An enhanced SPSA algorithm for the calibration of Dynamic Traffic Assignment models

摘要: Simultaneous perturbation stochastic approximation (SPSA) is an efficient and well established optimization method that approximates gradients from successive objective function evaluations. It especially attractive for high-dimensional problems has been successfully applied to the calibration of Dynamic Traffic Assignment (DTA) models. This paper presents enhanced SPSA algorithm, called Weighted (W-SPSA), which incorporates information spatial temporal correlation in a traffic network limit impact noise improve convergence robustness. W-SPSA appears outperform original algorithm by reducing generated uncorrelated measurements gradient approximation, DTA models sparsely correlated large-scale networks large number time intervals. Comparisons between have performed through rigorous synthetic tests application real world demonstrated with case study entire expressway Singapore.