Structure Learning for the Estimation of Non-Parametric Incident Duration Prediction Models

To minimize the effects of incident-induced congestion and to prevent secondary accidents, an effective incident management tool, which can reliably predict incident durations in real-time, is needed. One of the major challenges of real-time estimation of incident durations is that initial information about an incident is generally very limited and it may not be received in a pre-specified order. Most of traditional prediction models such as decision trees and linear regression models determine the order. The objective of this research is to develop a practical model, which is capable of accurately predicting incidents using Bayesian Networks (BNs) that can be easily incorporated into incident management activities of Traffic Management Centers (TMCs) to improve decision-making process. For this purpose, three structure learning algorithms are employed to construct BN structures and then the best performing one is chosen for the incident duration prediction using incident data collected in New Jersey in 2005. It is found that the proposed model can predict incident durations more accurately than the traditional approaches such as linear regression and Classification and Regression Trees (CART). In addition to significant improvements in prediction, BNs are shown to be useful to analyze important relationships among model variables using the concept of strength of links. Finally, through a systematic approach of randomly eliminating incident input data to mimic real-world conditions, it is shown that developed BN models can also be effective tools that can handle lack of input data, which is a common problem during real-time incident management operations due to their highly flexible and non-hierarchical nature. Future work should include the use of geographically more diverse incident data to obtain BNs to better understand transferability of these models among different locations.

• Supplemental Notes:
  • The DVD lists the title of this paper as: Structure Learning for Estimation of Nonparametric Incident Duration Prediction Models.
• Corporate Authors:

  Transportation Research Board

  500 Fifth Street, NW
  Washington, DC  United States  20001
• Authors:
  • Demiroluk, Sami
  • Ozbay, Kaan
• Conference:
  • Transportation Research Board 90th Annual Meeting
  • Location: Washington DC, United States
  • Date: 2011-1-23 to 2011-1-27
• Date: 2011

Language

• English

Media Info

• Media Type: DVD
• Features: Figures; References; Tables;
• Pagination: 19p
• Monograph Title: TRB 90th Annual Meeting Compendium of Papers DVD

Subject/Index Terms

• TRT Terms: Algorithms; Crash data; Data collection; Decision making; Incident management; Real time information; Time duration; Traffic congestion; Traffic control centers; Traffic incidents
• Uncontrolled Terms: Bayesian networks; Incident duration
• Subject Areas: Highways; Operations and Traffic Management; Planning and Forecasting; I72: Traffic and Transport Planning;

Filing Info

• Accession Number: 01336796
• Record Type: Publication
• Report/Paper Numbers: 11-3143
• Files: TRIS, TRB
• Created Date: Apr 18 2011 12:24PM