Structure and weight initialization of feed-forward networks for traffic flow-density relationships

Urban traffic is a complex process that is often described by macroscopic flow models. On one hand, this work justified how feedforward networks (NN) with two neurons in the hidden layer can fit real traffic data. On the other hand, two novel initializations are suggested. Both methods exploit the NN traffic model structure and the infrastructure parameters to ensure that the outputs of neurons are in the active region and speed up the learning convergence.