Application of neural network ensembles to incident detection

Traffic incident is an essential part of traffic control and management systems. This paper presents the application of Neural Network ensembles (NN ensembles) in incident detection. In addition, we proposed a new method to combine the outputs of networks, which made use of probability to improve further the performance of NN ensembles. Based on Boosting and Bagging, We generated neural network members, then employed several ensemble methods, including majority voting, weighted voting and our proposed method to combine the output of members to detect traffic incident. Several NN ensembles based detect incident models have been developed and tested with real 1-880 freeway traffic data collected in California. The performance of the neural network ensemble is compared to the single neural network. Empirical results indicated that neural network ensemble has advantages over single neural network, and incident detection based on neural network ensembles is a promising approach.