NETWORK-WIDE TRAFFIC SIGNAL CONTROL WITH RECURRENT NEURAL NETWORKS

摘要

This paper presents a network-wide traffic control model for urban road networksreferred to as NEURO-NESC, which is based on four cooperating recurrent neuralnetworks, three of which are error-propagation networks. In this model, a very closecooperation between network-wide dynamic traffic assignment and control isachieved. NEURO-NESC is based on NEUROMONET, which comprises tworecurrent neural networks: one that serves as a traffic simulation network, and theother as a dynamic route chooser that computes optimal flow splitting rates. Ourmodel NEURO-NESC extends the existing model by two other error-propagationnetworks that are related to the three essential parameters of traffic signal control:green, cycle and offset times. The time values correspond directly with link capacitiesthat shall adjusted according to the minimization of some given objective function.