GAPO: A Graph Attention-Based Reinforcement Learning Algorithm for Congestion-Aware Task Offloading in Multi-Hop...

Area of Science:

• Computer Science
• Artificial Intelligence
• Network Engineering

Background:

• Vehicular Edge Computing (VEC) networks face challenges in task offloading due to high vehicle mobility and dynamic network topologies.
• Delay-sensitive applications like autonomous driving require efficient solutions to minimize latency and avoid congestion.

Purpose of the Study:

• To propose an efficient and adaptive task offloading algorithm for multi-hop VEC networks.
• To address the challenges of dynamic network topologies and end-to-end congestion.

Main Methods:

• Developed GAPO, a graph attention-based reinforcement learning algorithm.
• Modeled the VEC network as an attributed graph using a Graph Neural Network (GNN) for state representation.
• Employed an attention-based Actor-Critic framework for joint offloading and resource allocation decisions.

Main Results:

• GAPO significantly reduces average task completion latency compared to traditional methods.
• The algorithm substantially decreases backbone link congestion in VEC networks.
• Demonstrated superior performance through comprehensive simulation experiments and ablation studies.

Conclusions:

• GAPO offers an efficient, adaptive, and congestion-aware solution for resource management in dynamic VEC environments.
• Integrating GNNs with Deep Reinforcement Learning (DRL) enhances VEC network performance.
• The proposed approach effectively tackles complex offloading challenges in VEC networks.