Federated Multi-Agent Reinforcement Learning for Dynamic Bandwidth and Computation Allocation in Vehicular Ad-Hoc Networks under Asymmetric Sensor Fusion Constraints

Federated Multi-Agent Reinforcement Learning for Dynamic Bandwidth and Computation Allocation in Vehicular Ad-Hoc Networks under Asymmetric Sensor Fusion Constraints

Reza Homayouni Amlashi1

1) Department of Computer Engineering, Islamic Azad University, Babol Branch, Babol, Iran

IOI Code : ACEF-ACAAGZ
Publication : 6th. International Congress On Engineering, Technology and Applied Sciences(etas2027.com)
Abstract :
    Cooperative perception in Ad-Hoc Networks can extend the sensing range of connected autonomous vehicles beyond direct line‑of‑sight and physical blockages. Sharing multi‑modal Bird s‑Eye‑View features from LiDAR, radar and camera streams however brings communication and computing costs on changing vehicle‑to‑everything links. Weather can degrade sensors unevenly making blanket feature sharing wasteful or harmful. This paper introduces Fed‑MAPPO, a multi‑agent reinforcement‑learning system for joint bandwidth allocation, edge computing offloading and modality‑aware feature pruning. A graph‑attention encoder captures the evolving vehicle neighbourhoods while a latency‑deficit queue penalises persistent delay violations. The study presents a CARLA–SUMO–OMNeT++/Simu5G co‑simulation setup. Gives early performance results compared to baseline methods. Because run‑level logs and raw data are not available, for audit the numerical results and statistical tests must be regarded as unverified and need reproduction before being taken as solid evidence. The queue design alone does not guarantee a fixed per‑packet latency; a guarantee would need assumptions and a separate proof.

Keywords : vehicular ad hoc networks (vanets) cooperative perception federated learning multi agent reinforcement learning (marl) lyapunov optimization resource allocation c v2x