A Bidirectional-AoI-Aware Multi-Agent Deep Reinforcement Learning Framework for Vehicular Platooning in Segmented Waveguide-Based Pinching Antenna Systems

Abstract

Ensuring reliable and low-latency vehicle-to-everything (V2X) communications in high-speed transport settings remains a significant challenge due to severe path loss brought about by non-line-of-sight (NLoS) and coverage gaps in conventional cellular infrastructure. While dielectric waveguide-based pinching antenna (PA) systems have been proposed to mitigate these physical limitations, they suffer from substantial in-waveguide attenuation over long distances. To address these challenges, we propose a segmented waveguide-enabled pinching-antenna (SWAN) architecture in platoon-based V2X networks. By employing dynamic segment selection, SWAN maintains robust line-of-sight (LoS) connectivity while mitigating the in-waveguide attenuation inherent in conventional PA structures. We formulate a joint resource allocation (RA) and mode selection problem to minimise the age of information (AoI) for both uplink platoon monitoring and downlink traffic broadcasting, whilst ensuring the exchange of intra-platoon cooperative awareness messages (CAMs) and minimising power consumption. To solve this high-dimensional problem, we propose a decomposed multi-agent deep deterministic policy gradient (DE-MADDPG) algorithm augmented with twin delayed (TD3) critics by considering each vehicle platoon (VP) as an agent. This approach decouples system-wide coordination from local executions of VPs, enabling efficient learning in dynamic environments. Extensive simulations demonstrate that the proposed framework significantly outperforms standard reinforcement learning (RL) baseline methods, achieving near-optimal uplink and downlink AoI performance, with an average gap of 4.8% to exhaustive search, and near-perfect CAM delivery probability (CDP), which approaches 100%, even under dense traffic conditions.