الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 133 |  |  |
| | | from | 133 | | |
Abstract
Nowadays, the unprecedented demands of high data rate and low latency
communications necessitate the efficient use of wireless communication resources.
Multiuser (MU)-MIMO scheduling allows multiple users to share the same time
frequency resources, exploiting the spatial diversity created by using multiple
antennas. This technique promises great enhancement of the spectral efficiency,
while at the same time reducing communication latency. However, high
computational complexity of the multi-user scheduler threatens these prospective
yields. This thesis aims to examine the MU-MIMO scheduler design and its challenges
and investigate the incorporation of various machine learning (ML) techniques into its
design to tackle those challenges.
First, a two-stage multi-user (MU)-MIMO scheduling algorithm, that balances
between the various communication system performance metrics, is proposed. The
performance and computational complexity of the proposed algorithm are analyzed.
The proposed algorithm has proven to increase the achieved system sum-capacity
with increasing the number of the multiplexed (i.e. grouped) users, achieving more
than 80%, on average, of the theoretical channel capacity upper bound in case of
grouping 4 users. However, the optimal solution of this algorithm implies high
computational complexity overhead which is impractical for real application.
Second, a ML-assisted MU-MIMO scheduler framework is structured and
proposed to handle the computational hurdle of the proposed scheduling algorithm.
Moreover, a low-dimensional feature set is designed for the ML models, that is based
on the domain-expert knowledge of the addressed user scheduling problem.
Third, two novel ML-based MU-MIMO uplink schedulers are implemented, by
employing the support vector machines (SVM) and deep neural network (DNN)
techniques. Numerical results are provided showing that the proposed ML-based
xi
schedulers achieve around 96% of the system sum-capacity achieved by the
conventional brute-force method, while reducing the computational complexity from
O(2
𝑁
× 𝑛
3
) in case of optimal conventional scheduler, where N is the number of the
scheduled users, to O(d × 𝑛
3
) in case of DNN-based scheduler and O(d × 𝑛
) in case of
SVM-based scheduler, where d refers to the ML feature set dimensionality.