Calibration of the demand structure for dynamic traffic assignment using flow and speed data. Exploiting the advantage of distributed computing in derivative-free optimization algorithms

Stochastic optimization algorithms have been used in the recent literature as a preferred way for calibrating Dynamic Traffic Assignment (DTA) models, as the computation of explicit gradients is numerically too cumbersome on real networks. However, early experiences based on the Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm have shown performance issues when the number of variables becomes large. This suggests to focus on structural demand variables rather than to consider all components of origin-destination (O-D) matrices. Moreover, with the possibility of distributed computing, many algorithms that where not efficient in a standard configuration (i.e. sequential objective function evaluations within each iteration) can become a viable alternative to SPSA. For example, parallelization can be especially beneficial for genetic algorithms, which require a large number of independent function evaluations per iteration. In this paper we examine several optimization algorithms applied to dynamic demand calibration using flow and speed field measurements. The problem is to minimize the distance between results of a dynamic network loading and traffic data observed on road links. This approach is investigated in the context of laboratory experiments, where known O-D matrices are perturbed after its dynamic assignment on the network, to prove the effectiveness of the proposed methodology.