High-dimensional offline origin-destination (OD) demand calibration for stochastic traffic simulators of large-scale road networks

This paper considers high-dimensional offline calibration problems for large-scale simulation-based network models. We propose a metamodel simulation-based optimization (SO) approach. The proposed method is formulated and validated on a simple synthetic toy network. It is then applied to a high-dimensional case study of a large-scale Singapore network. Compared to two benchmark methods, a derivative-free pattern search method and the SPSA method, the proposed method improves the objective function estimates by two orders of magnitude. Moreover, this improvement is achieved after only 2 simulation runs. Hence, the proposed method is computationally efficient. The main idea of the proposed approach is to embed, within the SO algorithm, information from an analytical (i.e., lower-resolution) yet differentiable and tractable network model. It is this analytical structural information that enables the SO algorithm to become both suitable for high-dimensional problems and computationally efficient. For a network with n links, the analytical network model is implemented as a system of n nonlinear equations. Hence, it scales linearly with the number of links in the network and independently of link attributes (such as link length) and of the dimension of the route choice set. (C) 2019 Published by Elsevier Ltd.