Calibrating Dynamic Traffic Assignment Models by Parallel Search using Active-CMA-ES

The widespread deployment of inductive-loop traffic detectors allows us to obtain a massive amount of traffic data in real time. A key step of utilizing the data is to use the data to fit a traffic model. Simultaneous perturbation stochastic approximation (SPSA) and its variants are popular techniques for the calibration of dynamic traffic assignment (DTA) models by searching for Origin-Destination (OD) matrices that fit the data. However, the performance of SPSA cannot scale with modern multi-core CPU architectures due to its sequential nature. This paper proposes the use of active covariance matrix adaptation evolution strategy (Active-CMA-ES) for optimizing OD matrices in microscopic traffic simulation. CMA-ES is a blackbox optimization algorithm that was found highly effective in many domains. According to our case study in Ulsan, South Korea, Active-CMA-ES outperforms Restart-WSPSA, one of the best SPSA algorithms, in terms of the calibration error and the running time. Moreover, the running time of Active-CMA-ES decreases as the number of parallel simulation processes increases.