Developing insight into effective SPSA parameters through sensitivity analysis

Traffic simulation models have seen increasing use during the past decades. One of the biggest challenges related to their successful application, is the appropriate set of their values, thus achieving the accurate representation of driving and travel behavior parameters´ diversity. Models´ calibration using optimization algorithms, and more specifically the Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm, is a crucial step. In this research, we study several aspects of SPSA´s algorithm. A sensitivity analysis is implemented, in the context of finding the appropriate set of parameters, that will significantly improve its performance. Through successive experiments, the most efficient set is selected, and some guidelines are presented.