Abstract
Urban mobility is a complex, partially observable system where only a fraction of vehicle movements are typically tracked. This paper presents a robust framework for city-scale traffic optimization that moves beyond simple prediction. By modeling traffic as a Partially Observable Discrete-Event Decision Process (PODEDP), the authors demonstrate how suggesting optimal departure times and routes to a small percentage of drivers can significantly reduce congestion and improve arrival punctuality across entire cities.
Introduction
Urban traffic management has traditionally been divided into two separate domains: prediction (forecasting jams) and control (adjusting infrastructure like traffic lights). While modern navigation apps generate vast amounts of "floating car data" (GPS traces), this information is rarely integrated into a closed-loop system that actively influences driver behavior for network-wide benefit.
The primary challenge lies in the nature of traffic itself—it is a partially observable, highly stochastic system. Traditional optimization methods often struggle to scale to the millions of agents in a metropolis, while black-box models like Deep Neural Networks frequently ignore the underlying physics of road networks. This paper proposes a framework that bridges this gap by using sparse data to inform a global optimization strategy.
Methodology
The proposed framework addresses urban congestion by treating the city as a Partially Observable Discrete-Event Decision Process (PODEDP). The system operates in a continuous loop of observation, state estimation, and optimization.
Stochastic Kinetic Modeling and Tracking
Traffic dynamics are modeled using a "stochastic kinetic model" where discrete events, such as a vehicle transitioning between road segments, drive the system. To estimate the global traffic state from noisy observations (e.g., GPS data from only 10% of vehicles), the researchers implemented a Particle Filter (PFSKM). This "observer" reconstructs the belief state of the entire network by leveraging known road topology and queuing dynamics, allowing the model to "fill in the gaps" where data is missing.
Optimization and Control
The "controller" utilizes an Expectation-Maximization (EM) algorithm to generate travel advice. Unlike standard shortest-path algorithms, this framework optimizes for the Charypar-Nagel score, a utility function that balances travel time with the value of performing activities at the destination and penalties for late arrivals. This allows the system to suggest not just the best route, but also the optimal departure time to effectively smooth out peak demand.
Datasets and Evaluation
The framework was benchmarked across five diverse scenarios to prove its versatility:
- SynthTown: A controlled synthetic scenario for detailed analysis.
- Berlin: A massive synthetic dataset representing 3 million inhabitants and 24,000 road links.
- Santiago de Chile: Large-scale modeling of 6 million inhabitants.
- Dakar: Real-world movement data derived from Call Detail Records (CDR) of 9 million mobile users.
- NYC Taxicab: Real-world data encompassing 1 million daily trips across 11,000 road links.
Key Results
The PFSKM-based framework consistently outperformed baseline re-planning and state-of-the-art co-evolutionary algorithms in both tracking accuracy and network performance.
- Tracking Accuracy: The PFSKM achieved superior reconstruction of city-wide traffic from limited probe data, outperforming standard Deep Learning models (RNNs/DNNs) which lack physics-informed constraints.
- Efficiency Gains: In the SynthTown scenario, average trip times were reduced from 45.57 to 31.49 minutes.
- Punctuality: In the Berlin dataset, the on-time arrival ratio improved from 72% to 86%.
- Scalability: The particle-based approach remained computationally performant on the massive Berlin and Santiago datasets, where Reinforcement Learning models often fail to converge due to high dimensionality.
Discussion and Future Work
The research highlights that physics-informed models can outperform black-box approaches in complex urban environments. By continuously correcting its internal model based on real-time observations, the system remains robust to unexpected surges or delays.
Future work may address the variance inherent in Monte Carlo integration and explore user compliance dynamics, as the system's effectiveness depends on drivers following suggestions. Additionally, the authors suggest using Inverse Reinforcement Learning to personalize utility functions, better reflecting diverse driver preferences beyond simple time minimization.
Summary updated