Resource Allocation for Out‑of‑Coverage V2V Using an Actor–Critic CTDE

Vehicular communications are critical for road safety, but out-of-coverage Vehicle-to-Vehicle (V2V) links face stringent latency and reliability demands in highly mobile, spectrum-congested environments. 5G New Radio-Vehicle-to-Everything (NR-V2X) Mode 2 lets vehicles semi-persistently reserve control (Physical Sidelink Control Channel) and data (Physical Sidelink Data Channel) resources, yet collisions and imperfect sensing of primary users (PUs) still lead to significant performance degradation. We propose a multi-agent resource allocation framework based on Centralized Training with Decentralized Execution (CTDE) and an Actor–Critic architecture. Each V2V link runs a lightweight Long Short-Term Memory (LSTM)-based Actor that outputs a continuous data subcarrier “budget” via a log-normal policy, while a reactive semi-persistent Physical Sidelink Control Channel/Physical Sidelink Data Channel scheduler handles control-block collisions and PU detection. During training, a global LSTM Critic observes all channel qualities and sensing masks, stabilizing value estimates despite aging Channel State Information (CSI) and sensing errors. We define a reward that balances Packet Reception Ratio (PRR) and a satisfaction index across three service levels (SSV+, C1, C2), fostering cooperation under a 100 ms latency constraint. Our simulations—integrating Simulation of Urban MObility (SUMO) mobility, 3rd Generation Partnership Project (3GPP) Urban Microcell (UMi) Line-of-Sight/Non-Line-of-Sight (LOS/NLOS) path loss with log-normal shadowing, multi-tap Tapped Delay Line (TDL) Doppler fading, Demodulation Reference Signals (DMRS)-based channel estimation, and Effective Exponential SINR Mapping (EESM) aggregation—demonstrate rapid convergence for critical services under fully decentralized execution.