Reinforcement Learning-Based Resource Allocation for IRS-Enhanced V2X Communication and Computation Systems

Abstract

Vehicle-to-everything (V2X) communication and computation face significant challenges in delivering ultrareliable, low-latency communication and optimizing energy efficiency in dynamic vehicular environments. To address these issues, intelligent reflecting surfaces (IRSs) have been introduced to enhance communication performance, improve reliability, and reduce latency and energy consumption. This article proposes an IRS-enhanced V2X system that employs multiple IRSs to improve vehicle-to-vehicle (V2V) communication and to offload computation by leveraging spectrum reuse. An effective system utility function that quantifies data versus energy consumption is developed, facilitating precise system evaluation and optimization. Given the complexity of the original optimization problem and the difficulty of direct solution approaches, we propose an improved reinforcement learning (IRL)-based resource allocation algorithm to maximize the system’s utility function under power, beam, and resource limitations, without requiring rigid system state acquisition or complicated computational complexity. A deep deterministic policy gradient (DDPG)-based neural network employing a reparameterization technique is constructed to jointly optimize continuous and discrete variables. Simulations show that maximizing the effective utility function significantly enhances system efficiency compared to benchmarks. Strategic IRS deployment and network adjustments reduce channel losses, yielding substantial performance gains (37.7% compared to no IRS and 20.5% compared to other learning methods) and supporting the advancement of intelligent, sustainable transportation networks.