Mean-Field Game Theory for Adaptive Predictive Traffic Signal Control . A Hybrid AI-Mathematical Framework for Equilibrium-Based Optimisation at Urban Intersections

2024- Ongoing Academic research Multiple National Awards

Video

About

My inspiration

Urban traffic congestion continues to challenge growing cities, yet many modern traffic management systems depend on expensive infrastructure upgrades, additional sensors, or large-scale data collection. For this project, I wanted to investigate whether meaningful improvements could be achieved without changing the existing traffic infrastructure.

Introduction

To explore this, I developed a hybrid adaptive traffic control framework combining Mean-Field Game Theory with Reinforcement Learning. Rather than modelling traffic as thousands of independent vehicles, the framework represents it as a continuous large-scale flow, allowing signal decisions to optimise the overall behaviour of the junction instead of reacting to individual cars.

Handwritten traffic signal modelling notes — Early research notes converting SCATS detector volumes into arrival rates, density, equilibrium, and green-time decisions.

Method

This network-level perspective provides a more scalable approach for complex urban environments while remaining computationally practical. A key design objective throughout the research was real-world deployment. The controller was built to operate using existing traffic sensors, making it a low-cost, low-risk solution that could integrate into current traffic systems without additional hardware.

Five-stage adaptive traffic control pipeline — The five-stage controller flow: sensor inputs, rate prediction, Mean-Field Game solver, reinforcement-learning safety net, and adaptive signal control.

Results

The framework was evaluated on a digital recreation of the Saint Patrick's Cathedral junction in Dublin using publicly available traffic data. Compared against the existing signal strategy, it achieved statistically significant reductions in average waiting time, queue length, and spillback probability, with all findings validated through ANOVA and paired two-tailed hypothesis testing.

Traffic junction simulation comparing signal states — Virtual junction simulation comparing baseline signal timing with adaptive control behaviour before moving toward real-world deployment.

Recognition pathway

Beyond the technical research, the project was recognised by the University of Limerick, PATCH, Stripe YSTE, and Intelligent Transport Ireland, where it was later developed into a business concept during the Stripe YSTE Bootcamp. The research is currently being extended under the mentorship of Amirreza Kandiri at University College Dublin towards academic publication.

Dublin City Council SCATS traffic volume dataset and junction location metadata — Public SCATS traffic volume data used to ground the simulation in a real Dublin junction.

Box plots comparing waiting time, queue length, and spillback before and after MFG-based control — Statistical comparison showing reduced waiting time, queue length, and spillback variability.

Summary of traffic signal control results before and after adaptive MFG control — A compact summary of the before-and-after gains across the core traffic metrics.