| dc.description.abstract | Reconfigurable intelligent surfaces (RISs) or known as intelligent reflecting surfaces (IRSs) have emerged as potential auxiliary equipment for future wireless networks, which attracts extensive research interest in their characteristics, applications, and potential. RIS is a panel surface equipped with a number of reflective elements, which can artificially modify the propagation environment of the electrogenic signals. Specifically, RISs have the ability to precisely adjust the propagation direction, amplitude, and phase-shift of the signals, providing users with a set of cascaded channels in addition to direct channels, and thereby improving the communication performances for users. Compared with other candidate technologies such as active relays, RIS has advantages in terms of flexible deployment, economical cost, and high energy efficiency. Thus, RISs have been considered a potential candidate technique for future wireless networks. In this thesis, a wireless network paradigm for the sixth generation (6G) wireless networks is proposed, where RISs are invoked to construct smart radio environments (SRE) to enhance communication performances for mobile users. In addition, beyond the conventional reselecting-only RIS, a novel model of RIS is originally proposed, namely, simultaneous transmitting and reflecting reconfigurable intelligent surface (STAR-RIS). The STAR-RIS splits the incident signal into transmitted and reflected signals, making full utilization of them to generate $360^{\circ}$ coverage around the STAR-RIS panel, improving the coverage of the RIS. In order to fully exert the channel domination and beamforming ability of the RISs and STAR-RSIs to construct SREs, several machine learning algorithms, including deep learning (DL), deep reinforcement learning (DRL), and federated learning (FL) approaches are developed to optimize the communication performance in respect of sum data rate or energy efficiency for the RIS-assisted networks. Specifically, several problems are investigated including 1) the passive beamforming problem of the RIS with consideration of configuration overhead is resolved by a DL and a DRL algorithm, where the time overhead of configuration of RIS is successfully reduced by the machine learning algorithms. Consequently, the throughput during a time frame improved $95.2\%$ by invoking the proposed algorithms; 2) a novel framework of mobile RISs-enhanced indoor wireless networks is proposed, and a FL enhanced DRL algorithm is proposed for the deployment and beamforming optimization of the RIS. The average throughput of the indoor users severed by the mobile RIS is improved $15.1\%$ compared to the case of conventional fixed RIS; 3) A STAR-RIS assisted multi-user downlink multiple-input single-output (MISO) communication system is investigated, and a pair of hybrid reinforcement learning algorithms are proposed for the hybrid control of the transmitting and reflecting beamforming of the STAR-RIS, which ameliorate $7\%$ of the energy efficiency of the STAR-RIS assisted networks; 4) A tile-based low complexity beamforming approach is proposed for STAR-RISs, and the proposed tile-based beamforming approach is capable of achieving homogeneous data rate performance with element-based beamforming with appreciable lower complexity. By designing and operating the computer simulation, this thesis demonstrated 1) the performance gain in terms of sum data rate or energy efficiency by invoking the proposed RIS in the wireless communication networks; 2) the data rate or energy efficient performance gain of the proposed STAR-RIS compared to the existing reflecting-only RIS; 3) the effect of the proposed machine learning algorithms in terms of convergence rate, optimality, and complexity compared to the benchmarks of existing algorithms. | en_US |
