Abstract
Traffic Signal Control (TSC) is essential for urban efficiency, yet Reinforcement Learning (RL) solutions often struggle with a "deployment gap" due to complex state representations. This paper introduces Efficient Pressure (EP), a simplified metric measuring the difference between average queue lengths of entering and exiting lanes. EP bridges the gap between traditional control theory and modern RL, enabling state-of-the-art performance with reduced computational complexity across real-world datasets.
Introduction: The Deployment Gap in Traffic Control
Traffic Signal Control (TSC) is a balancing act: moving the most vehicles through a network while minimizing wait times. Traditional methods, like fixed-time cycles, are robust but fail to adapt to the "bursty" nature of real-world traffic. Conversely, Reinforcement Learning (RL) has shown promise by learning optimal strategies from data. However, RL-based TSC faces a "deployment gap." These models are often computationally expensive, require long training periods, and rely on complex state representations—such as vehicle GPS coordinates or high-resolution images—that are not always available in existing infrastructure.
The core challenge addressed here is: How can we build an adaptive controller that is simple, cost-effective, and robust enough for real-world deployment?
Methods: Redefining Traffic Pressure
The researchers propose a new metric called Efficient Pressure (EP) and two primary implementations:
- Efficient Pressure (EP): Traditional "Max Pressure" (MP) controllers look at the difference in queue lengths between a specific upstream lane and a downstream lane. EP improves this by calculating the difference between the average queue length of all entering lanes for a movement and the average queue length of all exiting lanes. This "lanes-to-lanes" approach better accounts for real-world driver behavior, such as lane-changing within an intersection.
- Efficient-MP: A traditional, non-learning algorithm that chooses the signal phase with the highest EP at any given moment, requiring zero training.
- Efficient-XLight: A general RL template where the agent is given EP as its primary state instead of a complex list of features. This was applied to popular models like PressLight, MPLight, and CoLight.
The authors also standardized the "environment" by setting a fixed 15-second minimum phase duration and using a 4-phase signal structure.
Key Insights: Representation Over Complexity
A striking finding of the study is that a better representation of the traffic state (EP) is more important than the complexity of the algorithm itself. A simple greedy algorithm using EP (Efficient-MP) often outperformed sophisticated RL models that used inferior state representations.
By averaging across all lanes, EP captures the aggregate pressure of a traffic movement, making it robust to lane-changing and complex intersection geometries. Furthermore, the authors highlight that phase duration is a critical "hyperparameter." They found that a unified 15-second duration allowed even traditional models to outperform recent RL breakthroughs.
Experimental Results
The study utilized five real-world datasets from JiNan and HangZhou, simulated in the CityFlow environment. The authors compared their methods against standard baselines (Fixed-Time, Max Pressure) and state-of-the-art RL models.
- Efficient-MP: Achieved a new state-of-the-art among non-learning methods, reducing travel time by up to 6.12% compared to standard Max Pressure in HangZhou.
- Efficient-XLight: When EP was integrated into the CoLight model (Efficient-CoLight), it became the top performer, reducing average travel time by 6.25% in JiNan 1 compared to the original CoLight.
Conclusion and Future Work
Efficient Pressure demonstrates that simplifying state representation can lead to more effective and deployable traffic control solutions. While the "Basic Protocol" (15s duration, 4-phase) worked exceptionally well for the tested cities, future work will explore its universality. The team aims to incorporate additional data like vehicle speed into the EP calculation and develop adaptive RL algorithms that can dynamically adjust EP weights based on real-time traffic fluctuations.
Summary updated