Artificial neural network approach to crash modeling and prediction.

摘要

The focus of this dissertation is on the evaluation of artificial neural network modeling techniques for intersection crash prediction. Two types of artificial neural network modeling techniques were evaluated---the back propagation neural network modeling and the Bayesian neural network modeling. This dissertation is the first known study to evaluate the use of Bayesian neural networks to model the prediction of intersection crashes, although one previous study has investigated the use of Bayesian neural networks for predicting crashes along roadway segments with promising results. The ideas from that previous study were extended in this dissertation for intersection crash prediction modeling.;Several categorical as well as continuous explanatory variables were modeled in this dissertation to evaluate if both types of variables could be adequately handled by the neural network techniques. The neural network techniques were compared with the traditionally accepted negative binomial regression modeling technique for crash prediction. The models were trained with training datasets and their predictive capabilities were evaluated with test datasets. Predictive capabilities were evaluated using three measures---Mean Prediction Bias, Mean Absolute Deviation and Mean Square Prediction Error. Bayesian neural network models exhibited the most superior prediction capabilities on the datasets analyzed.;A new approach to rank the importance or the level of influence of explanatory variables for Bayesian neural networks was also developed in this dissertation. Two new concepts that utilized the information on active connections within the neural network were introduced to determine the ranking of the explanatory variables---Full Connection Frequency (FCF) and the Average Total Connections (ATC). The ranking of the variables is a step towards identification of significant variables with a Bayesian neural network.;A secondary contribution of this dissertation was the development of an algorithm to automate the mapping of the crash data into GIS maps. The implementation of the algorithm exhibited a high accuracy for the intersection crash mapping. Although the crash database and the roadway database used in the mapping algorithm are specific to the state of Wisconsin, the ideas generated from the algorithm can be generalized to other databases.