Fog-aided wireless communications and machine learning

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

The exponentially increasing demand for data, computation, low latency and reliable communications requires the balancing and exploitation of both edge and cloud processing in Fog-Radio Access Network (Fog-RAN) architectures to enable new use cases such as telemedecine, automated driving and privacy preserving distributed machine learning. This thesis considers enhanced Mobile BroadBand (eMBB), Ultra-Reliable and Low Latency Communications (URLLC), massive Ma-chine Type Communications (mMTC) as well as distributed edge learning, or edge-AI, in Fog-RAN architectures.
First, the coexistence of eMBB and URLLC services is considered in a multi-cell Fog-RAN architecture. Both the uplink and downlink are considered in the presence of practical assumptions such as fading channels and lack of channel state information at the URLLC transmitter. Several encoding and decoding schemes involving both the cloud and the edge are considered, namely, puncturing, treating interference as noise and successive interference cancellation. Orthogonal and non-orthogonal multiple access are analyzed and rate expressions are derived in addition to the reliability and latency for URLLC.
Second, mMTC is considered in Fog-RAN architectures. More specifically, In-ternet of Things devices are assumed to monitor a quantity of interest and transmit their observations to the edge node in the corresponding cell. The goal is to cor-rectly detect the underlying value of the quantity of interest. Both edge and cloud detection are considered and the performance is evaluated through the probability of error and error exponents as function of key system parameters such as interference power and fronthaul link capacity.
Finally, we turn our focus to edge-AI, commonly referred to as federated learning. We consider a Bayesian federated learning approach where devices aim to approximate a global posterior distribution over the model parameters instead of point estimates. We present a new non-parametric Bayesian federated learning algorithm that encodes the approximate posterior through deterministic samples, or particles. The algorithm is evaluated in terms of accuracy, convergence speed, communication load and calibration.
Date of Award1 Aug 2021
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorOsvaldo Simeone (Supervisor)

Cite this

'