Abstract
The general idea of caching is replicating data at various locations across the networks. This helps bringing the data sources closer the end users, allowing users to get their desired data locally from caching points. Therefore, caching reduces power consumption and enables the use of unlicensed bandwidth, offering a higher bandwidth efficiency. In addition, caching is a solution for shifting the network load from peak to off-peak time by pre-downloading data to caching points in off-peak time, then, in peak hours those caching points will be capable of serving users locally. By acknowledging the advantages of caching in future wireless communication networks, the aim of this dissertation is to investigate different caching models, in addition, to propose optimal caching and resource allocation strategies for maximizing the system performance. In our first model, caching is deployed at a base station (BS) aiming to mitigate the amount of video data downloaded from a remote server. This is due a fact that video data type dominates a significant portion of data flows nowadays. We propose a machine learning framework to anticipate the popularity of unpublished videos by analyzing the relation and similarity of the content among videos. For dealing with the uncertainty of published videos’ popularities, we predict the change in the number of views based on the prediction with expert advice method. Finally, the optimal caching strategies is obtained via a formulated linear optimization problem. Another caching strategy is storing potentially popular data right at the users’ devices. This model, with the aid of device-to-device communication (D2D) strategy, not only allows users to retrieve the desired content items right from their devices, but also enables the short-range data sharing among them. Aiming to jointly minimize the backhaul load and the system consumed power, our problem is formulated as nonconvex optimization problem where we propose two solving schemes based on primal decomposition method and the alternative direction method of multipliers (ADMM). Besides the centralized algorithms, distributed versions are constructed for exploiting the computational capability at the user side. Besides that, caching can also be integrated into a non-orthogonal multiple access (NOMA) system to further enhance the overall performance. In this model, we exploit the caching memory resource at users’ devices to perform the interference cancellation for the superposed signal. Our goal is maximizing the probability that all users can successfully decode their desired signals which addresses both the system quality and the users’ fairness. Regarding this objective, two methods are proposed, following optimization and deep reinforcement learning approaches that allocate optimal power to the component signals of all users. Each method has its own practical advantages.