Machine Learning based Optimization for Next Generation Wireless Networks

Abstract

The proliferation of connected devices and the emergence of applications, such as machine learning, virtual reality, and internet-of-everything (IoE), has brought massive data traffic. The next-generation (6G) wireless networks are expected to support unprecedented high data rates and highly diverse applications. Several key enabling technologies have been proposed in order to meet the demand of the highly challenging requirements of 6G. In this thesis, we develop machine learning-based algorithms to solve the highly complex resource allocation problems in 6G wireless networks. In this thesis, we present five main contributions.

In the first contribution, we lay the foundation for the subsequent chapters by developing general framework for analyzing the performance of a hybrid visible light communication (VLC) and Radio Frequency (RF) wireless system, in which both VLC and RF subsystems utilize non-orthogonal multiple access (NOMA) technology. The proposed framework is based on realistic communication scenarios as it takes into account the imperfect channel-state information (CSI). In this context, tractable closed-form expressions are derived for the corresponding average sum-rate of NOMA-VLC and orthogonal frequency division multiple access (OFDMA)-VLC. Respective computer simulations corroborate the derived analytic results, and interesting theoretical and practical insights are provided, which will be useful in the effective design and deployment of conventional VLC and hybrid VLC-RF systems.

In the second contribution, we propose a deep Q-learning (DQL) framework that aims to optimize the performance of an indoor NOMA-VLC downlink network. In particular, we formulate a joint power allocation and LED transmission angle tuning optimization problem, in order to maximize the average sum rate and the average energy efficiency. The obtained results demonstrate that our algorithm offers a noticeable performance enhancement into the NOMA-VLC systems in terms of average sum rate and average energy efficiency, while maintaining the minimum convergence time, particularly for higher number of users. Furthermore, considering a realistic downlink VLC network setup, the simulation results have shown that our algorithm outperforms the genetic algorithm (GA) and the differential evolution (DE) algorithm in terms of average sum rate, and offers considerably less run-time complexity.

In the third contribution, we design a two-stage resource management framework in an Intelligent Reflecting Surfaces (IRS)-assisted VLC system. During the first stage, a maximum possible fairness (MPF) algorithm is presented in order to maximize the fairness amongst the users. In the second stage, deep Q-learning is exploited in order to maximize the overall spectral efficiency (SE). The corresponding numerical results have shown that the proposed DQL-MPF framework exhibits superior performance in terms of the overall SE, achieved at fast convergence rate. More specifically, when the noise power is high and the number of users is relatively large, the DQL-MPF algorithm achieves more than tenfold overall SE compared to the Baseline scheme. Moreover, the synergy between the MPF and the DQL algorithms is investigated. To this end, we demonstrate that the MPF algorithm maximizes the fairness amongst the users while the DQL algorithm maximizes the overall SE and improves the robustness against the noise.

In the forth contribution, we propose an optimization problem for an unmanned aerial vehicle (UAV)-enabled IRS-assisted LiFi system. The aim is to jointly optimize the deployment of the UAVs, the orientation of the IRS mirrors, and the LED-user association. Due to the non-convexity of the optimization problem, the DQL-MPF algorithm is leveraged in order to minimize the total transmit power of the UAV-LiFi access points (APs). The simulation results have shown the effectiveness of our proposed algorithm to minimize the total transmit power at a relatively fast convergence rate. Moreover, the effectiveness of jointly optimizing the UAV-enabled LiFi and the IRS system was demonstrated. Further key insights were drawn which can aid the design of efficient IIoT networks.

Finally, we propose an efficient and robust hybrid precoding technique for mmWave massive multiple input multiple output (MIMO) (mMIMO) systems. The proposed scheme is based on the Perceiver Neural Network (PNN), where the PNN accepts a noisy channel matrix as an input and gives the analog precoder and combiners as an output. The proposed architecture learns to reshape the input data in order to achieve the best accuracy through a novel trainable reshaping module that works hand-in-hand with a dynamic cross-attention module. Our design involves an offline phase, in which we build our realistic ray tracing-based mMIMO dataset to train our Perceiver-based hybrid precoder (PBHP), and an online phase, in which the PBHP is deployed to perform real-time inference for the near-optimal beamforming indices of the mmWave mMIMO system.

Date of Award	Dec 2022
Original language	American English
Supervisor	Sami Muhaidat (Supervisor)