Internet of Things (IoT) networks have gained significant attention in recent years as it has the potential to transform various industries. A key concern, however, is that sensing devices have limited operational lifetime. Specifically, they have finite energy, which affects the amount of data they are able to collect and upload. One solution is to power these devices wirelessly, where devices harvest energy from Radio Frequency (RF) signals from transmitters such as a Hybrid Access Point (HAP). A key issue, however, is that energy delivery and data transmissions may be conducted on the same frequency band. This means a HAP has to determine a transmission schedule for energy or/and data transmissions. Another issue is that the channel gain of devices varies over time, which affects the amount of harvested energy and transmitted data. In this respect, a challenging issue is that an HAP has causal channel state information only, meaning it is not aware of energy arrivals or channel gains nor the data rate of devices in future time slots. Given the above issues, this thesis first proposes a novel mode-based structure for an RF-powered IoT network. Specifically, an HAP operates either in downlink mode to charge devices, or in uplink mode to collect data from them. Unlike previous works which only consider one time slot, this thesis aims to maximize the amount of collected data over multiple time slots. The problem at hand is to determine the mode of each time slot. To do this, this thesis presents a novel rolling horizon algorithm. In particular, it uses channel estimates and an Integer Linear Program (ILP) to determine the operating mode of a HAP in each time slot. This thesis then studies a multi-antenna SIC-enabled HAP that uses the said mode structure. Specifically, the HAP is able to receive multiple transmissions simultaneously and charge devices via beamforming. In this respect, the problem is to schedule a set of devices to transmit in uplink slots, and determine antenna weights in downlink slots. To do this, this thesis employs a rolling horizon algorithm to solve a Mixed Integer Linear Program (MILP) to determine its operating mode in each time slot. In addition, this thesis also proposes a data-driven approach to take advantage of the massive computational power at data centers to construct neural networks. Specifically, in the offline stage, the approach generates exhaustive collection of channel gain scenarios and stores the optimal