Joint Deployment and Beamforming Design in Aerial RIS-Assisted MU-MISO Systems Using Deep Reinforcement Learning

Reconfigurable intelligent surfaces (RIS) have emerged as a promising and cost-effective technology to enhance coverage, transmission rates, and energy efficiency in wireless communication systems. However, current research often treats the deployment and beamforming design of RIS as separate issues, resulting in suboptimal performance. This study aims to jointly optimize the deployment and beamforming of RIS to maximize the weighted sum rate of a multi-user multiple-input single-output (MISO) system. We model the control of RIS as a Markov decision process and develop a deep reinforcement learning (DRL) framework to address this optimization task jointly. Specifically, we design the action space of the agent using an enhanced discrete Fourier transform codebook, which reduces the dimensionality of the action space and facilitates the efficient training of the DRL. To exploit the graph topology inherent in the wireless communication optimization problem, we employ a graph neural network (GNN) structure to construct the DRL framework. Additionally, assuming the availability of location information, we utilize the location information of users and RIS to further enhance performance by designing an initialization network and modifying the state space. Simulation results validate the effectiveness of the proposed joint deployment and beamforming design, demonstrating up to twofold performance improvements compared to fixed RIS deployment strategies. Moreover, the proposed GNN-based DRL framework exhibits superior robustness, enabling flexible deployment across varying numbers of users. The incorporation of location information accelerates the training process and improves overall performance and robustness, underscoring the advantages of AI-enabled multimodal beamforming design.