A simulation-based optimization framework for urban traffic control

Microscopic urban traffic simulators embed numerous behavioral models that describe individual driver decisions (e.g. route choice, departure time choice, response to information) to capture the interactions between complex travel demand patterns and network supply phenomena. These disaggregate models make them detailed tools appropriate to perform scenario-based or sensitivity analysis. Nevertheless, this detail leads to nonlinear, stochastic and evaluation-expensive models. Their integration within an optimization framework to use them to identify traffic management schemes remains a difficult and challenging task. In order to perform computationally efficient simulation-based optimization for congested urban networks, information from the simulation tool should be combined with information from a network model that analytically captures the structure of the underlying problem. This paper presents a metamodel that combines the information from a microscopic traffic simulation model with an analytical queueing network model. It builds upon the classical approach of using a generalpurpose quadratic polynomial as a metamodel, by combining the polynomial with the analytical network model. We integrate this metamodel within a derivative-free trust region optimization framework. We evaluate the performance of this method with a fixed-time signal control problem for subnetworks of the city of Lausanne, considering different demand scenarios and tight computational budgets. The performance of the signal plans derived by the proposed method is compared to that of other methods, including an existing signal plan for the city of Lausanne. The method leads to well-performing signal plans for both small and larger samples. It leads to reduced, as well as more reliable, average travel times. This combination of a general-purpose metamodel, needed for its asymptotic properties, and an analytical network model, which provides tractable analytical information, leads to a simulation-based optimization framework that is computationally efficient and suitable for complex problems with tight computational budgets.