Deep Learning-based Channel Estimation And Beamforming For RIS-assisted Communication Systems

While the coming“Internet of Everything”era could significantly improve human connectivity and advance social progress,it also has urgent requirements for wireless com-munication capacity,coverage,and stability.However,the advancement and development of wireless communications may also be hampered by increasingly dense communication networks’high energy and hardware expenses.Recently,reconfigurable intelligent sur-face(RIS)stood out as a hot research topic among the many emerging technologies for future wireless communication scenarios and requirements due to its low cost,low energy consumption,programmability,and ease of deployment.Specifically,RIS is a real-time programmable artificial electromagnetic surface with many controllable electromagnetic units arranged.By applying control signals to these electromagnetic units,their electro-magnetic characteristics can be dynamically adjusted to create electromagnetic fields with controllable parameters such as amplitude,phase,and frequency.Therefore,RIS can ad-just and reconfigure the wireless communication environment.Since the mechanism of RIS to change the wireless environment generally does not need to be equipped with high-energy and high-cost devices such as RF chains,it can significantly save hardware cost and energy consumption in practical deployment.To exploit the advantages of RIS as-sisted communication systems,the channel estimation and the beamforming design of the system are two critical issues.Most current researches on these two issues are based on traditional communication and signal-processing techniques.However,the performance of these traditional techniques often depends on the accuracy of mathematical models,and thus hardly adapt and cope with the drastic changes in the environment and the in-terference in real communication.In addition,traditional technical means usually entail a sizeable computational overhead,which likewise restricts the application of RIS sys-tems.To this end,this dissertation investigates the channel estimation and beamforming problems of RIS in different scenarios.And by adopting deep learning techniques,this dissertation reduces the channel estimation overhead and solves the problem of nonlinear optimization of complex scenarios in beamforming.The main work and contributions of this dissertation are described as follows:1.To address the channel estimation problem in the RIS-assisted multiple-input single-output(MISO)communication system,this dissertation analyzes the two-timescale characteristic of the system’s channel firstly.Based on this characteristic,a three-stage joint channel decomposition and estimation framework is proposed.Then,based on the long short-term memory(LSTM)network technique in deep learning,an algorithm based on a sparse-connected LSTM(SCLSTM)network is developed to accomplish the chan-nel decomposition and estimation tasks.The theoretical analysis shows that the proposed algorithm has a negligible time overhead and computational complexity than the conven-tional channel estimation algorithms.In addition,the simulation experiments also demon-strate that the proposed algorithm can obtain the channel state information(CSI)more robustly and accurately.2.To address the beamforming problem in the RIS-assisted MISO communication systems,this dissertation constructs base station precoding and RIS reflection coefficient optimization models with finite discrete phase shift constraints for perfect and imperfect CSI cases.A deep quantization neural network(DQNN)algorithm is proposed to em-bed the finite discrete phase shift constraint into the beamforming problem for holistic optimization.An improved DQNN(I-DQNN)algorithm is proposed to further solve the beamforming problem with different phase shift quantization resolutions of RIS reflect-ing units.Simulation experiments show that the proposed DQNN and I-DQNN algorithms have comparable performance under different experimental settings and outperform the conventional comparison methods,which decouple the finite discrete phase shift optimiza-tion and the beamforming.The experiments also give guidance criteria for using DQNN and I-DQNN with different phase shift quantization resolutions of RIS reflecting units.In addition,the computational complexity and convergence of the proposed algorithms are analyzed,and it is shown that it has less computational complexity and faster convergence than the conventional algorithm.3.For the RIS-assisted multiple-input multiple-output(MIMO)communication sys-tems,this dissertation first proposes a CSI-based real-time joint beamforming algorithm based on Deep Reinforcement Learning(DRL),which jointly designs the transmit co-variance matrix at the base station and the reflection coefficient of RIS.Due to the di-mensionality of the channels to be estimated and the corresponding pilot overhead in the RIS-assisted MIMO system are much larger than that of the MISO system,this disser-tation proposed a beamforming algorithm based on DRL and devices’relative location information to avoid the complicated acquisition process of CSI.Meanwhile,an imitation environment network is built to replace the actual environment to reduce the energy and time required for the interaction between the agent and the actual environment in the DRL algorithm.Simulation experiments show that the performance of the proposed DRL-based algorithm with CSI is close to that of the traditional alternating optimization(AO)algo-rithm.The location-based DRL algorithm achieves a higher average achievable rate by significantly saving the overhead of the channel estimation.4.For the beamforming design in the RIS-assisted cell-free(CF)communication sys-tems,this dissertation first constructs a distributed weighted sum-rate(WSR)problem to jointly design each base station’s precoding and the RIS’s reflection coefficient by refor-mulating the WSR problem into a mathematically tractable form and optimizing it.Then,a distributed alternate direction method of multipliers(ADMM)algorithm is proposed for the reformulated problem.Inspired by model-based algorithm unrolling,this dissertation unrolls the distributed ADMM algorithm into a deep distributed ADMM(D~2-ADMM)network framework.The D~2-ADMM combines general domain knowledge with the in-ference capability of deep learning to obtain high-performance beamforming solvers and hyperparameters through end-to-end learning.Simulation experiments show that the pro-posed D~2-ADMM algorithm requires only a tiny number of iterations to converge to a per-formance approximating the centralized algorithm.Moreover,the proposed D~2-ADMM algorithm is also significantly improved compared to other distributed algorithms.