Accelerated on-line calibration of dynamic traffic assignment using distributed stochastic gradient approximation

Dynamic Traffic Assignment (DTA) system [Ben-Akiva et al., 1991] [Mahmassani, 2001] benefits travelers by providing accurate estimate of current traf-fic conditions, consistent anticipatory network information as well as reliable route guidance. Over the years, two types of model adjustment schemes have been studied - DTA off-line calibration [Balakrishna, 2006] [Toledo et al., 2003] [van der Zijpp, 1997] and DTA online calibration [Antoniou et al., 2007] [Wang et al., 2007] [Ashok and Ben-Akiva, 2000]. The on-line calibration of DTA system allows real-time model self-corrections and has proven to be a useful complement to off-line calibration. In this paper, we explore distributed gradient calculations for the speed-up of on-line calibration of Dynamic Traffic Assignment (DTA) systems. Extended Kalman Filter (EKF) and Stochastic Gradient Descent (GD) are examined and their corresponding distributed versions (Para-EKF and Para-GD) are proposed. A case study is performed on a 25-km expressway in Western Portugal. We empirically show that the application of distributed gradient calculation significantly reduce the computational time of online calibration and thus provide attractive alternatives for speed-critical real-time DTA systems.