MIT researchers' dynamic calibration boosts traffic simulations by 48% accuracy

Accurate calibration of macroscopic traffic flow models such as METANET is essential for reliable prediction and effective traffic control. However, gradient-based methods often struggle with highly nonconvex optimization landscapes, leading to parameter sets that produce unstable, unrealistic dynamics. Standard static calibration frequently undermines confidence in the estimated parameters, limiting the simulation's utility for counterfactual scenario analysis.

To address this, researchers Shreyaa Raghavan, Cameron Hickert, Monica Chan, and Cathy Wu from MIT propose a dynamic, rolling-horizon calibration framework. By treating parameter estimation as a closed-loop control problem, the framework maintains stability and accuracy even in the presence of measurement noise. Using real-world data from the I-24 MOTION testbed, the authors empirically characterize the instability of standard methods and demonstrate that their approach simultaneously enhances robustness to perturbations and achieves a 48% improvement in predictive accuracy over conventional static calibration. This work opens the door to more trustworthy traffic simulations for planning and real-time control applications.