Design management for smart urban traffic: an IoV-Based multi-agent AI framework for adaptive signal control

1 Introduction

Rapid urbanization has brought with it a multitude of challenges, among which traffic congestion remains one of the most pressing and persistent problems to be addressed. As urban populations grow and vehicle numbers rise, cities worldwide are increasingly turning to smart solutions to manage their transportation systems more effectively (Li et al., 2020). In the broader context of smart cities, intelligent data flow management has become a critical focus area. Understanding the vast amount of data gathered by smart infrastructure and technologies must be processed and interpreted swiftly to facilitate timely and effective decision making (Muntean, 2021). Intelligent Transportation Systems (ITS) utilize information and communication technologies to enhance traffic safety, efficiency, and sustainability. Core elements of ITS include connected and autonomous vehicles (CAVs), which help reduce transportation costs, human error, and traffic-related fatalities Through Vehicle to Everything (V2X) technologies including Vehicle-to-Vehicle (V2V), Vehicle-to-Infrastructure (V2I), Vehicle-to-Pedestrian (V2P), and Vehicle-to-Network (V2N) connected vehicles improve traffic flow and safety across transportation networks (Ashraf et al., 2020). Vehicles are being equipped with ITS to improve safety and passengers’ wellbeing. These include sensors, transmitters, receivers and GPS systems that improve vehicle to environment communication. Public safety applications and Vehicular Ad-Hoc Networks (VANETs) further support these efforts by lowering risks on the road (Alhaj et al., 2023).

One of the most important real-world challenges in ITS is traffic signal control, which aims to minimize vehicle waiting time at intersections. The management of public transportation, traffic control, security, and other matters are greatly influenced by ITS. An essential component of ITS is traffic flow detection, which uses real-time urban road traffic flow data to provide intelligent guidance to alleviate traffic congestion and lower environmental pollution (Hurtado-Gómez et al., 2021; Kumarasamy et al., 2024). The field of Intelligent Traffic Management (ITM) focuses on using advanced technologies and data analytics to ensure the sustainable, safe, and efficient movement of people and goods. Internet of Things (IoT)-connected devices gather real-time traffic condition data, enabling dynamic traffic pattern adjustments, hazard identification, and the promotion of eco-friendly transport modes such as public transit, cycling, and walking (Liu et al., 2020; Atassi and Sharma, 2023). Despite these advances, many nations still rely on conventional traffic lights, which are often hindered by fixing timing, lack of responsiveness to congestion, and vulnerability to weather conditions such as fog or rain. These systems cannot prioritize emergency vehicles or dynamically adapt to fluctuating traffic conditions, highlighting the urgent need for intelligent, flexible, and sustainable traffic management strategies (Elsagheer Mohamed and AlShalfan, 2021).

Recent international work also highlights the value of richer sensing and sustainability-oriented traffic monitoring. Damyanov et al. (Damyanov et al., 2025) demonstrated drone-based infrared thermographic monitoring of urban intersections and road-traffic thermal loads, illustrating how aerial sensing can complement conventional traffic observations. Rahman et al. (Rahman et al., 2025) integrated real-time vehicle detection, counting, speed estimation, and emission monitoring, while Halawani (Halawani, 2025) examined coordinated roundabout control for sustainable urban mobility. These studies motivate a broader sensing-aware traffic-management perspective; however, the numerical experiments in the present study are restricted to the traffic-state variables documented in Section 3.1.

1.1 Objective and key contributions

The primary objective of this research is to develop an intelligent multi-agent traffic-management strategy that can dynamically adapt to observed traffic conditions using IoV traffic-state information and AI-driven decision making. The broader SMATMS concept allows contextual information from remote-sensing systems to be incorporated in future deployments, but the present empirical evaluation does not claim quantitative remote-sensing fusion. Accordingly, sustainability is interpreted primarily through traffic-efficiency indicators such as delay, queue length, travel time, and pedestrian waiting, rather than through directly measured emissions or fuel consumption.

• To formulate an IoV-based multi-agent traffic-state framework for adaptive signal control, while treating remote sensing as a framework-level contextual extension rather than an experimentally validated input.
• To apply transparent normalization and state representation to the documented traffic-state variables, while explicitly separating the unverified historical PCA step from the empirically supported pipeline.
• To develop and clarify a Galactic Swarm Optimized Deep Q-Network in which the two-level swarm structure supports exploration and exploitation of candidate learning/control configurations and the DQN performs sequential control decisions.
• To compare the observed aggregate performance of GSO-DQN with MADDPG and MARDDPG in two- and six-intersection scenarios without overstating statistical significance or metropolitan-scale scalability.

Organization of the paper is as follows: Section 2 presents the related works; Section 3 describes the methodology; Section 4 presents the results; Section 5 discusses the findings; and Section 6 concludes the paper.

2 Related works

Hassan et al. (2023) presented an innovative Cooperative Multi-Agent Deep Q-Network (CMDQN) for traffic light control with the goal of reducing traffic congestion. For cooperation and the exchange of state information, the CMDQN utilized Decentralized Multi-Agent Network (DMN) and a Markov Game model. Utilizing Deep Q-Network (DQN) and reinforcement learning, it employed adaptive traffic signal regulation, utilizing Deep Computer Vision to obtain real-time traffic density data. Through both simulated and real-world tests, the system demonstrated enhanced performance on several important measures, including lowering wait times and enhancing traffic flow (Hassan et al., 2023).

Wu et al. (2020), suggested a unique multi-agent recurrent deep deterministic policy gradient (MARDDPG) algorithm for traffic light regulation in vehicle networks. Through the use of centralized learning across every critic network, the techniques facilitated agents to coordinate and estimate policies. According to experimental results, the technique operated steadily in a variety of situations and greatly decreased traffic for both cars and pedestrians (Wu et al., 2020).

Yang et al. (2020) suggested a Modified Proximal Policy Optimization (Modified PPO) algorithm to enhance traffic control in Smart City IoT (SD-IoT). It manages traffic signals and cars worldwide and modifies the clip hyperparameter to limit the distance between policies. According to experimental data, the algorithm improves SD-IoT performance and lessens traffic congestion while being more stable and competitive than the original algorithm (Yang et al., 2020).

Sharif et al. (2023) suggested a real-time traffic management service response time reduction approach using IoV systems. The path planning optimization issue was formulated as an integer linear program and the Ant Colony Optimization technique was applied. In IoV, fuzzy logic was utilized to estimate the level of congestion. In applications with strict Quality of Service (QoS) criteria, the suggested strategy performed better than current approaches and finds the quickest and most economical route (Sharif et al., 2023).

Hussain et al. (2023) proposed a deep neural network (DNN) architecture that predicted traffic flow performance across four intersections by utilizing gated recurrent units (GRU), bi-directional versions, and Long short term memory (LSTM) layers. One bi-directional LSTM layer and five GRU layers make up the structure’s sixteen layers. The suggested method showed the assure of contemporary technologies in traffic prediction by outperforming state-of-the-art techniques (Hussain et al., 2023).

Musa et al. (2023) focused on using ITS and IoT technology to traffic problems in smart cities. Traffic in hybrid cars was controlled by AI sensors and real-time traffic data. The framework had the ability to estimate traffic, decongest routes, predict traffic outcomes, and streamline urban transportation. By supporting low-carbon emission zones and public transit, pollution challenges was also tackled (Musa et al., 2023).

Hilmani et al. (2020) suggested a wireless sensor network (WSN)-based intelligent traffic control system for gathering information on traffic and parking availability in a smart city. The device allowed drivers to observe traffic rates and parking spots remotely using an Android smartphone application, avoiding traffic jams and offering alternative routes. The system connected citizens to a smart city by integrating three smart subsystems: parking spot management, crossroads management, and a mobile application (Hilmani et al., 2020).

Liu et al. (2023) proposed a deep reinforcement learning-based method for automating network setups in dynamic network contexts, such as the IoV. The method turns communication environment observations into essential features using neural networks to choose the best network configurations for cars. These features were subsequently trained on a DQN. Reliability, latency, and service satisfaction are all improved in IoV environments by the method (Liu et al., 2023).

More recent multi-agent traffic-control research has emphasized stable value estimation, graph-based coordination, and scalable decentralized learning. Alanazi et al. (2026) proposed DR-AGON, which combines a Dueling Double Deep Q-Network, graph attention, and adaptive learning for multi-intersection control. This line of work illustrates that scalability depends not only on the local policy learner but also on communication design and structured coordination among neighboring agents.

The methodological literature also shows the importance of efficient feature representation and decentralized computation. Tao et al. (2025) proposed a cross-dimensional and dual-domain feature-fusion architecture for traffic-sign detection, demonstrating the broader value of carefully designed feature integration in traffic perception. Gao et al. (2025) developed a decentralized optimal iterative-learning-control approach with efficient computation for collaborative tracking, highlighting how distributed coordination can reduce computational burden. Although these studies address different tasks, they inform the present discussion of feature processing, inter-agent coordination, and computational efficiency.

Galactic Swarm Optimization was originally introduced as a two-level population-based global optimizer in which multiple subswarms explore the search space and their best solutions form a superswarm for higher-level exploitation (Muthiah-Nakarajan and Noel, 2016). Relative to a single standard PSO population, this hierarchy is attractive for the present hybrid framework because it explicitly preserves multiple exploratory regions before global refinement. Genetic Algorithms can provide strong global exploration but require selection/crossover/mutation operators; Ant Colony Optimization is particularly natural for discrete path-construction problems and requires pheromone maintenance; standard PSO is computationally simple but uses a single swarm that may lose diversity in multimodal searches. GSO retains PSO-like vector updates while adding hierarchical diversity. Its theoretical update cost remains linear in the number of particles and parameter dimensions at each level, although the superswarm stage introduces additional overhead. The present study therefore motivates GSO structurally rather than claiming that it experimentally dominates every alternative optimizer.

3 Methodology

The Sustainable Multi-Agent Traffic Management Strategy (SMATMS) is an adaptive framework designed to improve urban traffic flow through IoV traffic-state sensing and multi-agent decision making. In the present empirical implementation, signal-control decisions are learned from traffic-state variables. Remote-sensing information is retained only as a possible contextual extension of the architecture and is not treated as a quantitatively fused experimental input.

This section describes data acquisition, preprocessing, the multi-agent reinforcement-learning formulation, and the hybrid GSO-DQN architecture. The revised presentation distinguishes the empirically evaluated traffic-state pipeline from framework-level extensions and provides corrected mathematical expressions and an explicit algorithmic interaction between GSO and DQN. The revised SMATMS empirical pipeline and optional remote-sensing extension are summarized in Figure 1.

3.1 Preprocessing and feature extraction

Effective data preparation is used to place traffic-state variables on comparable numerical scales before learning. Min-Max scaling is applied to continuous features. PCA is described as an optional feature-compression stage in the archived methodology; however, the exact retained component count and cumulative explained-variance ratio are not preserved in the available experiment record. Consequently, the revision does not assign reconstructed PCA percentages or claim a quantified PCA speed-up that cannot be verified.

Min-Max Scaling: For a feature value x with observed minimum and maximum values, scaling to [0,1] is defined by Equation 1.where x denotes the original value, x′ the normalized value, and the subscripts min and max denote the minimum and maximum values of the corresponding feature. The denominator in the previous version contained a typographical algebraic error and has been corrected.

x ′ = x − x min x max − x min (1)

Principal Component Analysis (PCA): PCA transforms correlated traffic-state variables into orthogonal components ordered by explained variance. In principle, a retention threshold can be selected from cumulative explained variance to trade information preservation against input dimensionality. Because the archived experimental record does not retain the actual threshold, retained component count, or explained-variance ratio used in the original run, these quantities are not reconstructed. The revision therefore limits the claim to the methodological role of PCA and identifies quantitative PCA ablation as future work. The empirical evaluation used the Urban Traffic Light Control Dataset (Ziya07, 2026), which provides traffic-state records for the two- and six-intersection scenarios.

3.2 Galactic Swarm Optimized Deep Q-Network (GSO-DQN)

The proposed structure combines Galactic Swarm Optimization (GSO) with a Deep Q-Network (DQN). DQN is the sequential decision learner, whereas GSO provides a higher-level population search over candidate learning/control configurations within the ranges documented in Table 1. This separation clarifies that the swarm optimizer is not itself the traffic-signal controller.

TABLE 1. Documented GSO-DQN hyperparameter ranges and fixed settings.

3.2.1 Deep Q-Network (DQN)

Q-learning is a model-free reinforcement-learning method. For intersection agent i at time t, the local state combines vehicle positions, vehicle velocities, queue length, encoded signal phase, and North–South/East–West pedestrian counts:where action 0 holds the current signal phase and action 1 switches the phase, consistent with Table 1. The Bellman Q-learning update is given in Equation 3.

s i , t = X i , t , V i , t , q i , t , ϕ i , t , p i , t N S , p i , t E W (2)

Q t + 1 s t , a t = Q t s t , a t + α r t + γ max a ′ Q t s t + 1 , a ′ − Q t s t , a t (3)

Here, α is the learning rate, γ is the discount factor, r_t is the immediate reward, and a′ indexes candidate next actions. In the multi-agent setting, each intersection maintains a local value function while interacting with neighboring traffic conditions. The archived manuscript defines the reward as a weighted traffic-performance signal. Because the original numerical reward weights are not retained, the revision does not reconstruct unverified coefficients. The generic reward dependence is represented as:

r i , t = f q i , t , d i , t , τ i , t , p i , t , v i , t (4)

Congestion-related quantities are penalized, whereas efficient traffic movement contributes positively.

Multi-Agent DQN: Each intersection is treated as an agent. The agent observes its local traffic state, receives neighboring state summaries where communication is available, chooses a signal action, stores the transition in a replay buffer, and updates its main Q-network using mini-batch gradient descent. Inter-agent communication is represented generically by:where j denotes a neighboring agent in the neighborhood of agent i. The available manuscript record does not support a more granular reconstruction of the historical message packet.

m j → i , t = g s j , t , j ∈ N i (5)

The agents:

• Observe the current local traffic state and available neighboring summaries.
• Select one of the two admissible actions: hold the current phase or switch the phase.
• Exchange compact neighboring traffic-state summaries to support coordinated congestion reduction.

Each agent operates with two key components.

• Main Network: Estimates Q-values for state-action pairs.
• Target Network: A stable copy of the main network, which is periodically updated to prevent instability during learning.

Action selection follows an ε-greedy policy. At each step, the agent explores with probability ε_t and otherwise exploits the action with the highest estimated Q-value. This is expressed in Equation 6.

a t = random admissible action , with probability ε t , arg max a ∈ A i Q s t , a , with probability 1 − ε t (6)

The exploration probability decays over training according to Equation 7, subject to a minimum exploration level. The documented decay-rate range is reported in Table 1.

ε t + 1 = max ε min , ρ ε t (7)

Each state–action–reward–next-state transition is stored in the replay buffer and sampled in mini-batches to update the main network. The target network provides a more stable bootstrap target and, according to the documented configuration, is synchronized every 1,000 steps. This fixed update frequency is now distinguished from the tunable parameter ranges in Table 1.

3.2.2 Galactic Swarm Optimization (GSO)

GSO was selected because its hierarchical multi-swarm structure explicitly separates broad exploration from higher-level exploitation (Muthiah-Nakarajan and Noel, 2016). In contrast with a single standard PSO population, multiple subswarms can preserve several candidate regions before their best solutions are aggregated. This is useful for a hybrid RL setting in which the performance surface over learning/control parameters may be multimodal. The choice is motivated by this structural fit and by PSO-like vector updates, not by an unreported head-to-head experiment against GA, ACO, or standard PSO.

Mechanism of Galactic Swarm Optimization (GSO) in SMATMS:

Level 1 (Subswarm Exploration): Let x and v denote the position and velocity of particle i in subswarm k at iteration t. Each particle represents a candidate configuration in the tunable parameter space. The local PSO-style updates are given in Equations 8, 9, where p denotes the particle-best position and g denotes the best position in subswarm k. The particle movement within a subswarm is illustrated in Figure 2.

v i , k t + 1 = ω v i , k t + c 1 r 1 p i , k t − x i , k t + c 2 r 2 g k t − x i , k t (8)

x i , k t + 1 = x i , k t + v i , k t + 1 (9)

Here, ω is the inertia weight; c₁ and c₂ are acceleration constants; and r₁ and r₂ are independent random variables in [0,1]. The notation has been standardized to resolve the inconsistent indexing in the previous version.

Level 2 (Superswarm Exploitation): The best solution from each subswarm is promoted to a superswarm. Let y and u denote the superswarm position and velocity associated with subswarm k. The higher-level refinement follows Equations 10, 11.

u k t + 1 = Ω u k t + C 1 R 1 P k t − y k t + C 2 R 2 G t − y k t (10)

y k t + 1 = y k t + u k t + 1 (11)

Here, P_k is the personal best of superswarm member k, G is the global best superswarm solution, Ω is the superswarm inertia weight, and C₁ and C₂ are acceleration constants. The two-level structure therefore adds a superswarm refinement stage to PSO-like updates. The corresponding superswarm particle dynamics are illustrated in Figure 3.

Computationally, a standard PSO update scales approximately linearly with particle count N and search dimension d per iteration, O(Nd), excluding objective-function evaluation. A two-level GSO adds a superswarm update over K promoted subswarm solutions, yielding an update cost of approximately O(Nd + Kd) per hierarchical cycle, again excluding the cost of evaluating candidate DQN configurations. GA has additional selection/crossover/mutation operations, whereas ACO requires pheromone updates and is particularly natural for discrete construction problems. GSO therefore incurs more coordination overhead than single-swarm PSO but preserves explicit hierarchical diversity.

The interaction between GSO and DQN is treated as a two-timescale process. During configuration search, each GSO iteration updates candidate parameter vectors after their DQN-based traffic-control fitness has been evaluated. The best available candidate is then passed to the DQN inner learner. During signal-control evaluation/deployment, GSO is not invoked at every environment step; the DQN agents execute the selected configuration and update/act according to the reinforcement-learning loop. The documented search domains include learning rate, discount factor, exploration-related settings, inertia/acceleration parameters, subswarm size, batch size, and replay-buffer size. The exact historical outer-loop refresh count and final particle vector are not reconstructed because the tuning log is unavailable.

Algorithm 1 Clarified GSO-DQN interaction in SMATMS.

4 Results

The following section reports the archived aggregate evaluation of GSO-DQN and the comparison algorithms. The implementation environment used 16 GB RAM, an Intel Core i7 processor, and an NVIDIA GTX 1660 Ti GPU, with Python 3.8 and deep-learning libraries. To improve transparency, Table 1 is relabeled to distinguish documented search/configuration ranges from the fixed target-network update frequency. The exact final historical values selected inside several ranges are not reconstructed from unavailable tuning logs.

Parameter-tuning strategy: In the revised methodological interpretation, GSO, is the automated outer-loop search mechanism over the documented ranges in Table 1, while the target-network synchronization interval is held fixed at 1,000 steps. Candidate configurations are evaluated using aggregate traffic-control performance and the best available candidate is passed to the DQN, learner. Because the archived tuning log does not preserve the final selected particle/vector, the revision reports verifiable search ranges and fixed settings rather than retrospectively assigning unsupported final values.

4.1 Evaluation metrics

The reported evaluation uses average reward together with average vehicle queue length, average travel time, average delay, and average number of waiting pedestrians. Higher reward is better, whereas lower values of delay, queue length, travel time, and waiting pedestrians indicate better traffic efficiency. The values in Tables 2, 3 are presented as mean ± standard deviation in the archived results. Several standard-deviation entries in the previous tables omitted a leading decimal point; these typographical errors have been corrected to match the plotted error-bar scale. Formal p-values or confidence intervals are not reconstructed because the aggregate tables do not preserve the independent-run sample size or run-level observations required for valid inferential testing.

TABLE 2. Performance comparison of algorithms in the two-intersection scenario (mean ± SD as reported in the archived aggregate results).

TABLE 3. Performance comparison of algorithms in the six-intersection scenario (mean ± SD as reported in the archived aggregate results).

4.2 Output Analysis

Figure 4 provides representative time-series outputs over 500 time steps for average queue length, vehicle velocity, and pedestrian count, together with the pedestrian-count/queue-length scatter plot. These plots show bounded fluctuations in the displayed trajectory and no obvious monotonic increase in queue length with pedestrian count. The revised text treats this observation descriptively; no statistical association is inferred from the scatter plot without an explicit correlation model.

4.3 Comparative Analysis

The proposed GSO-DQN model is compared with Multi-Agent Deep Deterministic Policy Gradient (MADDPG) (Lowe et al., 2017) and Multi-Agent Recurrent Deep Deterministic Policy Gradient (MARDDPG) (Wu et al., 2020). The comparison is descriptive and is restricted to the two archived network sizes.

Scenario 1 (2 Intersections): In the smaller scenario, GSO-DQN shows the highest reported mean reward (2.41). Relative to the strongest baseline for each lower-is-better metric, the reported mean values correspond to reductions of 20.65% in delay, 28.29% in queue length, 5.25% in travel time, and 16.81% in waiting pedestrians. The reward exceeds the stronger baseline reward by 0.53. These are descriptive differences calculated from the aggregate means and should not be interpreted as inferential significance. The comparative results for the two-intersection scenario are shown in Figure 5.

Scenario 2 (6 Intersections): In the six-intersection scenario, GSO-DQN again has the highest reported mean reward (−1.12, compared with −1.87 for the stronger baseline). Relative to the strongest lower-is-better baseline, the reported mean values correspond to reductions of 13.15% in delay, 29.17% in queue length, 7.41% in travel time, and 26.37% in waiting pedestrians. The reward improves by 0.75 in absolute score. The six-intersection case is therefore interpreted as a more complex multi-intersection validation than the two-intersection case, but not as evidence of metropolitan-scale scalability. The comparative results for the six-intersection scenario are shown in Figure 6.

5 Discussion

The revised interpretation emphasizes what the archived evidence can and cannot support. First, GSO is motivated by its two-level search structure: multiple subswarms preserve exploratory diversity and the superswarm refines the best candidate regions (Muthiah-Nakarajan and Noel, 2016). Compared with a single PSO, this adds an O(Kd) superswarm update to the O(Nd) particle-update structure, excluding objective evaluation. In a GSO-DQN implementation, objective evaluation is the dominant cost because candidate parameter configurations must be assessed through learning performance. GA and ACO remain reasonable alternatives, and the current study does not claim experimental superiority of GSO over those optimizers. Second, the aggregate results favor GSO-DQN in both archived scenarios. Table 4 quantifies the descriptive changes relative to the strongest baseline: queue length is lower by 28.29% and 29.17% in the two- and six-intersection scenarios, respectively, while delay, travel time, and pedestrian waiting also show lower reported means. These values strengthen the descriptive comparison but do not replace repeated-seed hypothesis testing.

TABLE 4. Descriptive improvement of GSO-DQN relative to the strongest baseline using the reported aggregate means (2-int. = two intersections; 6-int. = six intersections). These calculations do not constitute statistical significance tests.

Sustainability is interpreted in a similarly bounded way. Reduced delay, queueing, and travel time can support more efficient mobility and may indirectly reduce stop-and-go operation, but the present experiment does not directly measure fuel consumption, energy use, CO2, NOx, particulate emissions, or carbon footprint. Recent emission-monitoring work (Rahman et al., 2025) illustrates how vehicle type, speed, and emission factors could support a future environmental evaluation. Accordingly, all claims of demonstrated emission reduction have been removed, and environmental metrics are identified as a future extension.

Remote sensing is also separated from the empirical evidence. Studies such as Damyanov et al. (Damyanov et al., 2025) show how aerial thermal sensing can characterize intersection conditions, but no such feature was quantitatively fused in the archived experiment. The revised manuscript therefore treats remote sensing as a possible contextual layer rather than as an operationally validated component.

Qualitative GSO sensitivity: The two-level swarm parameters have interpretable computational effects even though a quantitative historical sweep cannot be reconstructed. A larger subswarm size broadens search diversity but increases the number of expensive DQN candidate evaluations. Larger inertia weights (ω, Ω) preserve momentum and favor exploration, whereas smaller values emphasize local refinement. Larger acceleration coefficients (c₁, c₂, C₁, C₂) strengthen attraction toward particle/subswarm/global best solutions and may accelerate convergence while reducing diversity. These statements are algorithmic sensitivity considerations rather than measured effect sizes; a controlled two-level parameter sweep is identified as a required follow-up experiment.

Limitations: The revision cannot reconstruct information that is not preserved in the aggregate manuscript record. In particular, the archived material does not provide the independent multi-seed observations required for valid confidence intervals or hypothesis tests, the numerical PCA retention log, the final selected values for all ranged hyperparameters, a version-specific dataset identifier and total row count, or a fully auditable training/testing split. Rather than impute these quantities, the manuscript now discloses these reproducibility limitations and removes inferential or scalability claims that would require them. Future work should repeat the experiment with at least three independent runs per configuration, report random seeds and confidence intervals, archive the exact dataset version and code, and include GSO ablation, optimizer comparisons, larger networks, and environmental metrics.

Scalability and communication overhead require cautious interpretation. With N intersection agents and an inter-agent communication graph containing |E| neighbor links, compact neighbor-state exchange grows approximately with O (|E|·d_m) per environment step, where d_m is the message dimension. DQN inference also scales with the number of active agents, while GSO adds outer-loop candidate evaluations during training. The six-intersection experiment therefore demonstrates a larger coordinated setting than the two-intersection case, but it does not establish metropolitan-scale real-time feasibility. Larger grid networks, heterogeneous traffic demand, explicit runtime profiling, memory consumption, communication latency, and comparison with recent graph-based MARL approaches such as DR-AGON (Alanazi et al., 2026) are necessary for a strong scalability claim.

6 Conclusion

In conclusion, this study presents an IoV-based SMATMS using a hybrid GSO-DQN formulation for adaptive multi-agent traffic-signal control. The archived aggregate results show favorable observed mean values for reward, delay, queue length, travel time, and pedestrian waiting relative to MADDPG and MARDDPG in two- and six-intersection scenarios. The revision clarifies the mathematical formulation, the two-level role of GSO, target-network updating, the limits of remote-sensing claims, and the descriptive nature of the performance comparison. The results should not be interpreted as formal evidence of statistical significance, direct emission reduction, or metropolitan-scale scalability. Future work should conduct repeated-seed experiments, ablation and sensitivity studies, larger-network validation, explicit communication/runtime profiling, and direct energy/emission assessment.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Rajesh Kumar, National Institute of Technology, India

Reviewed by: Mohammad Reza Kavianpour, K.N.Toosi University of Technology, Iran

Vladimir Stojanovic, University of Kragujevac, Serbia