Optimizing city-scale traffic flows through modeling isolated observations of vehicle movements

Fan Yang, Alina Vereshchaka, Bruno Lepri, Wen Dong

Abstract

The capability of traffic-information systems to sense the movement of millions of users and offer trip plans through mobile phones has enabled a new way of optimizing city traffic dynamics, turning transportation big data into insights and actions in a closed-loop and evaluating this approach in the real world. Existing research has applied dynamic Bayesian networks and deep neural networks to make traffic predictions from floating car data, utilized dynamic programming and simulation approaches to identify how people normally travel with dynamic traffic assignment for policy research, and introduced Markov decision processes and reinforcement learning to optimally control traffic signals. However, none of these works utilized floating car data to suggest departure times and route choices in order to optimize city traffic dynamics. In this paper, we present a study showing that floating car data can lead to lower average trip time, higher on-time arrival ratio, and higher Charypar-Nagel score compared with how people normally travel. The study is based on optimizing a partially observable discrete-time decision process and is evaluated in one synthesized scenario, one partly synthesized scenario, and three real-world scenarios. This study points to the potential of a "living lab" approach where we learn, predict, and optimize behaviors in the real world.

Introduction

Traffic congestion remains one of the most persistent challenges in urban planning. While modern systems can predict delays or optimize traffic signal timings, they often lack a holistic approach that connects real-time observations to individual driver decision-making. This research, published in IEEE Access, introduces a "closed-loop" framework that leverages sparse data from a fraction of vehicles to suggest optimal departure times and routes for an entire city.

The core challenge addressed is twofold: first, accurately inferring the state of a city's traffic when only a small percentage of vehicles (probe vehicles) provide GPS data; and second, translating these noisy observations into actionable advice that reduces overall system congestion rather than simply shifting bottlenecks.

System Architecture and Methodology

The authors treat city-scale traffic as a Partially Observable Markov Decision Process (POMDP). The system architecture is designed to bridge the gap between sparse observations and city-wide optimization through a continuous feedback loop.

The closed-loop system architecture for city-scale traffic prediction and control.

State Estimation via PFSKM

To handle partial observability, the researchers developed the Particle Filter Stochastic Kinetic Model (PFSKM). Unlike standard deep learning models that treat traffic as a generic sequence, the PFSKM explicitly incorporates the road network's topology. By modeling traffic as a series of discrete events (e.g., a vehicle moving between road links), the particle filter can "track" the likely distribution of all vehicles based on the movements of a few probe vehicles.

Visualization of the Particle Filter Stochastic Kinetic Model (PFSKM) tracking vehicle distributions across a road network.

Optimal Control via EM

To determine the best traffic policy, the framework employs an Expectation-Maximization (EM) algorithm. The city's optimal state is treated as a hidden variable, and the system iteratively refines suggestions for departure times and routes to maximize a reward function. This reward function utilizes Charypar-Nagel Scoring, which balances the utility of performing activities (like being at work) against the disutility of travel time and late arrivals.

Evaluation and Key Results

The framework was rigorously tested across five diverse datasets, ranging from fully synthesized towns to massive real-world datasets like Dakar (9 million users) and New York City (1 million daily taxi trips).

Key findings include:

  • Superior Tracking: The PFSKM outperformed deep neural networks (DNNs and RNNs) in tracking accuracy. The authors attribute this to the model's inherent knowledge of road topology, which neural networks must expend significant capacity to learn.
  • Significant Time Savings: In the SynthTown scenario, the framework reduced average trip times from 45.57 minutes to 31.49 minutes.
  • Improved Reliability: Across all datasets, the system consistently improved on-time arrival ratios. In the Berlin dataset, the ratio increased from a baseline of 0.72 to 0.86.
  • Scalability: While standard reinforcement learning algorithms often fail at city-scale, this POMDP approach remained computationally feasible for millions of agents.

Conclusion and Future Directions

This research demonstrates the potential of a "living lab" approach to urban mobility. By moving from open-loop simulations to closed-loop real-time optimization, cities can significantly improve traffic dynamics using existing mobile data.

Future work will focus on addressing the variance inherent in Monte Carlo integration and exploring Inverse Reinforcement Learning to better model human compliance with suggested routes. While computationally demanding, the framework provides a scalable path toward truly intelligent urban transportation systems.

Summary updated

Abstract

The capability of traffic-information systems to sense the movement of millions of users and offer trip plans through mobile phones has enabled a new way of optimizing city traffic dynamics, turning transportation big data into insights and actions in a closed-loop and evaluating this approach in the real world. Existing research has applied dynamic Bayesian networks and deep neural networks to make traffic predictions from floating car data, utilized dynamic programming and simulation approaches to identify how people normally travel with dynamic traffic assignment for policy research, and introduced Markov decision processes and reinforcement learning to optimally control traffic signals. However, none of these works utilized floating car data to suggest departure times and route choices in order to optimize city traffic dynamics. In this paper, we present a study showing that floating car data can lead to lower average trip time, higher on-time arrival ratio, and higher Charypar-Nagel score compared with how people normally travel. The study is based on optimizing a partially observable discrete-time decision process and is evaluated in one synthesized scenario, one partly synthesized scenario, and three real-world scenarios. This study points to the potential of a "living lab" approach where we learn, predict, and optimize behaviors in the real world.
BibTeX
@article{yang2022optimizing,
  author = {Fan Yang and Alina Vereshchaka and Bruno Lepri and Wen Dong},
  title = {{Optimizing city-scale traffic flows through modeling isolated observations of vehicle movements}},
  journal = {IEEE Trans. Intell. Transp. Syst.},
  year = {2022},
  doi = {10.1109/tits.2021.3094758},
}
Figure