Abstract
M.Com. (IT Management)
Vehicle traffic congestion on Gauteng’s metropolitan and national roads in South Africa remains a challenge, especially during peak hours and also when incidents occur. They hamper the flow of vehicle traffic. Traffic congestion has negative consequences for business and for commuters’ quality of life. The goal of this research project is to identify variables that influence the flow of traffic and to design a vehicle traffic prediction model which will predict the traffic flow pattern in advance, given a set of predictor variables that will enable motorists to make appropriate travel decisions ahead of time. The traffic flow was categorised into three classes, namely traffic jam, free flowing and flowing congestion. In this study, the artificial intelligence algorithms that include Bayesian Networks, K-Nearest-Neighbour, Multi-layer Perceptron (MLP) and C4.5 Decision Tree were used individually for predicting the vehicle traffic flow. The results obtained from these algorithms were compared using the predictive performance and prediction costs. The best predictive model is one that has lower cost and good performance.
The results show that the MLP was the best performing algorithm in terms of predictive performance and low prediction costs. In order to predict a novel instance, a feed forward Multi-layer Perceptron network was built using Matlab and was used to predict the unseen vehicle traffic instance, also called a novel instance. The (MLP) model accurately predicts vehicle traffic flow on a single novel instance with a prediction performance of 80% (16 out of 20) on novel instances.