Peer-Aware and Fair-Play Decentralized Task Offloading for Vehicular Edge Computing: A Multi-Agent Learning Framework

Venue: China Communications (under major revision)

Abstract

The rapid evolution of the Internet of Vehicles (IoVs) has brought vehicle-road collaboration systems to the forefront of intelligent transportation. Modern applications such as speech recognition, large language models, and intelligent assistants impose computational demands that far exceed the capabilities of typical onboard devices. Vehicular Edge Computing (VEC) mitigates this limitation by enabling vehicle task offloading to powerful edge servers. However, vehicular edge environments still face challenges such as high mobility, fluctuating wireless resources, and the need to balance system efficiency with long-term fairness across vehicles. To address these issues, this work introduces a multi-agent variable-length action reinforcement learning framework for task offloading. Each base station acts as an autonomous agent that executes independently without a central controller. The framework employs non-causal self-attention to aggregate observations from a changing set of vehicles, thereby supporting implicit inter-agent communication without excessive signaling. Furthermore, this framework can adapt decisions to dynamic vehicle populations and channel states. To achieve fair-play, a dynamic priority compensation module adjusts resource allocation based on each vehicle's historical Quality of Service (QoS), preventing persistent service deprivation while preserving overall efficiency. Extensive simulations based on a real-world trajectory dataset demonstrate its superiority compared to baseline methods. Specifically, the proposed framework improves task success rate by over 11%, reduces average latency by 6%, and lowers tail latency (99th-percentile delay) by 4% compared with representative baselines, while maintaining fair resource allocation.