Incident Duration Prediction Based on Bayesian Network

摘要

Incident duration is critical input for predicting the potential impact of incidents. Because of the ability to readily accommodate incomplete information, Bayesian network classifiers are chosen to predict incident duration. After analyzing the limitation of Naive Bayesian (NB) classifier and unrestricted Bayesian networks ( UBN) classifier, a new model is developed based on tree augmented Naive Bayesian (TAN) classifier. These models are calibrated and tested by incident records from the Georgia Department of Transportation. The results show that TAN classifier performs favorably compared to UBN classifier and NB classifier under a variety of information provision scenarios.