The Impact of Data Quantity on the Performance of Neural Network Freeway Incident Detection Models

Abstract

One of the difficulties in the development of artificial neural network (ANN) models is that, unlike statistical modelling where estimates of sample size can be initially computed, the number of samples or observations needed for training ANN models cannot be determined in advance. This is further complicated when dealing with ‘real world’ data that is not easily available or difficult and time consuming to collect. It is therefore desired that the impact of sample size on model performance be investigated such that the trade-off in performance using different sample sizes is evaluated. This issue is discussed in this paper in the context of a neural network freeway incident detection model that was developed using ‘real world’ incident and traffic data. From a practical perspective, the impact of sample size on model performance will provide an insight into the sample size of ‘real world’ data required to train ANN incident detection models. The results reported in this paper can also be used to make decisions about the sample size required for retraining the ANN incident detection model once it becomes out of date due to changed traffic conditions and/or upgrading of the facility.