چكيده به لاتين
Thousands of road accidents occur annually in Iran, resulting death of thousands of people and also excessive costs are imposed on the governments and he people. Several factors affect the occurrence and type of accidents. These factors can be categorize to vehicle characteristics, driver profiles, road profiles and weather conditions. Accident prediction and assessment of how each of the factors that affect the accidents can be a very important step in road safety management. For this purpose different mathematical and statistical models have been developed. One of these models, which has recently become so popular, is the Bayesian network. These networks are a combination of graph and probability theory. These networks are a directed acyclic graph that shows the conditional dependency or independency of the variables of model. Generally, the network can be constructed in two ways: first, by using an expert’s knowledge and second, automatically from data. In this thesis the focus is on learning Bayesian networks from data. In this thesis as a case study the data of accidents of Mashahd and Isfahan are collected and the accidents are analyzed using these data. Four different Bayesian networks have been constructed using four different methods and these models are compared with two criteria: 1-accuracy of prediction, 2-entropy. In Mashahd model, the model built with K2 algorithm had the best performance regarding to both criteria and in Isfahan model, the model built with AIC algorithm was the best model in comparison with other models. After learning the structure of the graphs according to data for each model, we analyzed the variable “type of collision” for Mashahd model and the variable “severity of accident” for Isfahan model. Variables that directly affect the type of collision are the median, the volume to capacity ratio and the vehicles involved in accident was determined and variables that have a direct impact on severity of accidents are type of vehicle, collision type and median presence. The effect of other variables that indirectly affects the type of collision and severity are analyzed using inference from the network.
