Future generation wireless channel characterisation using machine learning techniques

Wireless communication is one of the most important technologies that is used in our modern society. The technology plays a crucial role for our work, businesses, entertainment and connectivity with other members of society. The importance of wireless communication networks even becomes higher in this current pandemic environment where people need to keep social distances to prevent the virus spread. Different versions of wireless technology, known as generations have been continuously developed and deployed since 1980’s when first generation (1G) of wireless communication system was introduced. In current decade, fifth generation (5G) is the candidate for deployment
with researchers already discussing beyond 5G wireless systems.
One of the main components of any wireless communication system is the wireless medium, known as wireless channel, which is used for data transmission between sender (transmitter) and receiver. Wireless channel plays an important role in defining the accuracy of data being communicated and the amount of data that can be transmitted simultaneously between transmitter and receiver(s). Being one of the most important components of a wireless system, understanding and characterising wireless channel has been a topic of research since the introduction of wireless communication. While discussing about wireless channel, there are several contemporary and statistical
Machine Learning (ML) techniques can now be conveniently applied to the wireless channel estimation problem to reduce the estimation complexity and enhance the system’s overall data rate. Traditionally, the wireless channel is estimated at the base station (BS), where BS receives the channel information from the users via the feedback channel. This poses a potential problem with the privacy and security of users’ channel information as well as requires large bandwidth for the feedback channel in a system with a
large number of users. This dissertation is about to investigate the wireless channel characterisation using contemporary and statistical Machine Learning (ML) techniques. Traditionally, wireless channel is characterised using statistical estimation techniques such as Mean SquareError (MSE) or Least Error techniques for a particular wireless system such as Orthogonal Frequency Division Multiplexing (OFDM). In some instances, the results of these
estimated wireless channels are compared with the actual measured wireless channel models and a generalised model is then developed. In addition, to address, the dual issues of users’ privacy and lessening the required
bandwidth for the feedback channel, this thesis proposes a Federation Learning (FL) based wireless channel estimation solution for a commonly used wireless communication system. In this technique, each user estimates its own channel and only shares the model parameters with the BS which then combines the information from all users to construct a joint wireless channel to be used for transmission for the next time slot. To evaluate the accuracy of our proposed model, this dissertation focus on using Stochastic Gradient Descent (SGD) to optimise the solution and compared the results with the results presented in [1]. Our numerical results show that the proposed model renders more stable and accurate estimation results in conjunction with the use of SGD and FL.