Metamodel-based calibration of large-scale multimodal microscopic traffic simulation

Abstract

Agent-based microscopic traffic simulation models are gaining popularity over traditional trip-based traffic models due to their superiority in modeling household member interaction, car-sharing, vehicle interaction, etc. Calibrating these models is challenging due to the intrinsic complexity, e.g., nonlinearity, high simulation time, and lack of closed-form expressions. With multimodal large-scale networks, this problem becomes even more severe. This paper proposes a metamodel-based simulation optimization framework for calibrating an agent-based multimodal traffic microsimulator. The metamodel serves as a surrogate of the simulation model for faster and more tractable calibration optimization. In particular, a simplified multimodal traffic model is developed and incorporated in the metamodel. This paper also develops and compares three gradient-based metamodel schemes, taking account of the derivatives of link flows with respect to the calibration parameters in the metamodel fitting process. Numerical results on a toy network show that the proposed metamodel incorporating an analytical traffic model is more efficient and highly adaptable compared to the conventional pure polynomial metamodel schemes. Moreover, taking into account gradient information in the metamodels improves the performance of convergence and solution quality. The proposed metamodel-based method is further applied to calibrate the simulation of an actual network in Hong Kong. The results confirm the applicability of the proposed calibration method for large-scale networks. The proposed framework can be applied to other simulation-based optimization problems as well.