Scalable User Association in Cell-Free Massive MIMO With Matching Theory and Reinforcement Learning
Salvato in:
| Pubblicato in: | PQDT - Global (2025) |
|---|---|
| Autore principale: | |
| Pubblicazione: |
ProQuest Dissertations & Theses
|
| Soggetti: | |
| Accesso online: | Citation/Abstract Full Text - PDF |
| Tags: |
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| Abstract: | Cell-free massive MIMO (CF-mMIMO) is expected to play a key role in the sixth generation of mobile networks due to its ability to deliver enhanced spectral and energy efficiency and improved robustness. It implements massive MIMO in a distributed way by utilizing a large number of small, low power access points (APs) distributed over a large coverage area. The APs cooperate to jointly serve multiple user equipments (UEs) simultaneously using the same frequency-time resources. In the canonical form of CF-mMIMO, all APs were assumed to serve all UEs which is impractical because each UE is located close to a few APs. Moreover, the resultant fronthaul load and computational complexity would be intractable if each AP had to receive downlink data from the CPU, send uplink estimates to the CPU and process the signals for all UEs in the network. This highlights the need for user-centric CF-mMIMO in which each user selects a set of APs to serve itself and these AP-clusters can overlap; thereby resulting in a cell-free architecture. The selection of the optimal AP-UE association is a combinatorial problem with high complexity that has attracted significant research interest.In this thesis, three novel AP selection algorithms are proposed with the goal of improving the spectral and energy efficiency. The first approach utilizes matching theory with partially selfish agents who make socially conscious association choices. As such, they prefer matchings that improve not only their individual utility but also those of their neighbours. Numerical simulations show that the proposed approach can achieve close to optimal performance and outperform existing approaches.In the second study, the AP association problem is modelled as a Markov Decision Process (MDP) and deep reinforcement learning (DRL) is applied to solve it. Due to the large number of APs and UEs, the action space is combinatorial and grows exponentially with the number of APs and UEs. This makes it challenging for standard DRL approaches due to the difficulty in exploration. The proposed method deals with this by designing a simple and elegant mapping algorithm that is added to the policy network to shape the action space; thereby making it more manageable. Simulation results show that the proposed method can improve the sum SE and EE performance compared to existing DRL approaches that instead rely on reward engineering to guide learning.The third method enhances the DRL approach by adding Behaviour Cloning (BC) to speed up training. Here, demonstrations from a sub-optimal heuristic teacher are used to guide the agent to a region of the action space with high reward, and thereafter the agent learns on its own using DRL to improve the initial policy. Numerical simulations verify that the agent can quickly meet the performance of the expert and thereafter outperform it, as well as existing approaches in the current literature. |
|---|---|
| ISBN: | 9798297669062 |
| Fonte: | ProQuest Dissertations & Theses Global |