Deep Reinforcement Learning for BD-RIS-Assisted Multi-UAV RSMA Communications
Abstract
Beyond-diagonal reconfigurable intelligent surfaces (BD-RISs) extend conventional diagonal RISs by allowing controllable coupling among reflecting elements. This flexibility is attractive for aerial networks, where unmanned aerial vehicles (UAVs) face rapidly varying channels and strict power limits. In this paper, we study a hybrid BD-RIS-assisted multi-UAV rate-splitting multiple access (RSMA) downlink, in which UAVs within favourable BD-RIS coverage use both direct and BD-RIS-aided links, whereas the remaining UAVs are served through direct links only. We formulate a network sum-rate maximization problem that jointly optimizes UAV trajectories, RSMA common-rate allocation, beamforming, BD-RIS scattering matrices, and UAV-BD-RIS association. The resulting mixed-integer non-convex problem is solved by using deep deterministic policy gradient (DDPG) algorithm, equipped with a constraint-preserving action layer that generates reciprocal and lossless BD-RIS responses in closed form and handles the binary association inside the environment transition. Simulation results, averaged over three training seeds under a physically consistent channel model with blocked direct links near the BD-RISs, show that the proposed design attains a sum rate comparable to SAC and higher than twin delayed deep deterministic policy gradient ( TD3), proximal policy optimization (PPO), and deep Q-network (DQN), exceeds per-slot alternating optimization with fixed trajectories at about 1/65 of its per-slot runtime, and benefits mainly from RSMA, while the additional BD-RIS gain is modest.
Related
- Deep Reinforcement Learning-Aided Optimization of Multi-Interface Allocation for Short-Packet Communications
- Deep Reinforcement Learning for Blind Passive Beamforming in RIS-Assisted MIMO Systems
- Dynamic Multi-Pair Trading Strategy in Cryptocurrency Markets with Deep Reinforcement Learning
- Short-Length Code Designs for Integrated Sensing and Communications: A Deep Learning Approach
- Data-Assimilation-Assisted Reinforcement Learning for Power Grid Control under Load Uncertainty
BibTeX
@misc{huroon2026deep,
title = {{Deep Reinforcement Learning for BD-RIS-Assisted Multi-UAV RSMA Communications}},
author = {Aamer Mohamed Huroon and Ali A. Nasir and Daniel B. da Costa},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.03047},
eprint = {2611.03047},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.03047},
}
RIS
TY - UNPB
TI - Deep Reinforcement Learning for BD-RIS-Assisted Multi-UAV RSMA Communications
AU - Aamer Mohamed Huroon
AU - Ali A. Nasir
AU - Daniel B. da Costa
PY - 2026
DA - 2026/11/02
PB - arXiv
DO - 10.48550/arxiv.2611.03047
UR - https://arxiv.org/abs/2611.03047
AB - Beyond-diagonal reconfigurable intelligent surfaces (BD-RISs) extend conventional diagonal RISs by allowing controllable coupling among reflecting elements. This flexibility is attractive for aerial networks, where unmanned aerial vehicles (UAVs) face rapidly varying channels and strict power limits. In this paper, we study a hybrid BD-RIS-assisted multi-UAV rate-splitting multiple access (RSMA) downlink, in which UAVs within favourable BD-RIS coverage use both direct and BD-RIS-aided links, whereas the remaining UAVs are served through direct links only. We formulate a network sum-rate maximization problem that jointly optimizes UAV trajectories, RSMA common-rate allocation, beamforming, BD-RIS scattering matrices, and UAV-BD-RIS association. The resulting mixed-integer non-convex problem is solved by using deep deterministic policy gradient (DDPG) algorithm, equipped with a constraint-preserving action layer that generates reciprocal and lossless BD-RIS responses in closed form and handles the binary association inside the environment transition. Simulation results, averaged over three training seeds under a physically consistent channel model with blocked direct links near the BD-RISs, show that the proposed design attains a sum rate comparable to SAC and higher than twin delayed deep deterministic policy gradient ( TD3), proximal policy optimization (PPO), and deep Q-network (DQN), exceeds per-slot alternating optimization with fixed trajectories at about 1/65 of its per-slot runtime, and benefits mainly from RSMA, while the additional BD-RIS gain is modest.
ER -