Traffic Signal Control
Distributed traffic signal control with multi-agent reinforcement learning
Autonomous driving is the future, however autonomy will only be beneficial to society if vehicles and infrastructure can coordinate with each other to reduce traffic congestion and improve travel times. Intelligent decision making on the infrastructure side (e.g., traffic signals) is of paramount importance to regulate traffic flow. However, centralized control of the traffic infrastructure is infeasible, and in large metropolis that can count upwards of thousands of junctions, and decentralization will be necessary.
This research was aimed at intelligent decision making for traffic signal control based on the traffic conditions at a junction (i.e number and speed of incoming vehicles, queue lengths,etc,) as well as the conditions of adjacent junctions. We use distributed reinforcement learning approaches to tackle the coordination problem. In particular, how do agents learn to coordinate traffic flow with their neighbors such that the ensuing behavior of an agent to cooperate results in citywide traffic flow improvements.
Checkout out our recent full paper titled “SocialLight: Distributed Cooperation Learning towards Network-Wide Traffic Signal Control” accepted to AAMAS 2023 for more details.