Reinforcement Learning Strategies for Intelligent Traffic Signal Control

Reinforcement learning (RL) has emerged as a transformative approach to managing urban intersections through adaptive, data-driven decision-making. By modelling each traffic signal controller as an agent that learns from real-time vehicle flows and network feedback, RL methods can dynamically adjust signal phases to minimise delay, stops and emissions without explicit traffic flow models. Early implementations employed classical Q-learning for isolated junctions, while contemporary schemes leverage deep reinforcement learning (DRL) to handle high-dimensional state spaces encompassing vehicle counts, queue lengths and connected vehicle information. Multi-agent frameworks extend this paradigm by treating adjacent intersections as cooperating learners, enabling coordinated strategies that improve corridor throughput. Such intelligent control systems can integrate with connected and autonomous vehicles to harmonise speeds, prioritise emergency services and adapt to demand fluctuations, offering scalable solutions for sustainable urban mobility worldwide.

Research from Nature Portfolio

Recent work has demonstrated a city-wide, scalable re-timing system that relies solely on a small sample of vehicle trajectories rather than fixed detectors. By introducing a probabilistic time-space diagram under Newellian coordinates, researchers reconstructed spatio-temporal traffic states and formulated optimisation algorithms for signal parameters, achieving reductions of up to 20 % in delay and 30 % in stops in a real-world urban trial. This detector-free framework highlights the potential of trajectory-based RL strategies to deliver cost-effective, sustainable signal control at scale, while maintaining optimality guarantees under stochastic traffic dynamics.

Reinforcement Learning Strategies for Intelligent Traffic Signal Control publication trend

The graph below shows the total number of articles in reinforcement learning strategies for intelligent traffic signal control across all publications each year (not limited to Nature Index journals).

Technical terms

Reinforcement learning: A machine-learning paradigm in which an agent learns to choose actions that maximise cumulative rewards through trial and error in an environment.

Deep reinforcement learning (DRL): An extension of RL that uses deep neural networks to approximate value functions or policies, enabling handling of high-dimensional inputs.

Multi-agent system: A framework where multiple learning agents interact within a shared environment, coordinating or competing to achieve individual or collective objectives.

Signal phase optimisation: The process of adjusting the duration and sequence of traffic signal states to improve performance measures such as delay or throughput.

Probabilistic time-space diagram: A stochastic model linking vehicle trajectory data to queue dynamics, used to infer spatio-temporal traffic states for control optimisation.

References

1. Traffic light optimization with low penetration rate vehicle trajectory data. Nature Communications (2024).
2. Reinforcement learning in urban network traffic signal control: A systematic literature review. Expert Systems with Applications (2022).
3. Multi-Agent Deep Reinforcement Learning to Manage Connected Autonomous Vehicles at Tomorrow's Intersections. IEEE Transactions on Vehicular Technology (2022).
4. Deep Reinforcement Learning for Traffic Signal Control: A Review. IEEE Access (2020).