Ultra-fast Traffic Nowcasting and Control via Differentiable Agent-based Simulation
Fumiyasu Makinoshima, Yuya Yamaguchi, Eigo Segawa, Koichiro Niinuma, Sean Qian · 2026
Read Ultra-fast Traffic Nowcasting and Control via Differentiable Agent-based Simulation by Fumiyasu Makinoshima, Yuya Yamaguchi, Eigo Segawa, Koichiro Niinuma, Sean Qian (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-25068.txt.
Introduction
Makinoshima and co-authors build a traffic simulator that can be tuned by gradients, not guesswork. City digital twins need fine models fit to live data, but classic simulators are not differentiable. Their agent-based system keeps whole runs smooth enough for gradient methods. It covers single-car choices, random decisions, and car-to-car effects. On the Chicago network it runs over a million vehicles at 173 times real speed. Fitting the last half hour takes 455 seconds. A one-hour forecast takes 21 seconds, leaving time to act.
Catalog record
| CREATOR | Fumiyasu Makinoshima, Yuya Yamaguchi, Eigo Segawa, Koichiro Niinuma, Sean Qian |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 61262 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.25068
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Ultra-fast Traffic Nowcasting and Control via Differentiable Agent-based Simulation Fumiyasu Makinoshima, Yuya Yamaguchi, Eigo Segawa, Koichiro Niinuma, Sean Qian Abstract Traffic digital twins, which inform policymakers of effective interventions based on large-scale, high-fidelity computational models calibrated to real-world traffic, hold promise for addressing societal challenges in our rapidly urbanizing world. However, conventional fine-grained traffic simulations are non-differentiable and typically rely on inefficient gradient-free optimization, making calibration for real-world applications computationally infeasible. Here we present a differentiable agent-based traffic simulator that enables ultra-fast model calibration, traffic nowcasting, and control on large-scale networks. We develop several differentiable computing techniques for simulating individual vehicle movements, including stochastic decision-making and inter-agent interactions, while ensuring that entire simulation trajectories remain end-to-end differentiable for efficient gradient-based optimization. On the large-scale Chicago road network, with over 10,000 calibration parameters, our model simulates more than one million vehicles at 173 times real-time speed. This ultra-fast simulation, together with efficient gradient-based optimization, enables us to complete model calibration using the previous 30 minutes of traffic data in 455 s, provide a one-hour-ahead traffic nowcast in 21 s, and solve the resulting traffic control problem in 728 s. This yields a full calibration–nowcast–control loop in under 20 minutes, leaving about 40 minutes of lead time for implementing interventions. Our work thus provides a practical computational basis for realizing traffic digital twins. Introduction Rapid global urbanization represents a pivotal moment in human history, with urban dwellers now constituting the nearly half of the world’s 8.2 billion population [ 1 ] . The rapid urbanization has elevated the traffic congestion management to a critical global challenge, as it has significant effects on economic productivity [ 2 , 3 ] and environmental and public health [ 4 , 5 ] , beyond mere inconvenience. Given the difficulty and potentially irreversible impacts of implementing traffic policies in real-world, traffic simulation has become an indispensable tool for identifying effective strategies within a virtual space in computers. This has led to the development of various simulation models [ 6 , 7 , 8 ] simulating vehicle movements at a fine scale with micro or mesoscopic models, which are now widely applied to inform practical policy-making. While simulation is a powerful tool for policy-making, its real-world effectiveness hinges on calibration, the process of aligning the model with empirical data. A well-calibrated simulator enables the exploration of reliable and effective strategies [ 9 ] . This calibration process, however, becomes a critical bottleneck for fine-grained traffic simulation models. Because of their inherent stochasticity and complex vehicle interactions, such models are non-differentiable and must typically be calibrated using gradient-free optimization algorithms [ 10 , 11 , 12 , 13 , 14 ] , which are notoriously inefficient and require vast numbers of simulation runs. As the spatial scale of the simulation expands, the number of calibration parameters grows rapidly, further compounding the problem and making the optimization computationally intractable. As a result, the practical application of calibrated microscopic models remains confined to small-scale scenarios with only a few dozen parameters [ 15 ] , falling short of the needs for addressing complex, large-scale urban systems. On the other hand, real-time traffic observation data for calibration is now becoming readily available from various observation channels [ 16 ] . This situation underscores the urgent need for advanced simulation technologies capable of translating this wealth of data into actionable insights for timely transportation policy making. The emerging paradigm of differentiable simulation presents a promising pathway to overcome this challenge. Effectively using automatic differentiation developed mainly for machine learning [ 17 ] , the differentiable simulations calculate the dynamics of the equation-based white-box models while enabling the access to the gradient information of the simulated process and efficient gradient-based optimization [ 18 ] . Differentiable simulation techniques first demonstrated their effectiveness for rapid parameter calibration in the field of physical simulation [ 19 , 20 , 21 ] . More recently, their application has been extended to social simulations [ 22 , 23 , 24 , 25 , 26 ] , where comparable benefits have been reported. In traffic engineering, while efforts to apply this paradigm to parameter calibration are emerging, they have been largely confined to macroscopic, continuous models [ 27 , 28 , 29 , 30 , 31 , 32 ] . The few pioneering works on microscopic models [ 22 , 23 , 33 , 34 ] have come at the cost of simplification, often omitting complex behaviors such as agents’ stochastic decision-making and their interactions with other agents. The core challenge lies in consistently differentiating stochastic choice of vehicles, vehicle transfers between road links, and vehicle-to-vehicle interactions. Consequently, developing a differentiable fine-grained traffic simulator with the fidelity to capture traffic dynamics in complex, real-world networks has yet to be achieved. Here we introduce an end-to-end differentiable, agent-based model that simulates the dynamics of over one million vehicles in real-time. Its design is grounded in traffic flow theory and employs novel computational techniques to maintain full differentiability throughout the simulation, even with complex stochastic behaviors of vehicles. The model is fully tensorized to leverage massive parallelization on hardware accelerators, achieving simulation speed significantly exceeding real-time. We demonstrate its capability on a large-scale, real-world road network with over 10,000 calibration parameters. This exceptional speed, combined with efficient gradient-based optimization, enables model calibration, nowcasting, and intervention planning to be completed several times faster than real time, so that the full pipeline finishes with ample lead time for implementing interventions, thereby realizing a practical, operational traffic digital twin for a city. By bridging the gap between traditional simulators and the vast amount of available traffic data, our work paves the way for next-generation intelligent transportation systems to address pressing societal challenges. Results Differentiable agent-based traffic simulation We developed an end-to-end differentiable agent-based traffic simulator for real-time traffic nowcasting and control (see Methods and Supplementary Information A–D for details). The overall framework is illustrated in Fig. 1 . Given the initial traffic state and two types of parameters, namely behavioral model parameters \theta_{b} and environmental parameters \theta_{e} , the simulator computes the traffic state over a time horizon of length T . While the environmental parameters, such as road pricing, can be specified objectively and, in some cases, decided by policy makers, the behavioral parameters governing how travelers perceive the environment and make decisions are not directly observable and must be inferred from data. A single forward step of our simulator, denoted by \phi , can be viewed as the composition of a position update function f and a vehicle transfer function g . The position update function f advances vehicle positions over one time step based on traffic flow theory, whereas the vehicle transfer function g governs probabilistic next-link choices and stochastic merging behavior. In this study, we construct differentiable versions of these processes while still simulating vehicle movements in a manner that is faithful to traffic flow theory, without ad-hoc simplifications of inter-agent interactions or stochasticity (see Methods for details). Because these processes are differentiable, the composite function \phi and the simulated trajectories \phi^{T} over T time steps are end-to-end differentiable. Consequently, the gradient of a loss function, which quantifies the discrepancy between the simulated traffic state \hat{y} and the observed