在MATLAB/GUI (Graphical User Interface) 環境下的視窗作業系統。這個LA-SVM簡易視窗環境提供一個直覺、友善的操作介面，使用者不需任何圖表、公式及操作手冊，只須操作滑鼠游標點選視窗選項及輸入訓練資料及參數範圍即可迅速得到預測資料之分類結果及預測正確率。這套LA-SVM程式使得液化評估變得非常簡單，並且可以得到極高的準確率，值得後續之推廣及應用。|
Soil liquefaction is a vital topic of concern in the field of geotechnical engineering. Experts and scholars in this field seek methods to correctly determine whether liquefaction will occur at a site. The methods developed over the past several decades for evaluating soil liquefaction primarily involve simple empirical methods using in situ test data. Moreover, the current specifications and engineering practices adopted both domestically and internationally reference empirical formulae developed through these simple empirical methods. However, the great uncertainties inherent in earthquake mechanisms and soil properties complicate the selection of a suitable empirical formula for conducting regression analysis. Therefore, experts and scholars have attempted to discover an analytical method that is simpler, more reasonable, and better able to accurately determine soil liquefaction than traditional empirical formulae. With technological advances, artificial intelligence (AI) has become a popular topic in modern technological research and the development of computer applications technology. In a word, AI is an application of information technology that uses computing technology to simulate the thought processes of the human brain, thereby allowing a computer to conduct analyses and classification autonomously. Research into AI began in the 1950s and reached a peak at the end of the 1970s with the development of the Expert System. After this point, AI research met with certain obstacles and progressed slowly until the mid-1990s when several technological breakthroughs occurred. In particular, the invention of information processing methods such as machine learning and fuzzy logic incited widespread attention and excitement concerning AI. Artificial intelligence once again became a research topic of great interest in the international community. Among the branches of AI, machine learning receives the most attention. It currently demonstrates extensive applications, and has already been implemented successfully in word, voice, and image recognition systems. Applications of machine learning theory-for example, artificial neural networks (ANN) and support vector machines (SVM)—can simulate human thought and learning functions. These applications are advantageous for analytical processing of highly nonlinear relationships observed in geotechnical engineering domains such as soil liquefaction. Several scholars have confirmed that this technology is a powerful tool for handling liquefaction issues and is superior to traditional empirical assessment methods. Therefore, this study employs machine learning theory as a primary research tool to investigate the wider applications of current ANN and SVM methods. The study establishes a rapid and highly precise liquefaction evaluation model for Windows, which can serve as a reference for engineering design as well as disaster prevention and planning. The results of this study show that ANN and SVM surpass traditional simple empirical methods because of their accuracy in classifying liquefaction. Of the 2 methods, SVM is the most remarkable; it demonstrates a superior classification accuracy of 98.3%. SVM have proved to achieve good generalization performance and can be used as a practical tool for the prediction of soil liquefaction. Consequently, this study develops a Liquefaction Assessment program based on SVM (LA-SVM) that is constructed within a MATLAB/GUI (Graphical User Interface) Windows environment. This simple Windows-based LA-SVM provides an intuitive and user-friendly interface that obviates the use of graphs, formulae, or operating manuals. The user need only select items in Windows and enter training data and ranges to rapidly obtain classification results for forecast data and prediction accuracy. This LA-SVM program simplifies liquefaction assessment and achieves high accuracy rates, and should be promoted for further application.