Unsupervised Learning for Passive Beamforming

Abstract

Reconfigurable intelligent surface (RIS) has recently emerged as a promising candidate to improve the energy and spectral efficiency of wireless communication systems. However, the unit modulus constraint on the phase shift of reflecting elements makes the design of optimal passive beamforming solution a challenging issue. The conventional approach is to find a suboptimal solution using the semi-definite relaxation (SDR) technique, yet the resultant suboptimal iterative algorithm usually incurs high complexity, hence is not amenable for realtime implementation. Motivated by this, we propose a deep learning approach for passive beamforming design in RISassisted systems. In particular, a customized deep neural network is trained offline using the unsupervised learning mechanism, which is able to make real-time prediction when deployed online. Simulation results show that the proposed approach maintains most of the performance while significantly reduces computation complexity when compared with SDR-based approach.Index Terms-Reconfigurable intelligent surface, passive beamforming, deep learning, unsupervised learning

Smart Radio Environments Empowered by Reconfigurable Intelligent Surfaces: How It Works, State of Research, and The Road Ahead

“…In [268], the authors employ deep learning techniques to reduce the design complexity of RIS-based wireless networks. In particular, the authors propose an unsupervised approach to optimize the RIS phase shifts.…”

Channel Estimation for Reconfigurable Intelligent Surface Aided MISO Communications: From LMMSE to Deep Learning Solutions

“…Some other recent works have proposed alternative data-driven techniques for constructing such beamformers [56]- [59]. One such approach [56] proposed a deep reinforcement learning based method for active and passive beamforming design, assuming perfect knowledge of the channel matrices, while other approaches [57], [58] have proposed to employ a fully-connected deep neural network (FNN) to directly learn the beamformers from the received pilot signals, without explicitly estimating the channels. In another approach [59], the authors proposed a codebook-based passive beamforming design, where a deep neural network is used to learn the mapping from the received pilot signals to the optimum beamformers from a predefined set of candidate beamformers.…”

“…For the data-driven channel estimators that we propose, we chose CNN-based architectures in favor of other data-driven techniques like FNN [57], [58]. This choice was made for a few different reasons.…”

A Deep Learning Based Modeling of Reconfigurable Intelligent Surface Assisted Wireless Communications for Phase Shift Configuration

“…The major difference between our approach and [16] is that our approach predicts the achievable rate for "any" Rx location and RIS reflection beamforming vector combination, which allows the system to determine proper action based on the prediction result of all or a subset of candidate RIS configurations during online inference phase even if some configurations are not seen during the training phase. Another two related works as discussed in Section I: [17] presented an unsupervised learning approach for passive beamforming in RIS-assisted communication environment and takes estimated channel as input to predict the optimal RIS phase shift configuration; [15] proposed a DL-based approach that learns to configure the RIS phase shifts and the beamforming matrix at the BS to maximize the system sum rate based on the received pilot. The main difference between our approach and [17] [15] is that our approach does not use either estimated channel information or received pilot as input since one of our main goals is to remove the dependency on explicit channel information; instead, our model takes the fused feature maps that are constructed from a set of RIS phase shift configurations and Rx location attributes to predict the corresponding achievable rate at the Rx for a given RIS-assisted communication network.…”

“…In the supervised learning setting: [13] presented a deep neural network (DNN) model that uses the sampled channel knowledge from a few active RIS elements as input to train the proposed DNN model offline to predict the optimal RIS reflection beamforming matrix; [14] and [15] presented methods that use received pilots as input to train the proposed DNNs to predict the optimal RIS phase shifts and beamforming vector at the base station (BS) while bypassing the intermediate step of channel estimation; the authors in [16] proposed a DNN model that is trained offline to learn the implicit relationship between the measured Rx coordinates and the optimal RIS configuration. To avoid the overhead of collecting labelled data in the supervised learning setting, the authors in [17] leveraged unsupervised learning technique and designed an RIS beamforming neural network (RISBFNN) architecture to predict the optimal phase shift configuration using estimated channels at BS as input and the negated transmission rate as the loss function. Another learning alternative, deep reinforcement learning (DRL), which uses the data collected online to train the model, has gained momentum in various wireless network scenarios, especially for optimization problems [18].…”