Dynamic Calibration of a Carsharing System in Multiagent Transport Simulation

Abstract

When employing simulation platforms to reproduce user behaviors of a transport mode as close to reality as possible, the model used must be calibrated. With simulation software, several calibration methods can be used. Some models are calibrated manually: users modify the parameters to be calibrated, launch the simulation, and then evaluate the results. Users perform these steps until the simulation results satisfy the desired objective. In this study, the multiagent transport simulation (MATSim) platform was used to simulate user behavior across various transportation modes. In addition to known transportation modes (private cars, public transit, and walking), carsharing mode was added to the Siouxfalls scenario. This enhanced scenario was used as the baseline to simulate the impact of the chosen parameters on the agent’s behavior. Previous studies on the calibration of agent-based models such as MATSim have introduced two automated methods based on entropy and gradient descent. However, these methods can require a higher number of calibration iterations and longer simulations to reach the required parameters. To offer both an efficient method in relation to computation time, and a solution that is adapted specifically to the carsharing mode, this paper presents a calibration methodology based on a surrogate algorithm that will act as a substitute for MATSim during calibration to speed up the process. Assuming that the parameters related to the basic mode of transportation were calibrated, the process described in this paper was applied specifically to the calibration of the carsharing mode in MATSim.