Simulation-Based Optimization: Achieving Computational Efficiency Through the Use of Multiple Simulators

Transportation agencies often resort to the use of traffic simulation models to evaluate the impacts of changes in network design or network operations. They often have multiple traffic simulation tools that cover the network area where changes are to be made. These multiple simulators may differ in their modeling assumptions (e.g., macroscopic versus microscopic), in their reliability (e.g., quality of their calibration), as well as in their modeling scale (e.g., city-scale versus regional-scale). The choice of which simulation model to rely on, let alone of how to combine their use, is intricate. A larger-scale model may, for instance, capture more accurately the local-global interactions; yet may do so at a greater computational cost. This paper proposes an optimization framework that enables multiple simulation models to be jointly and efficiently used to address continuous urban transportation optimization problems. We propose a simulation-based optimization algorithm that embeds information from both a high-accuracy low-efficiency simulator and a low-accuracy high-efficiency simulator. At every iteration, the algorithm decides which simulator to evaluate. This decision is based on an analytical approximation of the accuracy loss as a result of running the lower-accuracy model. We formulate an analytical expression that is based on a differentiable and computationally efficient to evaluate traffic assignment model. We evaluate the performance of the algorithm with a traffic signal control problem on both a small network and a city network. We show that the proposed algorithm identifies signal plans with excellent performance, and can do so at a significantly lower computational cost than when systematically running the high-accuracy simulator. The proposed methodology contributes to enable large-scale high-resolution traffic simulation models to be used efficiently for simulation-based optimization. More broadly, it enables the use of multiple simulation models that may differ, for instance, in their scale, their resolution, or their computational costs, to be used jointly for optimization.