Power control with Machine Learning Techniques in Massive MIMO cellular and cell-free systems

Abstract

This PhD thesis presents a comprehensive investigation into power control (PC) optimization in cellular (CL) and cell-free (CF) massive multiple-input multiple-output (mMIMO) systems using machine learning (ML) techniques. The primary focus is on enhancing the sum spectral efficiency (SE) of these systems by leveraging various ML methods. To begin with, it is combined and extended two existing datasets, resulting in a unique dataset tailored for this research. The weighted minimum mean square error (WMMSE) method, a popular heuristic approach, is utilized as the baseline method for addressing the sum SE maximization problem. It is compared the performance of the WMMSE method with the deep Q-network (DQN) method through training on the complete dataset in both CL and CF-mMIMO systems. Furthermore, the PC problem in CL/CF-mMIMO systems is effectively tackled through the application of ML-based algorithms. These algorithms present highly efficient solutions with significantly reduced computational complexity [3]. Several ML methods are proposed for CL/CF-mMIMO systems, tailored explicitly to address the PC problem in CL/CF-mMIMO systems. Among them are the innovative proposed Fuzzy/DQN method, proposed DNN/GA method, proposed support vector machine (SVM) method, proposed SVM/RBF method, proposed decision tree (DT) method, proposed K-nearest neighbour (KNN) method, proposed linear regression (LR) method, and the novel proposed fusion scheme. The fusion schemes expertly combine multiple ML methods, such as system model 1 (DNN, DNN/GA, DQN, fuzzy/DQN, and SVM algorithms) and system model 2 (DNN, SVM-RBF, DQL, LR, KNN, and DT algorithms), which are thoroughly evaluated to maximize the sum spectral efficiency (SE), offering a viable alternative to computationally intensive heuristic algorithms. Subsequently, the DNN method is singled out for its exceptional performance and is further subjected to in-depth analysis. Each of the ML methods is trained on a merged dataset to extract a novel feature vector, and their respective performances are meticulously compared against the WMMSE method in the context of CL/CF-mMIMO systems. This research promises to pave the way for more robust and efficient PC solutions, ensuring enhanced SE and ultimately advancing the field of CL/CF-mMIMO systems. The results reveal that the DNN method outperforms the other ML methods in terms of sum SE, while exhibiting significantly lower computational complexity compared to the WMMSE algorithm. Therefore, the DNN method is chosen for examining its transferability across two datasets (dataset A and B) based on their shared common features. Three scenarios are devised for the transfer learning method, involving the training of the DNN method on dataset B (S1), the utilization of model A and dataset B (S2), and the retraining of model A on dataset B (S3). These scenarios are evaluated to assess the effectiveness of the transfer learning approach. Furthermore, three different setups for the DNN architecture (DNN1, DNN2, and DNN3) are employed and compared to the WMMSE method based on performance metrics such as mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE). Moreover, the research evaluates the impact of the number of base stations (BSs), access points (APs), and users on PC in CL/CF-mMIMO systems using ML methodology. Datasets capturing diverse scenarios and configurations of mMIMO systems were carefully assembled. Extensive simulations were conducted to analyze how the increasing number of BSs/APs affects the dimensionality of the input vector in the DNN algorithm. The observed improvements in system performance are quantified by the enhanced discriminative power of the model, illustrated through the cumulative distribution function (CDF). This metric encapsulates the model's ability to effectively capture and distinguish patterns across diverse scenarios and configurations within mMIMO systems. The parameter of the CDF being indicated is the probability. Specifically, the improved area under the CDF refers to an enhanced probability of a random variable falling below a certain threshold. This enhancement denotes improved model performance, showcasing a greater precision in predicting outcomes. Interestingly, the number of users was found to have a limited effect on system performance. The comparison between the DNN-based PC method and the conventional WMMSE method revealed the superior performance and efficiency of the DNN algorithm. Lastly, a comprehensive assessment of the DNN method against the WMMSE method was conducted for addressing the PC optimization problem in both CL and CF system architectures. In addition to, this thesis focuses on enhancing spectral efficiency (SE) in wireless communication systems, particularly within cell-free (CF) mmWave massive MIMO environments. It explores the challenges of optimizing SE through traditional methods, including the weighted minimum mean squared error (WMMSE), fractional programming (FP), water-filling, and max-min fairness approaches. The prevalence of access points (APs) over user equipment (UE) highlights the importance of zero-forcing precoding (ZFP) in CF-mMIMO. However, ZFP faces issues related to channel aging and resource utilization. To address these challenges, a novel scheme called delay-tolerant zero-forcing precoding (DT-ZFP) is introduced, leveraging deep learning-aided channel prediction to mitigate channel aging effects. Additionally, a cutting-edge power control (PC) method, HARP-PC, is proposed, combining heterogeneous graph neural network (HGNN), adaptive neuro-fuzzy inference system (ANFIS), and reinforcement learning (RL) to optimize SE in dynamic CF mmWave-mMIMO systems. This research advances the field by addressing these challenges and introducing innovative approaches to enhance PC and SE in contemporary wireless communication networks. Overall, this research contributes to the advancement of PC optimization in CL/CF-mMIMO systems through the application of ML techniques, demonstrating the potential of the DNN method, and providing insights into system performance under various scenarios and network configurations.