銀行在評分模型建置完成後需進行驗證工作，以瞭解評分模型是否能有效評出客戶的風險層級，穩健地估計區別鑑別力指標為驗證工作中的重點。在先前的文獻中假設正常授信戶與違約戶分數分配為常態分配。但在實際資料中，分配未必定為常態。因此本文接著探討在正常授信戶與違約授信戶之分配為混合分配，即兩分數分配為偏斜常態分配下，何種方法可以對於估計AUC具有較高的穩定性。本文比較五種估計AUC的方法，分別為常態核，經驗分配，曼惠尼近似，最大摡似法和EM演算法。模擬結果呈現(1)投信戶組合分配為兩常態分配下，最大摡似法在大部分違約率下都可以得到較窄的信賴區間。(2)組合分配為一常態與一偏斜常態及兩偏斜常態分配下，EM演算法在大部分情況有較窄的信賴區間，其中在兩偏斜常態分配下，表現更佳。(3)曼惠尼近似建構的信賴區間寬度最大，代表曼惠尼近似是較保守的估計方法。 / Banks face discrimination after constructing the rating systems to figure out whether the systems can discriminate defaulting and non-defaulting borrowers. Literature assumed the two score distribuion are normal distributed. However, the real data may not be normal distribuions. We assum the two score distribuions are skewed normal distribuions to discuss which method has more robustness to estimate the AUC value.Under skewed distribution, we propose EM algorithm to estimate the population parametric. If used properly, information about the population properties may be used to get better accuracy of estimation the AUC value.Numerical results show the EM algorithm method , comparing with other methods, has robustness in detect the rating systems have discirmatory power.

隨著經濟的高度成長及電廠興建的日益困難，電力的需求問題是愈來愈嚴重。如何有效率地安排核能機組進行例行性的停機大修及燃料再裝填工作是重要的課題。本論文中考慮核能發電機組五年時程的大修排程問題，我們將這個大修排程問題描繪成一個大型混合型整數線性規劃模型。由於問題的龐大與複雜，此問題的最佳解難以求出。因此，我們發展數個邏輯條件式有效地縮小解集合空間；另外並發展出一個啟發性演算法，採用合併變數法將0/1決策變數合併，使原問題轉成較小的合併模型。先解合併後的合併模型，利用合併模型答案的資訊來固定原始模型的部分變數值之後，再解原始問題。幾個實例計算顯示此演算法的可行性。 / Since the growth of economics and the difficulty to build a new power plant, the supply of electric power has become very tight. It is important to ensure the efficient operation of nuclear power plants, including timely shutdown, refueling and maintenance schedule. In this thesis, we deal with the scheduling shutdown and maintenance of nuclear power plants for a five-year time period. This problem can be formulated as a large-scale mixed integer linear problem. The difficulty of solving this problem is due to the large number of binary variables. We then develop several valid logical constraints to reduce the complexity in processing using the branch and bound technique. Also, a heuristic based on the aggregation and dis-aggregation techniques has been developed to yield a good solution. Several examples are given to show the applicability of the algorithm.

在代理人基計算建模(agent-based computational modeling)被拿來廣泛應用的同時，多數學者發現模擬的結果會高度取決於人工適應性個體的設計方式或者是個體的學習方法上，所以如何挑選合適的演算法就成為我們應用代理人基計算建模時首要面臨的課題。 本文挑選了兩個常出現於文獻當中但是卻甚少一起比較的演算法，分別是遺傳演算法(genetic algorithms)與強化學習(reinforcement learning)。我們透過將演算法與學習理論(learning theory)結合的方式，歸納出這兩個高使用頻率的演算法各自有其適合描述的個體行為以及議題，最後並套用到代理人基彩券市場當中，而模擬的結果也證實符合真實彩券市場上多數人學習特性(個人式學習)的強化學習比起遺傳演算法更能完整地捕捉彩券市場上的特性。

電離層延遲為精密GPS定位及導航的主要誤差來源之一，為了減弱電離層延遲對GPS定位及導航的影響，可以利用雙頻GPS觀測量構建即時的區域電離層模型，以提供即時的電離層延遲誤差改正參數，修正因電離層延遲效應造成的定位及導航誤差。 本研究以台灣地區雙頻GPS觀測量，採用相位水準技術估算全電子含量（TEC）、修正的單站演算法估計各GPS衛星及接收儀之L1/L2差分延遲及以UNSW網格式演算法構建區域的電離層模型。並進而求得適合台灣南部地區網格式電離層模型之較佳網格大小及探討使用那些內政部衛星追蹤站的觀測資料，便可有效建立台灣地區的電離層模型。 / The ionospheric delay is one of the main sources of error in precise GPS positioning and navigation. The magnitude of the ionospheric delay is related to the Total Electron Content (TEC) along the radio wave path from a GPS satellite to the ground receiver. The TEC is a function of many variables, including long and short term changes in solar ionising flux, magnetic activity, season of the year, time of day, user location and viewing direction. A dual-frequency GPS receiver can eliminate (to the first order) the ionospheric delay through a linear combination of L1 and L2 observables. However, the majority of civilians use low-cost single-frequency GPS receivers that cannot use this option. Consequently, it is beneficial to estimate ionospheric delays over the region of interest, in real-time, in support of single-frequency GPS positioning and navigation applications. In order to improve real-time regional ionosphere modelling performance, a grid-based algorithm is proposed. Data from the southern Taiwan region GPS network were used to test the ionosphere modelling algorithms. From the test results described here, it is shown that the performance of real-time regional ionosphere modelling is improved significantly when the proposed algorithm is used.

量子擾動及挫折性兩者均可破壞絕對零溫的磁序，為近代凝態物 理關注的有趣現象。在外加橫場下的三角晶格易辛反鐵磁兼具量子臨 界現象(quantum criticality)及幾何挫折性，可謂量子磁性物質之一典 範理論模型。本論文利用平衡態及非平衡態量子蒙地卡羅(quantum Monte Carlo)方法探測三角晶格易辛反鐵磁之量子相變，其界定零溫 時無磁性的順磁態及具 Z6 對稱破缺的有序態(所謂時鐘態)。這裡的 量子蒙地卡羅方法為運用算符的零溫投射(zero-temperature projector) 及隨機序列展開(stochastic series expansion)演算法。在非平衡模擬 中，我們分別沿降溫過程及量子絕熱過程逼近量子相變點，藉此我們 得到動力學指數，及其它相關臨界指數。 / The destruction of magnetic long-range order at absolute zero temperature arising from quantum fluctuations and frustration is an interesting theme in modern condensed-matter physics. The triangular lattice Ising antiferromag- net in a transverse field provides a playground for the study of the combined effects of quantum criticality and geometrical frustration. In this thesis we use quantum Monte Carlo methods both in equilibrium and non-equilibrium setups to study the properties of the quantum critical point in the triangular lattice antiferromagnet, which separates a disordered paramagnetic state and an ordered clock state exhibiting Z6 symmetry breaking; The methods are based on a zero-temperature projector algorithm and the stochastic series ex- pansion algorithm. For the non-equilibrium setups, we obtain the dynamical exponent and other critical exponents at the quantum critical point approached by slowly decreasing temperature and through quantum annealing.

本文研究美國退休福利保險公司(PBGC)保險價值的計算,延伸 Marcus (1987)模型,提出狀態轉換過程保險價值模型計算,也就是將市場分為兩種情況,正成長率視為正常狀態,負成長率為衰退狀態,利用狀態轉換過程評價 PBGC 契約在經濟困難而終止和介入終止下合理的保險價值。在參數估計方面,本文以 S&P500股價指數和一年期國庫券資料參數估計值及Marcus(1987)和Pennacchi and Lewis(1994)的方式給定參數,以 EM-PSO-Gradient 延伸 EM-Gradient 方法並以最大概似函數值、AIC 準則和 BIC 準則比較估計結果。最後固定其他參數, 探討狀態轉換過程保險價值模型對參數調整後保險價值的影響之敏感度分析。 / In this paper, we evaluate Pension Benefit Guaranty Corporation insurance values through regime switching models, which is the extension of the models of Marcus (1987). That is, we can separate periods of economy with faster growth from those with slower growth when observing long-term trends in economy and calculate the reasonable PBGC insurance values under distress termination and intervention termination by regime switching processes. We set parameters by estimating S&P 500 index and 1-year treasury bills by EM-PSO-Gradient, which is the extensive method of EM-Gradient and refer the methods of setting parameters from Marcus (1987) and Pennacchi and Lewis (1994). After that, we compare the maximum likelihood estimates, AIC and BIC of the estimative results. Finally, we do sensitivity analysis through given the other parameters and look into what would impact on our models of insurance values when adjusting one parameter.

技術面分析是使用過去市場資料包含股票價格與交易量來預測未來市場動態。技術分析將股價與交易量經由數學轉換成易懂且能繪製成圖表的技術分析指標，幫助技術分析投資人預測未來股價。本文的決策過程有別於傳統的技術面分析，使用相似度模型以貼近現實技術分析投資人的決策過程。此策略使用多個技術指標作為相似度技術指標交易策略的依據，用以捕捉市場動態與預測未來股價報酬，且即便不同的技術指標提供不同的買賣訊號，技術分析投資人依然可以藉由相似度技術指標交易策略進行投資決策。相似度技術指標交易策略所預測的未來報酬是根據過往價格圖形出現相似情境的報酬加權平均作為未來預測報酬。當預測報酬為正則買；預測報酬為負則賣。本文使用S&P500指數期貨來檢測相似度技術指標交易策略的獲利能力，發現在不同的技術指標下，相似度技術指標交易策略報酬顯著異於零也高於S&P500指數期貨在樣本期間內的B/H報酬。為使本文相似度技術指標交易策略更能模擬現實投資人的真實情況，導入機器學習改善相似度技術指標交易策略，分別使用貪婪演算法與模擬淬鍊法(Simulated Annealing)來模擬現實投資人會根據交易策略表現的好壞變更決策過程的策略。其報酬顯著異於零也高於S&P500指數期貨在樣本期間內的B/H報酬。本研究發現投資人會參考不同的混合技術指標策略，且會依照不同混合策略的過往績效，篩選出參考策略，進而決定投資策略，這也呼應混合技術指標的相似度技術指標交易策略比單一技術指標的相似度技術指標交易策略擁有較好的預測能力。因此使用混合技術指標的相似度技術指標交易策略作為機器學習篩選的策略可有效的改善原本的相似度技術指標交易策略。

我們提供一個新的可靠度模型，DwACM，並提供一個模式選擇準則CCP，我們利用DwACM和CCP來選擇衰變量。 / We propose a new reliability model, DwACM (Dichotomous-data with Auxiliary Continuous Measurements model) to describe a data set which consists of classical dichotomous response (Go or No Go) associated with a set of continuous auxiliary measurement. In this model, the lifetime of each individual is considered as a latent variable. Given the value of the latent variable, the dichotomous response is either 0 or 1 depending on if it fails or not at the measuring time. The continuous measurement can be regarded as observations of an underlying possible degradation candidate of which descending process is a function of the lifetime. Under the assumption that the failure of products is defined as the time at which the continuous measurement reaches a threshold, these two measurements can be linked in the proposed model. Statistical inference under this model are both in frequentist and Bayesian frameworks. To evaluate the continuous measurements, we provide a criterion, CCP (correct classification probability), to select the best degradation measurement. We also report our simulation studies of the performances of parameters estimators and CCP.

為了提升競爭優勢與生產能力，並進而達到永續經營的目的，突破現況、持續改善產品品質、降低產品成本與服務成本則成為提昇競爭力的重要因素之一，因此產品在設計開發階段就必需要考量品質與成本的問題。 本研究以電源轉換器為對象。該電源轉換器目前已設計完成且已通過美國UL安規認證，並已在國內量產銷售，但因為該電源轉換器的溫升及其變異很大，仍然會導致該產品的壽命過短，因此降低電源轉換器的溫升及其變異是一急需解決的問題。 透過了田口與實驗設計的方法規劃及進行實驗並收集數據。並利用十二種分析方法(包括：田口方法、主成份分析、主成份＋倒傳遞類神經網路＋基因演算法、主成份灰關聯＋倒傳遞類神經網路＋基因演算法、指數型理想函數＋倒傳遞類神經網路＋基因演算法、MSE方法、MSE方法＋倒傳遞類神經網路＋基因演算法、SUM方法、SUM方法＋倒傳遞類神經網路＋基因演算法、重要零件加總法、重要零件加總法＋倒傳遞類神經網路＋基因演算法)對實驗數據進行分析，以決定最適因子水準組合。 由改善後的確認實驗得到：雖然平均溫升下降的程度不大，然而大部份量測點的溫升標準差都顯著變小了。因此本研究在降低該電源轉換器溫升變異的效果十分顯著。對於電源轉換器的生產者而言，品質提升就是提升銷售量的保證，因此本研究所得到的最適因子水準組合，雖然產品在成本上有些微的增加，但品質改善後之產品將可為生產者帶來更多有形與無形之利益。

協同規劃、預測與補貨(Collaborative Planning, Forecasting and Replenishment,CPFR)，在歐美經過一些企業的採用後已經有顯著的成效，目前國內已經有一些企業相繼採用或即將採用CPFR，期望能因此降低供應鏈作業成本及提升供應鏈作業績效，以提升企業競爭力。在CPFR流程與供應鏈協同作業環境下，一個供需雙方協同，且績效良好的的銷售預測具有關鍵的重要性，是管理決策與協同合作時的的重要依據；但是多數的企業並沒有一個結構化、系統化的預測流程及方法，而是各部門透過簡單時間序列方法、天真預測法或人為經驗法則估算需求，進行多點且不同方法之預測，這樣的銷售預測較無穩定的品質，亦較難提供管理者合理的數據解釋。本研究結合時間序列、多元回歸模型與基因演算法發展出一個CPFR流程下之三階段混合預測方法，以買賣方直接之銷售資料、銷售計畫等資訊進行以「週」為單位之個別商品銷售預測。同時本研究中，亦以國內某製造業公司與其顧客（一國際大型零售連鎖店通路商）之產品銷售資料進行方法的驗證；實驗顯示，本研究所提出之預測方法之預測結果較Jeong等人(2002)所提結合多元回歸模型與基因演算法之二階段預測系統之預測結果佳；亦較傳統使用普通最小平方法求解之一般統計回歸方法預測結果佳。 / It has been verified in pilot projects by many European and American Corporations that Collaborative Planning, Forecasting and Replenishment (CPFR) can improve supply chain performance. Enterprises nowadays in Taiwan are implementing or going to implement CPFR, with hopes to reduce their supply chain operation cost, enhance logistic performance and increase their competition capability consequently. Under CPFR process and supply chain collaboration environment, a supply and demand both sides promised identical sales forecast with well forecasting performance for order decision making and cooperation is very important. Due to the dynamic complexities of both internal and external co-operate environment, many firms resort to qualitative, navie forecasting or other simple quantitative forecasting techniques and have many forecasts in their organization. However, these forecasting techniques lack the structure and extrapolation capability of quantitative forecasting models or without stable performance, while multi-forecasts providing different views of demand. Forecasting inaccuracies exist and typically lead to dramatic disturbances in sales order and production planning. This paper presents a hybrid forecasting model for sales forecasting requirements in CPFR. A three stage model is proposed that integrate the time series model, regression model and use genetic algorithm to determine its coefficients efficiently. Direct sales information and related planned events in both collaborated sides is used for individual product’s “week” sales forecasting. To verify this model, we experiment on two different products and produce forecasts with datum from one manufacturer in Taiwan and its international retailer. The results shows that the hybrid sales forecasting model has better forecasting performance than not only the causal-genetic forecasting model proposed by Jeong et al. (2002), but also ordinary regression model with no genetic training process.