Abstract
We consider the traffic light control problem for a single intersection modeled as a stochastic hybrid system. We study a quasi-dynamic policy based on partial state information defined by detecting whether vehicle backlogs are above or below certain controllable thresholds. Using Infinitesimal Perturbation Analysis (IPA), we derive online gradient estimators of a cost metric with respect to these threshold parameters and use these estimators to iteratively adjust the threshold values through a standard gradient-based algorithm so as to improve overall system performance under various traffic conditions. Results obtained by applying this methodology to a simulated urban setting are also included.
1. Introduction
Traffic Light Control (TLC) remains a critical challenge in urban planning, balancing the need for high throughput with the reality of stochastic vehicle arrivals. Traditional systems typically fall into two categories: static control, which uses fixed green/red cycles, and dynamic control, which reacts to individual vehicle data. While static systems fail to adapt to traffic spikes, full dynamic control often requires high-fidelity sensor infrastructure that is cost-prohibitive for many municipalities.
This paper proposes a quasi-dynamic control approach that bridges this gap. By utilizing partial state information—specifically, whether a queue has crossed a predefined threshold—controllers can adapt to real-time conditions using simple sensors like inductive loops. The central problem addressed is the optimization of these thresholds to minimize average queue lengths across the intersection.
2. Methodology
2.1 Stochastic Hybrid System Modeling
The intersection is modeled as a Stochastic Hybrid System (SHS), capturing both the discrete events of signal switching and the continuous flow of vehicle backlogs. To simplify the mathematical complexity while maintaining aggregate accuracy, the researchers employed a Stochastic Flow Model (SFM), treating traffic as a fluid rather than individual particles.
2.2 Infinitesimal Perturbation Analysis (IPA)
The core innovation is the application of IPA to derive online gradient estimators. Unlike traditional optimization methods that require a known probability distribution of traffic arrivals, IPA calculates gradients based on a single observed "sample path." This makes the method distribution-agnostic and highly robust for real-world deployment where traffic patterns are unpredictable.
As shown in Figure 4, the thresholds are iteratively adjusted in real-time. The algorithm "slides" toward the optimal performance point by calculating the direction of improvement (the gradient) of the average queue length with respect to the threshold parameters.
3. Experimental Results
The methodology was tested using simulated urban traffic data under various scenarios, including unbalanced traffic (Scenario A) and high-intensity unequal traffic (Scenario B).
3.1 Performance and Convergence
The IPA-driven optimization demonstrated significant improvements over static and initial quasi-dynamic settings. In high-intensity scenarios, the method achieved a 66% reduction in mean queue length compared to unoptimized thresholds. Furthermore, the sequential quasi-dynamic approach outperformed standard static controllers by 51%.
A key finding is the robustness of the convergence. As illustrated in the cost surface plot (Figure 5), the algorithm consistently reaches the same optimal threshold point regardless of the initial configuration, confirming the stability of the IPA-based gradient descent.
4. Discussion and Future Work
While the current model provides a powerful framework for single-intersection optimization, it currently ignores complexities such as turning movements and vehicle acceleration physics. The most significant limitation is the focus on an isolated intersection. Real-world traffic operates as a network where a green light at one node creates a surge at the next. Future research will focus on extending this IPA methodology to a network of interconnected intersections and exploring the simultaneous optimization of both light cycle lengths and queue thresholds.
Summary updated