Cell-Free Massive MIMO NOMA Network with Machine Learning aided User Association and Power Allocation

Abstract

This dissertation explores the applications of Machine Learning (ML) in Cell Free Massive Multiple-Inputs-Multiple-Output (CF-mMIMO) Non-Orthogonal Multiple-Access (NOMA) Network that incorporates Unmanned Ariel Vehicles (UAVs) as mobile Access Points (APs) and Simultaneous Wireless Information and Power Transfer (SWIPT) for energy harvesting User Equipment (UE). Firstly, we propose an objective function to maximise Downlink (DL) Sum Spectral Efficient (SSE) whilst considering UE Spectral Efficiency (SE) requirements for both downlink and uplink transmission. To maximise the objective function, we propose using a Deep Reinforcement Learning (DRL) algorithm, namely, a Deep Deterministic Policy Gradient (DDPG), to do joint power allocation and user as sociation by using a hybrid action space. We find the optimal time allocation for energy harvesting for the frame structure through a heuristic approach. We show the DDPG model performs better than random power allocation and user association schemes. Secondly, we improve the performance of our DDPG model by using various bootstrapping methods described in literature. We also introduce a novel bootstrapped DDPG, we call a Multi-Head-Elastic-Step (MH-ES) DDPG. We compare the various bootstrapped DDPG models to the base DDPG and show that bootstrapping can improve the training performance and inference performance of the base DDPG. Lastly, we conclude our work by exploring future work to extend on our research.