Context
To avoid traffic conflicts at intersections, traffic signal control allocates green times to various vehicle movements at signalized intersections. However, when poorly optimized, these strategies can lead to severe congestion, increased energy consumption, and higher pollution levels. Optimizing traffic signal strategies across a network is particularly challenging in dense urban areas, where it is crucial to address disruptions caused by congestion, accidents, or equipment failures.
In such urban grid networks, local traffic signal control strategies often result in gridlocks under oversaturated conditions. Perimeter Control (PC) has been proposed as a solution to protect a specific region (the Protected Network) and mitigate the spread of congestion within this region (Li et al., 2021). While existing approaches typically assume a fixed perimeter for the protected region, the real challenge lies in dynamically identifying and adjusting this perimeter in real time as traffic conditions evolve, ensuring a well-fitted and effective traffic flow management. The objective is to avoid over-reacting to gridlock and applying well-suited network protection, when necessary.
Recent advancements have explored the potential of data-driven methods, particularly Deep Reinforcement Learning (DRL) algorithms, for decentralized traffic control systems, demonstrating strong scalability and adaptability (Yu et
al., 2023).
In this context, this internship will focus on leveraging Deep Reinforcement Learning (DRL) with Graph Neural Networks (GNN) to integrate the local road network configuration and interaction in the state description of the agent dedicated to optimising the traffic crossing at one intersection. The goal is to develop a strategy where traffic control systems consider the interactions and impacts of traffic flows over time and across space. The integration of attention mechanisms within the graph modeling, coupled with reinforcement learning, will be explored to assess conditions both at individual intersections and across neighboring areas (Munikoti et al., 2023), creating a more adaptive approach. Moreover, this strategy will be tested as a key element for dynamically determine when a protected zone is required and define its boundaries.Objectives and Activities
The primary goal of this internship is to address the joint problem of traffic signal control and dynamic cordon adjustment for perimeter control, with an emphasis on scalability and resilience. A data-driven, deep learning framework will be considered.
