以貝氏方法構建與求解路徑基礎之時變交通指派模式

Abstract

現今之交通運輸系統不再是以「大規模的建設」來解決所有的運輸問題，而是朝向一個更細緻化、對環境更友善且能夠永續經營的方式。智慧型運輸系統，整合了電子、通訊、資訊處理等技術，希望能夠透過有效之管理方式來減少交通擁擠狀況。於智慧型運輸系統當中的先進交通管理系統，則需要即時交通狀況之資訊，才能據以進行分析與控制。因此，本研究發展一路徑基礎之動態交通量指派模式，並據此推估路網現況。 大多數動態交通量指派之研究，均專注於探討使用者均衡或系統最佳狀態。但路網現況是否滿足使用者均衡，仍有許多學者存疑。是故，本研究不以使用者均衡指觀點描述交通量指派問題，而是透過線性動態系統構建路徑基礎之動態交通量指派模式。由最基礎之非時變且不考慮旅行時間之模式，逐漸放鬆成為時變且考量旅行時間之模式。過去針對此種動態系統所構建之模式，大多需要歷史之起迄流量資訊、路徑選擇矩陣或狀態轉移矩陣；但在現實環境中，這些資訊不一定能夠順利取得。本研究透過貝氏方法以及卡門濾波兩者之結合，放鬆上述之假設條件，並且提出一整合型演算法求解此模式。為確認此演算法之收斂，本研究亦提出一平行數列收斂性確認方式，作為此演算法之收斂停止條件。由於此演算法需要大量之計算，為增進計算效率，本研究將此演算法以通訊量最小化的目標進行平行化，並執行於平行電腦上。透過真實路網流量資訊，本研究得以驗證此模式之估計與執行效率。

Transportation system nowadays is no longer “extraordinary construction” but becomes more elegant, environmental-friendly, and sustainable. Intelligent Transportation System (ITS) integrates the telecommunications, automation, electronics, and information processing system, is considered possessing the potential to solve the traffic congestion problem. Advanced Traffic Management Systems (ATMS), requires real-time traffic condition, is one of the key issues of ITS. Therefore, we suggest a path-based traffic assignment model to describe the network flow status. Most existing research works on dynamic traffic assignment focus on the user-equilibrium or system optimal. Nevertheless, the existence of such assumption in real world network is questionable to many researchers. An approach without these assumptions while keeping the basic traffic relationship might be useful facing the disequilibria issue. Therefore, we model the dynamic traffic assignment problem with dynamic system approach. Existing researches with this approach usually assume the prior information of O-D matrix, link-proportion matrix, or state transition matrix. In this paper, we relax such assumption by combining Gibbs sampler and Kalman filter in a state space model. A solution algorithm with parallel chain convergence control is proposed and implemented. To enhance its efficiency, a parallel structure is suggested with efficiency and speedup demonstrated using PC-cluster.