A data-driven approach for learning traffic assignment using an artificial neural network

Solving transportation network design problems using traffic assignment models becomes increasingly impractical as the size of the considered network grows, due to the need to evaluate numerous candidate solutions. This study aims to model the traffic assignment process through a data-driven approach, and introduces a model based on an artificial neural network (ANN) to predict the total travel time in a network under various link capacity modifications for a specified set of links. ANN-based prediction models are trained and validated on a synthetic data set generated from a medium-sized network to evaluate the feasibility of the proposed approach. Models developed using different ANN architectures highlight the importance of selecting the most suitable architecture specific to the problem. The practical value of the ANN-based prediction model becomes apparent when used in optimising a transportation network design problem instead of the traditional traffic-assignment model: the ANN-based model reduced by up to 60 000 times the computation time required for the optimisation.