Machine learning-based radio resource allocation for Beyond-5G massive MIMO networks
Thesis event information
Date and time of the thesis defence
Place of the thesis defence
OP auditorium (L10)
Topic of the dissertation
Machine learning-based radio resource allocation for Beyond-5G massive MIMO networks
Doctoral candidate
Master of Science (Technology) Nuwanthika Sandeepani Rajapaksha Rajapakshage
Faculty and unit
University of Oulu Graduate School, Faculty of Information Technology and Electrical Engineering, Centre for Wireless Communications - Radio Technologies
Subject of study
Communications Engineering
Opponent
Professor Jyri Hämäläinen, Aalto University
Custos
Professor Nandana Rajatheva, University of Oulu
Machine learning-based radio resource allocation for Beyond-5G massive MIMO networks
The goal of this thesis is to design novel low-complexity radio resource allocation algorithms in cellular multi-user multiple-input multiple-output (MU-MIMO) and cell-free massive MIMO (mMIMO) systems using machine learning.
First, cell-free mMIMO uplink power control via unsupervised learning is proposed. There, a deep neural network (DNN) is trained to directly optimize over the optimization objective to learn optimal power allocation outputs. This novel approach results in a simpler and more flexible data collection and model training stage and enables online learning to further improve performance. A more practical and low-cost cell-free network setup is then considered with transceivers with hardware impairments and limited fronthaul capacity links between the access points and the central processing unit.
Unsupervised learning-based joint uplink power control and fronthaul capacity allocation is proposed, which has a similar performance and significantly lower computational complexity compared to the optimization-based solution. The flexibility and generalizability of the DNN in changing network configurations is also shown to have promising results.
Next, transmit antenna muting in cellular MU-MIMO is investigated, to reduce the base station power consumption by utilizing only a subset of antenna elements while satisfying the per-user quality of service requirements. A neural antenna muting (NAM) solution is proposed to address the computational complexity of the combinatorial optimization problem, along with several heuristics including a greedy-based solution. In NAM, a classification DNN trained in a supervised manner using a custom loss function is implemented to approximate the outputs of one of the baseline solutions. NAM is shown to result in significant energy savings compared to the full array transmission, via numerical simulations performed using a 3GPP-compliant system-level simulator. It is also shown to have significantly lower computational complexity than the heuristic solutions showing the potential of the NAM approach for practical implementation.
First, cell-free mMIMO uplink power control via unsupervised learning is proposed. There, a deep neural network (DNN) is trained to directly optimize over the optimization objective to learn optimal power allocation outputs. This novel approach results in a simpler and more flexible data collection and model training stage and enables online learning to further improve performance. A more practical and low-cost cell-free network setup is then considered with transceivers with hardware impairments and limited fronthaul capacity links between the access points and the central processing unit.
Unsupervised learning-based joint uplink power control and fronthaul capacity allocation is proposed, which has a similar performance and significantly lower computational complexity compared to the optimization-based solution. The flexibility and generalizability of the DNN in changing network configurations is also shown to have promising results.
Next, transmit antenna muting in cellular MU-MIMO is investigated, to reduce the base station power consumption by utilizing only a subset of antenna elements while satisfying the per-user quality of service requirements. A neural antenna muting (NAM) solution is proposed to address the computational complexity of the combinatorial optimization problem, along with several heuristics including a greedy-based solution. In NAM, a classification DNN trained in a supervised manner using a custom loss function is implemented to approximate the outputs of one of the baseline solutions. NAM is shown to result in significant energy savings compared to the full array transmission, via numerical simulations performed using a 3GPP-compliant system-level simulator. It is also shown to have significantly lower computational complexity than the heuristic solutions showing the potential of the NAM approach for practical implementation.
Last updated: 8.1.2025