Hybrid-RIS-Assisted Cellular ISAC Networks for UAV-Enabled Low-Altitude Economy via Deep Reinforcement Learning with Mixture-of-Experts

• Published in: IEEE transactions on cognitive communications and networking p. 1
• Main Authors: Ma, Zhangfeng, Liang, Yongzhe, Zhu, Qiuming, Zheng, Jiakang, Lian, Zhuxian, Zeng, Linzhou, Fu, Chengwei, Peng, Yifei, Ai, Bo
• Format: Journal Article
• Language: English
• Published: IEEE 15.10.2025
• Subjects: Antennas; Array signal processing; cellular integrated sensing and communication; deep reinforcement learning; hybrid reconfigurable intelligent surface; Integrated sensing and communication; Interference cancellation; Low-altitude economy; mixture of experts; Optimization; Quality of service; Radar antennas; Reconfigurable intelligent surfaces; Signal to noise ratio; Vectors
• ISSN: 2332-7731; 2332-7731
• DOI: 10.1109/TCCN.2025.3622130

This paper investigates cellular base station (BS)-enabled wireless systems featuring downlink integrated sensing and communications, a critical enabler for the low-altitude economy. To serve a sensing target and multiple communication users simultaneously, a hybrid active-passive reconfigurable intelligent surface (RIS) is employed. A key challenge arises from the inherent downtilt of cellular BS antennas, which often results in high-altitude targets (e.g., unmanned aerial vehicles) being sensed predominantly via antenna sidelobes, thereby substantially constraining sensing performance. To jointly enhance both communication and sensing capabilities, we formulate a radar signal-to-noise ratio maximization problem. This formulation involves the co-optimization of beamforming vectors and RIS phase shifts, subject to constraints on transmit power, hybrid RIS power consumption, discrete RIS phase shifts, and communication user quality of service requirements. Addressing the non-convexity and NP-hardness of this problem, we propose a novel mixture-of-experts (MoE)-based proximal policy optimization (PPO) approach. Specifically, the MoE architecture is integrated within PPO to alleviate the learning complexity and mitigate gradient interference inherent in a single policy network. This is achieved by employing expert networks that specialize in distinct regions of the state space, orchestrated by a gating network that dynamically weights their contributions, thereby promoting accelerated convergence and enhanced optimization efficiency. Numerical results demonstrate that even for a target flying above the cellular BS, the deployment of a hybrid RIS with as few as 4 active elements, each providing an amplitude amplification of 2, yields a remarkable 18.5% improvement in sensing SNR, which translates to substantially enhanced sensing accuracy and reliability for low-altitude targets.