Machine learning methods for future-generation wireless networks

Student thesis: Phd


The wireless communication systems, particularly the vehicle to everything (V2X) wireless networks, take advantage of machine learning (ML) to improve and overcome the issues facing classical methods of solving complex scenarios. This thesis assesses the new development in ML techniques for V2X in resource allocation, power efficiency, mobile intelligent reflecting surfaces (IRS), downlink-uplink decoupling, and end-to-end (E2E) optimised communication systems. A survey summary has been produced on applying machine learning to wireless communications. In addition, deep Q-learning (DQL) was investigated with the classical maximum weight matching (MWM) to improve the radio resource allocation in vehicle-to-vehicle (V2V) communications, where approximately 50% increase in the cellular users' throughput in comparison with only using the DQL at dense V2V links and three power transmission levels scenario. In addition, the DQL and QL research investigate unmanned aerial vehicles (UAV) energy efficiency (EE). Showing ML techniques power in the proposed QL and DQL algorithms that perform better than the baseline fractional power control (FPC) scheme using the UHF and mmWave bands compared to conventional FPC schemes, where the QL(DQL) EE improvement are around 80%(100%) for the UHF band, and by around 160% (170%) for the mmWave band, in comparison to conventional FPC scheme. Furthermore, the research explored the E2E wireless model by proposing embedded variational autoencoders (EVAE). The result shows that the new technique learns the wireless environment autonomously through latent random variables (LRV), and the performance for the V2X scenario has improved, where the EVAE 8dB SINR bit error rate (BER) has approximately 3 dB improvement in comparison to the 8-QAM uncoded theoretical value and 1 dB in comparison to the 8-QAM VAE scheme. Moreover, the IRS has been modelled as an E2E communication system. The inferring of the signal constellations representation pattern reduces the wireless environment contamination impact on the received signal and improves the SER performance of the proposed E2E system compared to the conventional AE and classical IRS, this improvement can be quantified by 8 dB gain in SER performance comparing to classical IRS for 64 meta-surfaces and -20 dB and RS assisted AE-based scheme Eb/No is -20 dB. Lastly, the federated learning (FL) for distributed ML has been proposed to assist in offloading data from the mobile network using the mmWave band to improve mobile IRS selection. The FL utilises the UHF band to create global and local deep neural network (DNN) models to enhance the performance of MIRS-assisted-V2V communication throughput and reliability. Mobile IRS (MIRS), as a newly proposed type of mobile network resource, improves the network's performance when the number of MIRS classified as LOS increases in the system the quantification of the result shows that with four MIRSs the system can achieve BER of 10-6 in compare to 0.011 BER when only one MIRS used.
Date of Award1 Aug 2023
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorKhairi Hamdi (Supervisor) & Emad Alsusa (Supervisor)


  • Machine Learning
  • Wireless communications

Cite this