Smart Signal Control Using Reinforcement Learning to Ease Urban

Traffic

Sivakumar Ponnusamy

, M. L. M. Prasad

, G. Swarnalakshmi

, S. Karthikeyan

S. K. Lokesh Naik

and Moses Raja A.

Department of Computer Science and Engineering, K.S.R. College of Engineering, Tiruchengode, Namakkal, Tamil Nadu,

India

Department of CSE (AI&ML), Joginpally BR Engineering College, Moinabad, Hyderabad, Telangana, India

Department of Mechanical Engineering, New Prince Shri Bhavani College of Engineering and Technology, Chennai,

Tamil Nadu, India

Department of Electrical and Electronics Engineering, J.J. College of Engineering and Technology, Tiruchirappalli, Tamil

Nadu, India

Department of Computer Science and Engineering, MLR Institute of Technology, Hyderabad, Telangana, India

Department of MCA, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Keywords: Reinforcement Learning, Adaptive Traffic Signals, Urban Congestion, Multi‑Agent Systems, Vehicle Flow.

Abstract: This study offers an innovative proposal for the management urban traffic through the implementation of

reinforcement learning for adaptive traffic signal systems to overcome the traffic jam problem and increase

vehicle flow. The model, based on deep learning algorithms and real-time data, dynamically modifies timing

in accordance with traffic conditions, facilitating smooth and hang-up-free traffic. Compared to classical static

or rule-based methodologies, the approach is flexible enough to adapt to changing patterns and is shown that

outperforms in simulations based on realistic urban networks. In addition, the system uses multi-agent

cooperation methodology to optimize signal coordination across intersections, thus being scalable and

responsive. The results show that the learning instantiation can lead to significant gains in traffic efficiency,

delay reduction, and emission mitigation, demonstrating the potential of reinforcement learning as a

promising approach to enhance intelligent transportation systems.

1 INTRODUCTION

Congestion in the urban areas is a perennial challenge

in fast-growing cities causing higher travel time,

more pollution, inefficiency and economic loss.

Conventional traffic signal control systems, which are

time of day or "rule" based, cannot fully adapt to the

dynamic and often random flows of real-time traffic.

With the popularity of intelligent transportation

systems, an increasing number of researchers are

endeavouring to exploit machine learning, especially

reinforcement learning, for adaptive traffic control

protocols that can learn optimal traffic control

strategies incrementally. Reinforcement learning is a

strong framework for decision making in complex

environments and is capable of providing traffic

signal systems that learns about their environment

from real-time traffic. We propose a traffic signal

control system mobilized by deep reinforcement

learning to dynamically adjust signals based on traffic

flow and intersection conditions. The method

increases coordination among signals, and

consequently throughput and decreases global

congestion by representing intersections as agents in

a multiagent model. Finally, both the learning-based

traffic flow estimation and control manifest the

vision of smart city, and potentially provide a

scalable, data-driven solution to challenges that are

faced in the era of mobility.

102

Ponnusamy, S., Prasad, M. L. M., Swarnalakshmi, G., Karthikeyan, S., Naik, S. K. L. and A., M. R.

Smart Signal Control Using Reinforcement Learning to Ease Urban Trafﬁc.

DOI: 10.5220/0013858100004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

102-108

ISBN: 978-989-758-777-1

2 PROBLEM STATEMENT

Traffic in urban communities has been getting worse

over time as a result of old style traffic signal systems

that are too often unbendable and unable to respond

to the fluidity of everyday traffic. The systems do not

have any intelligence to adapt dynamically to

changes in vehicle flow, with the consequence that

significant inefficiencies result, such as long waiting

times, waste of fuel, and pollutants being released

into the environment. Even in the presence of traffic

infrastructure improvements, current operational

tools fail to effectively respond to the sheer size and

heterogeneity of urban mobility. An intelligent

adaptive road traffic signal control system is urgently

needed, capable of learning and adapting to real-time

traffic situations so as to maximize the motion of

vehicles while reducing gridlock in complex road

networks.

3 LITERATURE SURVEY

The application of reinforcement learning to traffic

signal control has received considerable attention

over the last decade as it overcomes shortcomings of

conventional systems using adaptive, data-driven

approaches. Rafique et al. (2024) presented a turn-

based and time-based RL method for enhancing the

traffic signal actions, and alleviates the congestion in

simulated scenarios. Similarly, Swapno et al. (2024)

implemented dynamic intersection control using deep

Q-learning and reported significant reductions in

delays and queue lengths of vehicles. Li et al. (2024)

investigated the adaptive signal control in the

presence of mixed traffic, where CTAVs operate

along mission with CAVs, and reported improved

coordination and reduced intersection delays. Fan et

al. (2025), to make cooperative control of signals

scalable, introduced sparse deep reinforcement

learning by sharing knowledge. Bie et al. (2024)

studied the problem of variable intersection difficulty

in the context of multi-agent systems for a more

localized and economic control in a variety of

network structures. In improving learning depth and

convergence, Li et al. (2024) devised an enhanced

deep RL model and surpassed a higher level of real-

time traffic condition adaptation. Qi et al. (2022)

developed a kind of ensemble models for RL,

enhancing robustness either in per-episode or

fluctuating environments. Wang et al. (2023) Design

UniTSA – A Universal V2X-Based Reinforcement

Learning Framework to Promote its Deployment in

Connected Environments. Peng et al. {2023})

proposed heterogenous agent collaboration using

GNN-based MARL for traffic guidance around

intersections. Busch et al. (2023) proposed a

combination of RL model for vehicles speeds and

signals control based on throughput and safety. Zhu

et al. (2021) used policy-based RL to reduce

decision-making complexity with high efficiency.

Further, Xie et al. (2022) stressed decentralized,

schedule-based control for scalable urban

coordination. Wang et al. (2024) verified practical

operability of RL approaches with the real vehicle

trajectory data. Feng et al. investigated dense RL for

AV safety testing. (2023), and considering

frameworks applicable to traffic systems. Yan et al.

(2011) simulated statistically realistic learning

validation settings. The previous seminal work of

Liu et al. (2009) and El-Tantawy et al. (2013) first

described the queue estimation and MARLIN-ATSC

control. Taken together, studies in this thesis confirm

the viability and versatility of reinforcement learning

for intelligent traffic signal control, but also reveal

existing and persistent real-world operational,

scalability and cross-agent coordination problems.

4 METHODOLOGY

The proposed approach is based on deep

reinforcement learning (DRL) and is aimed at

establishing an intelligent decentralized traffic signal

control system for dynamic city traffic. The main

idea is to simulate each cross-traffic intersection as

self-actuated intelligent agent in multi-agent system.

These agents are constantly monitoring their

environments (e.g., numbers of vehicles, lengths of

queues, current signal phases, and the durations of

previous waits) as high-dimensional state vectors.

The environment is addressed as Markov Decision

Process (MDP), in which at discrete time slots, the

agent chooses a behaviour (i.e. a change in the traffic

light phase) and is rewarded in terms of traffic

performance. Reward function the reward function is

carefully designed as a compromise between several

conflicting objectives including minimizing the

waiting time and stopping frequency of vehicles,

reducing queue lengths, and indirectly optimizing

fuel consumption and emissions.

Smart Signal Control Using Reinforcement Learning to Ease Urban Trafﬁc

103

Table 1: Agent Configuration.

Component Configuration

Algorithm Deep Q-Network (DQN)

State Representation Queue length, Phase time

Action Space Signal Phase Change

Reward Metrics Delay, Queue, Stops

Learning Rate 0.001

To accomplish efficient policy learning in such a

complex environment, every agent uses a DQN that

can approximate the optimal action-value function

using deep neural networks. The networks are trained

using experience replay and target network updates

for convergence. The agents are trained with

simulated rollouts in a high-fidelity traffic simulator

such as SUMO (Simulation of Urban Mobility) with

realistic traffic composition, which gives rich,

continuous feedback. The policy is gradually trained

by repeatedly entering the simulated environment,

associating traffic conditions with optimal phase

decisions.

Table 1 shows the Agent Configuration.

Figure 1: Deep Reinforcement Learning-Based Traffic

Signal Control Workflow.

And, a MARL framework is introduced to solve the

coordination problem among multiple intersections

and to avoid the problem of conflicting actions. This

setting provides methods for sharing partial

observations or cooperating with shared rewards or

policies. At intersections with a higher level of

interaction, agents communicate information about

their intentions or situational summaries which

contributes to the global system coherence and traffic

flow. In the testing phase, the learned model is tested

under different traffic densities and conditions (e.g.,

rush hours, accidents, and abnormal events), so as to

guarantee the generalization of the system to unseen

scenarios. Performance is compared with the optimal

performance achieved by the fixed-time and actuated

signal control, and a number of key performance

measures including average delay, throughput,

intersection utilization and total emissions are used.

This all-encompassing approach that integrates

DQN with MARL and traffic simulation guarantees

that the system is scalable, robust to sudden traffic

dynamics and adaptive to the increasing pressure

placed on urban transportation networks. It is an

important move toward a real-time data-driven

intelligent traffic signalling system that may be used

in smart cities of the future. Figure 1 shows the Deep

Reinforcement Learning-Based Traffic Signal

Control Workflow.

5 RESULTS AND DISCUSSION

The system of adaptive traffic signal control based on

deep reinforcement learning (DRL) greatly improved

traffic management efficiency in simulated urban

areas. The system was evaluated in different traffic

conditions, including hours of peak traffic with high

volumes and hours between peaks with lower flows.

In all cases the developed model provided improved

performance to conventional fixed-time and actuated

control strategies. Results showed that the DRL

system was able to reduce average vehicle delay

around 35% when compared to fixed-time systems,

and 20% when compared to the actuated signal

control. These improvements were even greater

during peak hours when the model was successfully

able to adjust to surges of traffic and reduce the

amount of time vehicles idled at intersections.

Figure

2 shows the Average Traffic Delay Comparison.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

104

Figure 2: Average Traffic Delay Comparison.

Furthermore, the signalized intersection average

queue lengths were significantly reduced by up to

30%, indicating an increased vehicular flow and

better road space utilization. Reduced queue length

means that system effectively reduced the overflow

effects that are caused by spill back - especially in

crowded intersection and substantially contributes to

urban congestion. Smoother driving profiles with

less jerky deceleration and acceleration were also

implemented by minimizing the number of vehicles

stops. The net result of this was a more even and

consistent traffic flow, both of which are important

in terms of travel time predictability and driver

satisfaction.

Table 2 shows the Performance Metrics

Comparison.

Figure 3: Reward Convergence Over Training Episodes.

Besides, a great advantage of our proposed model

is that we proposed MARL architecture, which

connected one intersection with its neighbours. This

coordination made sure that phase decisions did not

battle against each other and that signal timing was

coordinated for several intersections at the same

time. It was a process that resulted in network-wide

optimization that eliminated bottlenecks and took

vehicles off of the road as a constant flow,

particularly on arterials. The reward function

employed by our model giving preference to

minimising delay, maximising the positive queue

length, and having a small number of stops proved

effective at steering the learning into lift off policies

that balance efficiency against fairness in vehicle

flow.

Figure 3 shows the Reward Convergence Over

Training Episodes.

Figure 4 shows the Fuel

Consumption Across Control Types.

Figure 4: Fuel Consumption Across Control Types.

Table 2: Performance Metrics Comparison.

Metric Fixed-

Time

Control

Actuated

Control

Proposed

stem

Average

Delay (s)

45.2 38.4 27.6

Queue

Length

(vehicles)

15.6 12.3 8.9

Stops per

Vehicle

2.8 2.1 1.2

Throughp

(

)

74.5 79.6 89.3

Environmental footprint is with respect to the

DRL-based visited system another points of interest.

Given the reduced stops and steadier acceleration

profiles, the system resulted in measurable savings in

fuel and emissions. This places the model as a

mobility-efficient tool, and not only, also sustainable

and respectful with the urban environment purposes.

The possibility to connect the system with connected

vehicle applications makes the system even more

scalable and future proofed, which can be used to

control the traffic more proactively by vehicle-to-

infrastructure communication.

Table 3 shows the

System Performance under Multi-Agent vs. Single-

Agent.

Smart Signal Control Using Reinforcement Learning to Ease Urban Trafﬁc

105

Table 3: System Performance Under Multi-Agent Vs.

Single-Agent.

Setup

Ave

rage

Dela

(

)

Queue

Length

(vehicles)

Stop

Frequ

ency

Overall

Efficiency

(%)

Single

ent

32.5 10.4 1.6 82.1

Multi-

Agent

27.6 8.9 1.2 89.3

However, the system efficiency is dependent upon

some factors despite these good results. Accurate

real-time traffic data is crucial for both training and

deployment. Learning of the system is

computationally demanding and entails a training that

must be carried out in a very well performing

simulation environment, particularly on larger

networks. Moreover, the reward function needs to be

delicately designed such that it does not lead to races

among traffic streams by flocking all resources to a

single stream. Although the system emerged as robust

in simulation, practical issues including hardware

integration, data transmission latency, and public

policy regulation would need to be addressed for real-

world deployment.

Table 4 shows the Traffic

Conditions Used for Testing.

Table 4: Traffic Conditions Used for Testing.

Scenario

Vehicle

Rate

(veh/hr)

Signal

Coordination

Low

Traffic

200 Minimal

Moderate

Traffic

500 Moderate

Peak Hour 1000 High

Incident 800 High

Variable

Flow

300-900 Dynamic

In conclusion, our experimentation has firmly

established the feasible to apply RL to the problem of

urban traffic signal control, and provides an

intelligent and scalable one solution for today's urban

congestion problem. The method indicates good

potential for practical application, especially with the

development of smart city infrastructure and vehicle-

to-vehicle/vehicle-to-infrastructure (V2X)

communication technology.

Table 5 shows the

Emission Reduction Statistics.

Table 5: Emission Reduction Statistics.

Control

Type

Fuel

Consumpti

on (L/km)

CO₂

Emissio

(g/km)

Stop Count

per Vehicle

Fixed-Time 0.12 280 2.8

Actuated 0.10 240 2.1

RL-Based 0.07 190 1.2

6 CONCLUSIONS

This paper offers a detailed investigation of an

adaptive traffic signal controller on the basis of deep

reinforcement learning, aiming at mitigating the

urban traffic congestion and flow inefficiency

problems that have lingered. By looking past, the

conventional implementation strategies for traffic

control, the model described here illustrates how

intelligent software systems, trained by actual

interaction with the environment, could provide much

improved vehicular flow through traffic and alleviate

the effects of traffic congestion. The fact that the

system can learn to determine the best signal timings

by taking into consideration real-time conditions

instead of using pre-set rules or past patterns to

control traffic flow, provides a paradigm shift for

cities to handle traffic in more and more complex and

unpredictable city life.

The use of a multi-agent approach is conducive

to decentralized control, where the interaction of each

intersection is autonomic, although these traffic

signals are still interacting with their neighbours. This

distributed but cooperating structure contributed

greatly to the improvement of traffic flow throughout

the whole network and helps eliminate problems such

as phasing or local spill over. Therefore, significant

improvement in performance measurement values,

including average delay and queue length with the

frequency of stopping of vehicles, was observed

under various traffic conditions. Naris/Picas The seal

geometry has been specially developed to offer an

energy-efficient solution that also provides reduced

fuel consumption and lower vehicle emissions.

Moreover, the model is flexible in design and it

holds in a dynamic environment which considers the

variation of traffic and the urban expansion. The fact

that this model is occurring while cities are in the

process of development and vehicle numbers are

projected to rise, suggests that the model is well-

equipped to scale its operation and remain effective

without extensive manual reconfiguration. The model

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

106

also paves the way for integration with next

generation technologies, especially connected and

automated vehicles, in which real-time data

exchange can further improve system responsiveness

and predictability.

Although it is highly performant when running in

simulated environments, the development of an actual

deployable version of this system does have to

overcome some challenges such as data validity,

infrastructure status, and computational resources.

However, these challenges are not insurmountable

and can be addressed gradually through staggered

integration, policy assistance and technology

enhancement. To the extent that cities are moving

towards smarter infrastructure, systems similar to the

adaptive, learning-based traffic control proposed in

this paper could be a linchpin for smart & sustainable

urban mobility.

In summary, the framework of reinforcement

learning for the control of traffic signal is a

revolutionary method, consistent with the

requirements of the current urban traffic system. It

helps to maintain the intelligence, scalability and

environment friendliness of the traffic management

system, creating opportunity for the smart city and

intelligent transportation network for further

development.

REFERENCES

Bie, Y., Ji, Y., & Ma, D. (2024). Multi-agent deep

reinforcement learning collaborative traffic signal

control method considering intersection heterogeneity.

Transportation Research Part C: Emerging

Technologies, 164, 104663.

Busch, J. V. S., Voelckner, R., Sossalla, P., Vielhaus, C. L.,

Calandra, R., & Fitzek, F. H. P. (2023). Deep

reinforcement learning for the joint control of traffic

light signaling and vehicle speed advice. arXiv preprint

arXiv:2309.09881.

El-Tantawy, S., Abdulhai, B., & Abdelgawad, H. (2013).

Multiagent reinforcement learning for integrated

network of adaptive traffic signal controllers

(MARLIN-ATSC): Methodology and large-scale

application on downtown Toronto. IEEE Transactions

on Intelligent Transportation Systems, 14(3), 1140–

1150.

Fan, L., Yang, Y., Ji, H., & Xiong, S. (2025). Optimization

of traffic signal cooperative control with sparse deep

reinforcement learning based on knowledge sharing.

Electronics, 14.

Feng, S., Yan, X., Sun, H., Feng, Y., & Liu, H. X. (2021).

Intelligent driving intelligence test for autonomous

vehicles with naturalistic and adversarial environment.

Nature Communications, 12, 748. https://doi.org/10.10

38/s41467-021-21007-8

Feng, S., Sun, H., Yan, X., Zhu, H., Zou, Z., Shen, S., &

Liu, H. X. (2023). Dense reinforcement learning for

safety validation of autonomous vehicles. Nature, 615,

620–627. https://doi.org/10.1038/s41586-023-05732-2

Li, D., Zhu, F., Wu, J., Wong, Y. D., & Chen, T. (2024).

Adaptive urban traffic signal control based on enhanced

deep reinforcement learning. Scientific Reports, 14,

Article 64885. https://doi.org/10.1038/s41598-024-

64885-w

Li, D., Zhu, F., Wu, J., Wong, Y. D., & Chen, T. (2024).

Managing mixed traffic at signalized intersections: An

adaptive signal control and CAV coordination system

based on deep reinforcement learning. Expert Systems

with Applications, 238, 121959.

Liu, H. X., Wu, X., Ma, W., & Hu, H. (2009). Real-time

queue length estimation for congested signalized

intersections. Transportation Research Part C:

Emerging Technologies, 17(4), 412–427.

Peng, X., Gao, H., Wang, H., & Zhang, H. M. (2023).

Combat urban congestion via collaboration:

Heterogeneous GNN-based MARL for coordinated

platooning and traffic signal control. arXiv preprint

arXiv:2310.10948.

Qi, R., Wang, Y., Liu, Y., & Zhang, Y. (2022). Random

ensemble reinforcement learning for traffic signal

control. arXiv preprint arXiv:2203.05961.

Rafique, M. T., Mustafa, A., & Sajid, H. (2024). Adaptive

traffic signal control using reinforcement learning.

ResearchGate.

Rafique, M. T., Mustafa, A., & Sajid, H. (2024).

Reinforcement learning for adaptive traffic signal

control: Turn-based and time-based approaches to

reduce congestion. arXiv preprint arXiv:2408.15751.

Rahman Swapno, S. M. M., Nobel, S. M. N., Meena, P.,

Meena, V. P., Azar, A. T., Haider, Z., & Tounsi, M.

(2024). A reinforcement learning approach for reducing

traffic congestion using deep Q learning. Scientific

Reports, 14, Article 30452. https://doi.org/10.1038/s4

1598-024-75638-0

Wang, M., Xiong, X., Kan, Y., Xu, C., & Pun, M.-O.

(2023). UniTSA: A universal reinforcement learning

framework for V2X traffic signal control. arXiv

preprint arXiv:2312.05090.

Wang, X., Jerome, Z., Wang, Z., Zhang, C., Shen, S.,

Kumar, V., ... & Liu, H. X. (2024). Traffic light

optimization with low penetration rate vehicle

trajectory data. Nature Communications, 15, 1306

https://doi.org/10.1038/s41467-024-45427-4

Xie, X.-F., Smith, S. F., & Barlow, G. J. (2022). Scalable

urban traffic control: Decentralized and schedule-

driven coordination. Transportation Research Part C:

Emerging Technologies, 135, 103465

Yan, X., Zou, Z., Zhu, H., Sun, H., & Liu, H. X. (2023).

Learning naturalistic driving environment with

statistical realism. Nature Communications, 14, 2037.

https://doi.org/10.1038/s41467-023-37677-5

Smart Signal Control Using Reinforcement Learning to Ease Urban Trafﬁc

107

Zheng, J., & Liu, H. X. (2017). Estimating traffic volumes

for signalized intersections using connected vehicle

data. Transportation Research Part C: Emerging

Technologies, 79, 347–362.

Zhu, Y., Cai, M., Schwarz, C., Li, J., & Xiao, S. (2021).

Intelligent traffic light via policy-based deep

reinforcement learning. arXiv preprint arXiv:2112.138

17.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

108