Date of Award

Fall 2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering (ENAS)

First Advisor

Tassiulas, Leandros

Abstract

Recent years have witnessed a rapid growth in both the number of connected devices and the volume of data transmitted among them. The unprecedented demands in networked systems bring imperative needs for resource allocation optimization approaches to avoid performance failures such as traffic congestion, overload of computational tasks, excessive energy consumption, etc. However, existing resource allocation approaches are significantly challenged by the increasing heterogeneity (e.g., device heterogeneity and data heterogeneity) in networks and the unpredictability of user demands. In this dissertation, we leverage state-of-the-art tools, such as machine learning, to overcome those challenges. We investigate three specific resource allocation scenarios in networked systems. First, we introduce a reinforcement learning based autoscaling algorithm to achieve a balance between the resource reconfiguration rate and the amount of resource consumption in data centers. Second, we propose a two-stage model pruning approach to accelerate federated learning without compromising the model accuracy. This is done by adaptively and automatically finding a small subset of important model parameters to use, based on the gradients observed from each parameter. Third, we improve the robustness of viewport prediction in 360-degree video streaming by personalizing prediction models to each user using meta learning. We adopt two meta models on each user for the prediction of the viewing directions and the video prefetch sizes, respectively. The two meta models are sensitive to new training data, so they can be quickly adapted to users' behaviors during the playing of videos. All proposed approaches are validated through extensive simulations, emulations, experiments, or proof-of-concept prototypes.

Download

Share

COinS