Comparing the efficiency of different computation intelligence techniques in predicting accident frequency
Amir Mohammadian Amiri, Navid Nadimi, Amin Yousefian
Abstract
Until now, considerable efforts have been made to determine which modelling technique performs the best for predicting accident frequency based on crash data. In this regard, the presented study seeks to compare four types of Computational Intelligence (CI) modelling techniques in accident frequency prediction in urban segments, including Artificial Neural Network (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), Hybrid ANFIS-Genetic Algorithm (H-ANFIS-GA), and Hybrid ANFIS-Particle Swarm Optimization (H-ANFIS-PSO). Accordingly, different variables relating to traffic condition and road specifications were employed as independent variables, using the dataset consisting of 1370 crash occurred in Mashhad (Iran), in 2014. According to the results, H-ANFIS-GA exhibited the best performance in forecasting accident frequency. In contrast, PSO did not improve ANFIS performance, and it caused a negative influence on its prediction accuracy. Although the ANFIS model performed better than the developed ANN, it came in the third most accurate models. Additionally, the effect of each independent variable on predicted crash frequency was evaluated using sensitivity analysis.