都市旅運行為受到社會變遷與家戶型態轉變等因素影響,使得都市旅運行為變得更加複雜且難以預測。本研究旨在探討在旅次鏈假設下之都市多活動旅運決策過程。根據台北市一日旅運資料,分析出較顯著之工作相關旅程類型,並從此資料中挑選出影響較大之輸入因子,針對具不同社經特性之個體,利用倒傳遞網路模擬都市通勤旅程之活動—旅運選擇傾向。將先前研究中針對特定活動—旅運類型之二項分析,擴展為多選擇模式,以更符合旅運者之實際決策過程。然而多項模式可能面臨推估結果之傾向度偏低,選擇方案數目愈多,此問題愈嚴重,造成目標方案之傾向度往往低於0.5。本研究除了倒傳遞網路之外並嘗試以不同分析方法建構都市多活動旅程選擇模式。 首先利用巢式倒傳遞網路,將被選擇問題分層處理,在各層利用二項模式傾向度保證之基本優點,建立被選方案多層二項模式。利用89年台北都會區住戶交通旅次調查資料,進行可能旅程之判別分析,作為各層巢化之分類依循,分層巢化之標準根據整體推估正確率及目標方案正確率決定。此多層二項模式之推估結果和傳統單層多選擇模式比較,其雖無法同時顯示各方案間之相關參數值,但就多選擇決策而言,其推估結果可提高目標方案之傾向度,並有效降低推估結果傾向偏低之缺點。除了巢式倒傳遞網路之外,並以自適應共振理論結合倒傳遞網路模式建構都市多活動旅程選擇模式。利用自適應共振理論網路之自我分類優點,建立被選方案之重新分類,從其中獲得較接近實際情況之分類,再將新分類帶入倒傳遞網路中進行比較。其結果顯示以旅運相關變數帶入自適應共振理論網路中所得到之新分類在倒傳遞網路中學習的效果為最佳。在倒傳遞網路驗證方面,本研究嘗試使用凍結部份權重之方式來進行驗證模式適應性之提高。雖然準確改善情況不如原先所預期,但其結果對於網路之學習速率有相當之幫助。
The travel behavior in urban area becomes more complex and hard to predict than before due to changes of social and economic structure. The aim of this study was to explore the decision-making process of daily journeys comprised with multiple activities based on trip-chaining hypotheses. By using the one-day trip survey in Taipei metropolitan area, distinct work-related tours were first categorized. The analysis was followed by selection of influence factors and simulation of urban activity-travel choice tendency modeled by back-propagation network (BPN). The previous works of analyzing binary choices were firstly expanded into a multiple-choice model in this study to better explain the decision-making process in reality. However, the problems of low tendencies estimated by multiple-choice models are frequently seen, more choices to be selected, worse the problem. As a result, the estimated choice tendencies of selected travel tours sometimes were barely fifty percents or lower. Several forms of neural network were attempted here in addition to the traditional BPN models in order to improve the problem. A nested back propagation network was tried thereafter by separating choices with layers to compare with the tradition multi-choice model; utilizing the merit of securing choice tendencies, a multi-layer binary model was therefore constructed. The probable tours were classified and the distinct one was specified in each later. Criteria of the nested structure were based on the total as well as objective correct rates. Compare to the traditional multi-choice model, the multi-layer binary model is unable to reveal the simultaneous relationship if all choices presented at the same time. Nevertheless, the sequential structure identified the choice hierarchy and promised acceptable choice tendencies. Additionally, the adaptive resonance theory (ART) was introduced here to combine with BPN as a pre-processing submodel for classification. By profiting the self-grouping capability of ART, the previously given tour choices were re-examined and re-categorized; the regrouped choices were then used in BPN. The results indicated the new categories generated by ART with travel attributes provided the best training results in BPN. The transferability based on model tests in the past reports was usually less emphasized or performing poorly. In this study, certain weights exhibiting least influences were frozen after initial analysis to increase model generalization. The results showed this method was helpful for increasing learning speed.
為了持續優化網站功能與使用者體驗,本網站將Cookies分析技術用於網站營運、分析和個人化服務之目的。
若您繼續瀏覽本網站,即表示您同意本網站使用Cookies。