Traffic lights are a familiar example of a system that looks local but behaves as a network. A decision that clears one intersection can simply push a queue towards the next. That coordination problem becomes harder as traffic demand changes across a city, which is why researchers have increasingly investigated whether individual intersections can learn to adapt while still contributing to better network-wide flow.
A new peer-reviewed study in Scientific Reports tested a decentralised multi-agent reinforcement learning framework for this problem. In simulation, the proposed system reduced average travel time by 26.9%, cut average queue length by 46.4%, improved the congestion index by 33.3% and increased network throughput by 27.4% compared with the study’s benchmark approaches.
Turning intersections into learning agents
The framework is called Decentralized Agent-Based Multi-Agent Reinforcement Learning Traffic Optimization, or DAB-MARLTO. Rather than placing all traffic decisions under one central controller, the researchers modelled vehicles, road segments and traffic signals as interacting autonomous agents within a directed road network.
Each signalised intersection was formulated as a Markov Decision Process. In practical terms, this means an intersection observes the traffic state available to it, chooses a signal action, receives feedback about the resulting traffic conditions and gradually learns which choices produce better outcomes. The study used decentralised tabular Q-learning for this process.
Traffic demand was represented using a Poisson arrival process, while vehicle movement incorporated car-following, lane-changing and queue models. The researchers implemented the framework in the Simulation of Urban Mobility environment, commonly known as SUMO, and used its Traffic Control Interface to execute control policies.
This design matters because decentralisation changes where the computational burden sits. A central optimisation system can use a broad view of a network, but its decision problem grows as more intersections and vehicles are added. A decentralised system instead allows local controllers to respond to their immediate conditions, although this creates a different challenge: local decisions still need to produce sensible behaviour across the wider network.
Four benchmark strategies were compared
The researchers evaluated DAB-MARLTO against fixed-time signal control, rule-based adaptive control and centralised traffic optimisation. This gave the study benchmarks ranging from conventional predetermined timing to approaches that already react to traffic conditions or optimise across the network.
Performance was not judged on a single measure. The analysis considered average travel time, network throughput, average queue length, a congestion index, fuel consumption and emission intensity. This broader set of measures is important because a traffic controller can appear successful under one metric while moving costs elsewhere. For example, clearing one queue quickly is less useful if it produces longer journeys or worsens congestion downstream.
Across the reported comparisons, the decentralised learning framework produced the strongest overall traffic performance. Average travel time fell by 26.9%, while average queue length was 46.4% lower. The congestion index improved by 33.3%, and network throughput increased by 27.4%.
These are substantial differences within the simulated environment. The queue-length result is particularly notable because queues are both an outcome of congestion and a mechanism through which congestion propagates. A long queue can block turning movements, interfere with neighbouring intersections and reduce the effective capacity of the surrounding network.
Why decentralised learning may help
Fixed-time traffic signals are built around predetermined cycles and therefore cannot fully respond to unexpected changes in demand. Rule-based adaptive systems can respond, but their behaviour remains constrained by the rules designers specify. Centralised optimisation can coordinate a larger system, yet it may face increasing computational and communication demands as the network grows.
Multi-agent reinforcement learning takes another route. Each controller can learn from repeated interaction with its environment instead of relying entirely on a fixed schedule or manually specified response. In the study’s framework, that learning is decentralised, allowing decisions to be made locally rather than requiring every action to be calculated by a single central optimiser.
The results suggest that this combination can improve the way capacity is used under changing simulated demand. Higher throughput means the network moved more vehicles over the evaluation period, while shorter queues and travel times indicate that this did not simply come from pushing more traffic into longer delays elsewhere.
The results are simulated, not a city deployment
The size of the reported improvements should not be interpreted as a prediction that installing the framework in a real city would automatically cut travel times by 26.9%. The experiments were conducted in SUMO, where traffic demand, driver behaviour, road interactions and control conditions can be represented systematically but cannot reproduce every source of uncertainty found on real streets.
Real traffic networks contain pedestrians, cyclists, public transport priority, emergency vehicles, incidents, roadworks, unusual driving behaviour, imperfect sensors and communication failures. Deployment would also have to satisfy safety rules and integrate with existing signal infrastructure. These practical constraints can alter both the information available to an agent and the actions it is allowed to take.
Tabular Q-learning also becomes more demanding as the number of possible states and actions increases. That makes scalability an important question when moving from a controlled network model to a large, heterogeneous urban system. A decentralised architecture reduces dependence on one central decision-maker, but decentralisation does not by itself remove the need for robust coordination and validation.
The study therefore provides evidence about comparative performance inside its simulation framework, rather than evidence of realised travel-time savings on public roads. Field trials would be needed to establish whether the improvements persist with noisy sensors, real drivers, different intersection geometries and rapidly changing demand.
A useful direction for increasingly complex road networks
Despite those limitations, the work illustrates why decentralised traffic control is attracting research attention. Urban traffic is inherently distributed: congestion emerges from many vehicles and intersections interacting at once. Giving individual intersections some capacity to learn locally may offer a way to respond at the same distributed level.
The strongest contribution of the study is therefore not simply one percentage improvement. It is the simultaneous movement of several traffic indicators in the same direction. Travel times and queues declined while throughput increased, suggesting that the simulated system was improving network flow rather than optimising one isolated measure.
Whether those gains survive outside simulation is the next, harder question. Testing across more varied networks and ultimately under real traffic conditions will be essential before the approach can be treated as a deployable alternative to established adaptive signal systems.
Source Information
Study: A decentralized multi-agent-based optimization framework for intelligent traffic systems
Author: Mohammed Salem Basingab
Journal: Scientific Reports
Published: 2 October 2026
DOI: 10.1038/s41598-026-74078-2
