Open optimization problem in urban mobility digital twins: high-dimensional OD demand estimation

Establish scalable and sample-efficient optimization methods for estimating high-dimensional origin–destination (OD) travel demand matrices in large urban road networks, formulated as minimizing the squared discrepancy between ground-truth (field) link traffic counts and the expected simulated counts from a stochastic, non-differentiable, and computationally expensive traffic simulator (e.g., SUMO), under sparse and noisy sensor observations. The objective is to enable reliable OD calibration at metropolitan scale, including instances with up to 10,100 OD variables and realistic nonlinear dynamics.

Background

The paper frames origin–destination (OD) demand estimation as a central, difficult inverse optimization task for urban mobility digital twins. The problem requires inferring a high-dimensional OD vector so simulated traffic statistics align with ground-truth sensor measurements. Each evaluation necessitates running a stochastic, non-differentiable, and expensive traffic simulator, rendering the task a black-box optimization problem.

BO4Mob introduces five benchmark scenarios derived from San Jose, CA freeway networks, with dimensionality scaling up to 10,100 OD variables. The setting incorporates real-world sensors and realistic stochastic dynamics. Empirical results show existing Bayesian optimization methods improve performance on small and mid-scale networks but struggle at full metropolitan scale, highlighting the open challenge of scalable, sample-efficient OD calibration.