Deep-learning based optimization framework for wireless powered communication networks

Net-zero-energy networks enable communication when the connection to the electric grid or changing batteries is not feasible for the information sources by balancing harvested and consumed energy. Radio Frequency Energy Harvesting (RF-EH) is a key enabler for net-zero-energy networks as RF signals are independent from climate conditions, more predictable, and controllable. This thesis focuses on low-complexity resource allocation and optimization for RF-EH networks. In particular, the joint relay selection, scheduling, and power control problem in multiple-source-multiple-relay RF-EH network is formulated with the objective of minimizing the total schedule length and the constraints on data demand, energy causality, and maximum transmit power. The formulated problem is a mixed-integer non-linear problem and proven to be NP-hard. As a solution strategy, a bottom-up approach is followed by starting from a simpler scheduling and power control subproblem, then extending the problem with additional relay selection variables. First, the iterative algorithms for optimal and suboptimal solutions to the problem based on conventional optimization theory techniques are proposed. To address runtime concerns inherent in iterative algorithms, then, this thesis presents a novel approach integrating deep learning methods with established optimization principles. Feed-forward Deep Neural Network (DNN) architectures are presented to solve the scheduling and power control subproblem while leveraging optimality analysis for feature set extension and output layer simplification in the proposed DNN models. The remaining relay selection problem is reformulated as a classification task, successfully addressed through deep learning models, including a feed-forward DNN and two Convolutional Neural Network (CNN) based architectures. This thesis further introduces teacher-student learning, a process that transfers knowledge from a complex teacher network to a simpler student network, for an even lower complexity solution to the relay selection problem. A novel dichotomous-based neural architecture search algorithm designs the student network architecture. The results demonstrate the noteworthy reduction in runtime complexity achieved through deep learning models while maintaining optimality for the solution of resource allocation problems in RF-EH networks.