Urban Transport Decision Making: Improving Traffic Prediction with Symbolic Regression, Transfer Learning and Deep Learning

Abstract

Transport administrators make decisions about road infrastructure and traffic management that significantly impact people’s lives.To improve efficiency, an increasing number of decision makers complement their expertise with the automated insight afforded by robust and reliable computational models of traffic. To produce such models, we propose an innovative intelligent algorithm that combines Symbolic Regression with Transfer Learning and Neural Networks. The algorithm learns historical and real time traffic patterns from several areas of the road network and transfers that knowledge to the location where a prediction is needed (i.e., the target zone), by injecting it into the model trained on local data. We enhanced our evolutionary algorithm with a Deep Learning component that automates the selection of areas to transfer knowledge from (i.e., the source zones). Its role is to identify those regions of the road network where pre-training models significantly increases predictive accuracy at target locations. Preliminary experiments show our approach is more likely to identify adequate transfer learning sources than algorithm variants where source selection is manual and, respectively, performed by a standard Artificial Neural Network.