Multiobjective Optimization of Wireless Powered Communication Networks Assisted by Intelligent Reflecting Surface Based on Multiagent Reinforcement Learning

• Published in: IEEE transactions on antennas and propagation 2024-04, Vol.72 (4), p.3274-3281
• Main Authors: Guan, Xiangrui, Xue, Jianbin, Jiang, Hengjie, Tian, Guiying
• Format: Article
• Language: English
• Subjects: Array signal processing; Beamforming; Communication; Communication networks; Communications networks; Communications systems; Energy harvesting; Intelligent reflecting surface (IRS); Internet of Things; multiagent reinforcement learning; Multiagent systems; multiobjective optimization; Multiple objective analysis; Optimization; Reconfigurable intelligent surfaces; Reflection; Resource management; Throughput; Wireless communication; Wireless communications; Wireless networks; wireless powered communication network (WPCN)
• ISSN: 0018-926X; 1558-2221
• DOI: 10.1109/TAP.2024.3370195

Intelligent reflecting surface (IRS) is expected to be an important enabling technology for future wireless communication networks due to its capacity for reconfiguring wireless propagation environments. In this article, we consider a multiuser communication system for wireless powered communication network (WPCN) with IRS assistance. To overcome the low-quality communication problem of remote Internet of Things (IoT) devices in WPCN, we propose a multiobjective optimization scheme for IRS-assisted WPCN to optimize jointly throughput and remaining energy of remote IoT devices. We present a multiobjective optimization problem by jointly designing the hybrid access point (HAP) transmit beamforming, HAP receive beamforming, IRS phase shift beamforming, the IoT device transmit power, and energy-harvesting (EH)/information transmission (IT) time allocation to maximize system throughput and remaining energy. To address the aforementioned multiobjective optimization problem, the original optimization problem is first transformed into a Markov game model, and then, a multiobjective optimization scheme based on a multiagent deep deterministic policy gradient (MADDPG) is proposed. We centrally train the MADDPG model offline, and the two optimization objectives throughput and remaining energy are abstracted as two agents to execute decisions online. According to the results of the simulation, the multiobjective optimization scheme based on multiagent reinforcement learning can guarantee the performance of WPCN and enhance the throughput and remaining energy overall.