Gnn-Enhanced MARL Achieves Aoi-Aware Queue Management for V2V Networks

Researchers are tackling the challenge of reliable communication in rapidly changing vehicle-to-vehicle (V2V) networks, crucial for cooperative driving and safety applications. Hao Fang, from The National Mobile Communications Research Laboratory at Southeast University, alongside Xiao Li and Chongtao Guo from The Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen University, et al., present a novel approach to managing data freshness and network resources. Their work focuses on minimising the ‘age of information’ , how quickly status updates become stale , by intelligently prioritising and dropping packets, alongside dynamic power control. This research is significant because it introduces a Graph Neural Network (GNN) to understand the complex interference patterns in V2V networks, enabling vehicles to make smarter, more coordinated decisions without constant communication, ultimately improving the speed and reliability of critical status updates.

This breakthrough research addresses the challenge of maintaining up-to-date information about vehicle status, represented by multiple interdependent data packets, amidst the constraints of Wireless communication. Researchers formulated a joint optimization problem, simultaneously learning strategies for actively dropping less important packets and controlling transmit power, enabling fine-grained queue management at the packet level. A key innovation lies in the design of a hybrid action space, accommodating both discrete dropping decisions and continuous power adjustments for optimal performance.

The team achieved substantial progress by exploiting the inherent graph-structured interference present in V2V network topologies. To capture these topological dependencies without constant message exchange, a graph neural network (GNN) was introduced to aggregate slowly varying large-scale fading information. This allows agents within the network to implicitly understand the relationships between vehicles and adjust communication strategies accordingly. The overall framework leverages multi-agent proximal policy optimization (MAPPO) with a centralized training and decentralized execution (CTDE) paradigm, facilitating efficient learning and real-time adaptation.
This approach enables vehicles to make independent decisions based on a globally informed understanding of the network state. Experiments reveal that the proposed method consistently outperforms several baseline approaches in reducing average Age of Information (AoI) across a wide range of network densities, channel conditions, and traffic loads. AoI, defined as the time since the last received packet, is a crucial metric for assessing the freshness of information and ensuring timely decision-making in cooperative driving scenarios. The study establishes that this GNN-enhanced MARL approach provides a significant improvement in maintaining low AoI, which is vital for safety-critical applications like coordinated lane merging and collision avoidance.

This work opens new avenues for enhancing the reliability and responsiveness of autonomous vehicle systems. Furthermore, the research demonstrates the adaptability of the system to dynamic vehicular environments where channel state information and network topology are constantly changing. By integrating GNNs into the reinforcement learning architecture, the framework can extract rich representations of the network state, encapsulating both local interactions and global dependencies. Simulations confirm the robustness of the proposed method, consistently delivering lower AoI even under challenging conditions.

The. Experiments revealed that this method consistently outperforms baseline approaches across diverse scenarios, including varying packet sizes, traffic loads, and channel conditions. The core of the breakthrough delivers a hybrid action space supporting both discrete dropping decisions and continuous power control, enabling fine-grained queue management at the packet level under resource constraints. Results demonstrate a substantial decrease in average AoI across a wide range of densities, channel conditions, and traffic loads. The system model considers both large-scale and small-scale fading, acknowledging that large-scale fading remains constant within intervals of N slots, while small-scale fading is updated in every slot. This allows for long-term topology-aware representation alongside adaptation to short-term channel variations, crucial for scalable and adaptive resource management.

The achievable channel capacity of the m-th V2V link in time slot n is given by Cm [n]=Blog2 1+ |hmm [n]|2pm [n] σ2 + M P i=1,i=m |him [n]|2pi [n], where pm [n] denotes the transmit power. Tests prove that the number of packets successfully transmitted over the m-th V2V link during time slot n is calculated as ym [n] = t · Cm [n] L, where t is the time slot duration and L is the packet size in bits. Data shows that the system addresses a critical scheduling dilemma: whether to continue transmitting remaining packets of a current batch or drop it in favour of a newly arrived batch. A packet dropping factor γm [n] ∈{0, 1} is introduced, where 0 indicates dropping the current batch and 1 signifies continued transmission. Simulations consistently show performance gains, particularly in scenarios where a new status batch arrives before completion of the previous one, highlighting the effectiveness of the proposed packet management strategy.