Deep Reinforcement Learning based Joint Spectrum Allocation and Configuration Design for STAR-RIS-Assisted V2X Communications

TL;DR

This paper proposes a deep reinforcement learning approach combined with optimization techniques to enhance spectrum allocation and configuration in STAR-RIS-assisted V2X communications, improving coverage and data rates.

Contribution

It introduces a novel joint optimization framework for STAR-RIS-assisted V2X systems, integrating deep reinforcement learning with real-time optimization for better performance.

Findings

Proposed DDQN with attention mechanism effectively models control variables.

Joint optimization improves data rate and coverage in V2X systems.

Algorithm outperforms benchmark schemes in numerical simulations.

Abstract

Vehicle-to-Everything (V2X) communications play a crucial role in ensuring safe and efficient modern transportation systems. However, challenges arise in scenarios with buildings, leading to signal obstruction and coverage limitations. To alleviate these challenges, reconfigurable intelligent surface (RIS) is regarded as an effective solution for communication performance by tuning passive signal reflection. RIS has acquired prominence in 6G networks due to its improved spectral efficiency, simple deployment, and cost-effectiveness. Nevertheless, conventional RIS solutions have coverage limitations. Therefore, researchers have started focusing on the promising concept of simultaneously transmitting and reflecting RIS (STAR-RIS), which provides 360\degree coverage while utilizing the advantages of RIS technology. In this paper, a STAR-RIS-assisted V2X communication system is investigated. An optimization problem is formulated to maximize the achievable data rate for vehicle-to-infrastructure (V2I) users while satisfying the latency and reliability requirements of vehicle-to-vehicle (V2V) pairs by jointly optimizing the spectrum allocation, amplitudes, and phase shifts of STAR-RIS elements, digital beamforming vectors for V2I links, and transmit power for V2V pairs. Since it is challenging to solve in polynomial time, we decompose our problem into two sub-problems. For the first sub-problem, we model the control variables as a Markov Decision Process (MDP) and propose a combined double deep Q-network (DDQN) with an attention mechanism so that the model can potentially focus on relevant inputs. For the latter, a standard optimization-based approach is implemented to provide a real-time solution, reducing computational costs. Extensive numerical analysis is developed to demonstrate the superiority of our proposed algorithm compared to benchmark schemes.