張斐章臺灣大學：生物環境系統工程學研究所邱建堯Chiu, Chien-YaoChien-YaoChiu2007-11-272018-06-292007-11-272018-06-292005http://ntur.lib.ntu.edu.tw//handle/246246/56084典型的倒傳遞類神經網路(Back Propagation Neural Network ,BPNN)以最陡坡降法為搜尋機制，初始權重值採用隨機亂數為其初始值，此法容易造成搜尋時間上的浪費與容易落入區域解等問題。而且在複雜的可行解空間中，不同的初始值將產生不同的區域解，導致搜尋結果良莠不齊，使得傳統求解最佳權重經常必須以大量次數的搜尋方式進行。 本研究提出遺傳演算法(Genetic Algorithm, GA)結合共軛梯度演算法(Conjugated Gradient Algorithm, CG)之複合型搜尋機制，以求取BPNN權重的最佳解。藉由GA在高維度空間的強大搜尋能力來解決傳統以隨機亂數設定權重初始值，而使得求解過程耗時繁複的問題；接著，再透過CG快速簡易的演算特性，對GA搜尋結果進行更進一步的修正，期望提升BPNN預測模式的表現。 本研究以台北市文山區集水區中港下水道系統之一階段BPNN水位預測模式為例，比較以隨機亂數初始化網路權重、GA優選網路權重與複合型搜尋機制等三種求解方式的優劣，結果證明複合型搜尋機制能有效且快速的求得BPNN網路權重最佳解。研究並以複合型搜尋機制建構中港下水道水位二階段BPNN預測模式，並證實亦有相當良好的表現。The standard back propagation neural network (BPNN) uses the steepest descent method to search the optimal solution for the random initial value of connecting weights. However, the search result of this approach is highly dependent on the initial weights. It is difficult to tell whether the initial weights are close to the global minima and the searched solution could easily reach a local minimum when the weight space is complex. To solve this problem, the search process usually is run with a large number of sets of initial weights. That consumes lots of time for try-and-error and it is not an effective searching strategy. In this study, we propose a hybrid searching strategy, combining Genetic Algorithms (GA) with the Conjugate Gradient Algorithm (CG) as the search engine of BPNN, to improve the standard searching strategy. In this hybrid strategy, GA can globally search the weight space to get a number of better candidate solutions in its iterative generations. After GA process reached a stable condition, CG is then used to optimize the weights of BPNN. This hybrid searching strategy is not only effective but also has high possibility to reach the global optima. For demonstrating the performance of the proposed searching strategy, the urban drainage system of Zhong-Gang Catchment located in Wenshan District of Taipei City is used to evaluate its applicability and efficiency. We apply the proposed model to search the optima weights of BPNN to predict one-step-ahead and two-step-ahead sewer stage during flood events. The results show that the proposed strategy is robust and efficiency.章節目錄 摘 要 I ABSTRACT II 章節目錄 IV 表目錄 VII 圖目錄 VIII 第一章 前 言 1 1.1研究動機 1 1.2研究方法 2 第二章 文獻回顧 4 2.1類神經網路的發展 4 2.2遺傳演算法的演進 6 2.3不同模式在水文預測上之應用 6 第三章 理論概述 8 3.1類神經網路的特性 8 3.2倒傳遞類神經網路 9 3.3誤差倒傳遞演算法 12 3.4共軛梯度演算法 17 3.5遺傳演算法 19 3.5.1遺傳演算法求解問題架構 20 3.5.2遺傳演算法運算元 21 3.5.3菁英策略 22 3.6複合型搜尋機制 24 第四章 研究案例 25 4.1研究區域概況 25 4.2資料整理與分析 28 4.3模式測試與比較 31 4.3.1一階段預測模式 32 4.3.2模式比較 34 4.4模式評比指標 36 4.5結果討論 38 4.5.1隨機亂數初始網路權重 38 4.5.2 GA優選BPNN網路權重 40 4.5.3複合型搜尋機制 47 4.6二階段預測模式 53 第五章 結論與建議 64 5.1結論 64 5.2建議 65 第六章 參考文獻 671616757 bytesapplication/pdfen-US倒傳遞類神經網路下水道系統遺傳演算法共軛梯度演算法複合型搜尋機制Back Propagation Neural Networksewer drainage systemGenetic AlgorithmConjugated Gradient Algorithmhybrid searching strategythesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/56084/1/ntu-94-R92622009-1.pdf