Ph.D
Group : Networking & Stochastic and Combinatorial Optimization
Optimized beamforming and user-to-cell association for future dense networks
Starts on 01/10/2015
Advisor : BOUKHATEM, Lila
Funding : Contrat doctoral uniquement recherche
Affiliation : Université Paris-Saclay
Laboratory : LRI - ROCS
Defended on 30/09/2019, committee :
Directrice de thèse :
- Lila BOUKHATEM, Maître de Conférences, UPSud
Co-encadrante de thèse :
- Steven MARTIN, Professeur des Universités, UPSud
Rapporteurs :
- Nadjib AIT SAADI, Professeur, ESIEE Paris
- Mylene PISCHELLA, Maître de conférences, CNAM Paris
Examinateurs :
- Megumi KANEKO, Associate Professor, NII Tokyo
- Xavier LAGRANGE, Professeur, IMT Atlantique
Research activities :
Abstract :
Recently, mobile operators have been challenged by a tremendous growth in mobile data traffic. In such a context, Cloud Radio Access Network (CRAN) has been considered as a novel architecture for future wireless networks. The radio frequency signals from geographically distributed antennas are collected by Remote Radio Heads (RRHs) and transmitted to the cloud-centralized Baseband Units (BBUs) pool through fronthaul links. This centralized architecture enables a global optimization of joint baseband signal processing and radio resource management functions for all RRHs and users.
At the same time, Heterogeneous Networks (HetNets) have emerged as another core feature for 5G network to enhance the capacity/coverage while saving energy consumption. Small cells deployment helps to shorten the wireless links to end-users and thereby improving the link quality in terms of spectrum efficiency (SE) as well as energy efficiency (EE). Therefore, combining both cloud computing and HetNet advantages results in the so-called Heterogeneous-Cloud Radio Access Networks (H-CRAN) which is regarded as one of the most promising network architectures to meet 5G and beyond system requirements.
In this context, we address the crucial issue of beamforming and user-to-RRH association (user clustering) in the downlink of H-CRANs. We formulate this problem as a sum-rate maximization problem under the assumption of mobility and CSI (Channel State Information) imperfectness. Our main challenge is to design a framework that can achieve sum-rate maximization while, unlike other traditional reference solutions, being able to alleviate the computational complexity, CSI feedback and reassociation signaling costs under various mobility environments. Such gain helps in reducing the control and feedback overhead and in turn improve the uplink throughput.
Our study begins by proposing a simple yet effective algorithm baptized Hybrid algorithm that periodically activates dynamic and static clustering schemes to balance between the optimality of the beamforming and association solutions while being aware of practical system constraints (complexity and signaling overhead). Hybrid algorithm considers time dimension of the allocation and scheduling process rather than its optimality (or suboptimality) for the sole current scheduling frame. Moreover, we provide a cost analysis of the algorithm in terms of several parameters to better comprehend the trade-off among the numerous dimensions involved in the allocation process.
The second key contribution of our thesis is to tackle the beamforming and clustering problem from a mobility perspective. Two enhanced variants of the Hybrid algorithm are proposed: ABUC (Adaptve Beamforming and User Clustering), a mobility-aware version that is fit to the distinctive features of channel variations, and MABUC (Mobility-Aware Beamforming and User Clustering), an advanced version of the algorithm that tunes dynamically the feedback scheduling parameters (CSI feedback type and periodicity) in accordance with individual user velocity. MABUC algorithm achieves a targeted sum-rate performance while supporting the complexity and CSI signaling costs to a minimum.
In our last contribution, we propose to go further in the optimization of the CSI feedback scheduling parameters. To do so, we take leverage of reinforcement learning (RL) tool to optimize on-the-fly the feedback scheduling parameters according to each user mobility profile. More specifically, we propose two RL models, one based on Q-learning and a second based on Deep Q-learning algorithm formulated as a POMDP (Partially observable Markov decision process). Simulation results show the effectiveness of our proposed framework, as it enables to select the best feedback parameters tailored to each user mobility profile, even in the difficult case where each user has a different mobility profile.
