چكيده به لاتين
Introduction: In Iran, in 1969, 469,950 accidents occurred in the country, of which 3% of the accidents occurred at the moment of the collision. 29 percent of these deaths occurred in urban areas, and in general, 70% of accidents occurred in the year of 1395 occurred in urban areas. Therefore, the need for further study on the assessment of accidents in urban segments is felt more and more. The occurrence of urban accidents is dependent on various and complex factors and this complexity makes it difficult to analyze. Using a variety of mathematical and statistical methods in dealing with this category can be fruitful. Over the past years, many efforts have been made to develop a variety of mathematical and statistical methods for predicting crashes, leading to the introduction of new models and even improvements to previous models.
Methods: Accordingly, two general goals are pursued in this doctoral thesis. First, we examine how different parameters, especially traffic parameters, as well as geometric design in the occurrence of accidents are affected. The secondary and final goal is to find out which modeling method will result in more accurate results in assessing the various aspects of the occurrence of accidents, considering the data and limitations in this study. The present doctoral thesis can even compile all of the known modeling methods in the two main subcategories of the crash assessment (severity and frequency of accidents), and then compare their performance in each of these subcategories It works. The present treatise predicts the frequency of accidents in urban roads in two separate stages.
Findings: In the first stage, four ELM, RBF, PNN and MLP models were selected and their performance was compared. At this stage, the ELM model with R ^ 2 equaled 81% at the training stage and 75% in the validation step, provided the most accurate and fastest prediction. The second phase used the MLP, ANFIS, Hybrid ANFIS-PSO and HybridANFIS-GA models as four subsystems of the computing intelligence approach. At this stage, the Hybrid ANFIS-GA model with MSE equaled 1.55 at the training stage and 2.78 at the validation stage led to the highest predictive accuracy. This doctoral thesis predicts the severity of accidents, as well as predicting the frequency of accidents in two separate stages. In the first step, four Multinomial Logit models (MNLs), Nested Multinomial Logit (NMNL), Mixed Logit (ML), and Multinomial probit (MNP) were developed and compared with each other. The results were consistent with the results of recent studies, in which the ML model with the AIC was 4266 was recognized as the best model. In addition to discrete choice methods, this doctoral thesis uses various data mining methods such as decision tree, Bayesian network and neural network to predict the severity of accidents and compare each other's performance. Based on the results, the C5 algorithm with the highest AUC is proposed as the best model.
Conclusion: According to the data used, it can be concluded that the use of genetic algorithm and its integration with other evolutionary intelligence methods in accident prediction is highly effective . Also, in order to predict the severity of accidents, data mining methods, especially the decision tree, lead to more accurate results.
Keywords: Traffic safety, Intensity prediction, Frequency prediction, Artificial intelligence models, Macroscopic Traffic Characteristics.
