Application Of Artificial Neural Networks For Exploratory Analysis Of Small Dataset

Abstract

This study aims to explore the use of Artificial Neural Networks (ANNs) for estimating the relationship between accidents and other variables with a small dataset. ANNs have not been used to explore relationships between variables, especially for road accidents which have small datasets. Analysis of road traffic accidents is often hampered due to insufficient datasets. Especially, for the cases when specific highway facilities are considered. This issue is also gaining importance for analyzing traveler behavior with the advent of new technologies and implementation of concepts of smart cities. The accident sites selected for this study comprise of unsignalized intersections in Penang State of Malaysia. Accidents in Malaysia have become a major concern for the authorities. However, the data collection is a major issue hindering its analysis because of limited datasets. The safe operation of traffic on unsignalized intersections mainly depends on drivers’ judgement and decision making. Hence, the safety considerations on such locations are peculiar in comparison to other facilities. These facts led to carrying out the study for these sites. Two types of ANNs were used i.e. Multilayer Feedforward (MLFF) and linear. In addition, regression model and Mann-Whitney test were also used to compare the results from ANNs. It was found that regression model as well as Mann-Whitney test gave inconclusive results for the available dataset. On the other hand, ANNs were able to approximate the relationship in conformity to the previous studies. It was found that major road width and near-to-far-volume ratio increases the accidents while near-to-far-gap ratio reduces the accidents. The accuracy of ANNs was also better than regression model with an average error of less 40% for ANNs compared to 46% by regression model. However, larger datasets are expected to give better accuracies for regression as well as ANNs.