The next generation wireless communication systems are anticipated to provide orders of magnitude increase in capacity and connect billions of devices constituting the future Internet of Things, due to the tremendous popularity of handhold smart devices, wearable electronics, sensors, and so on. Sustainable and cost-effective power supply for these IoT devices is critical to fulfill this objective. To this end, RF-based wireless power transfer (WPT) is proposed to replenish batteries of the IoT devices, because of a few remarkable advantages. Firstly, it is based on far-field radiation and thus allows for better mobility and scalability of the charging devices over a large geographic area. Besides, it can reuse the same set of antennas for information communication, enabling simultaneous information and power transfer. This supports the design of always-online wireless systems for IoT applications. The reuse of antennas also supports highly integrated system design to minimize the size and improve the reliability of IoT devices.
One of the main challenges in wireless powered communication lies in that the low efficiency in energy harvesting cannot fulfill the high-power consumption in legacy transceivers. Such unbalanced power supply and demand can be resolved by the development of wireless backscatter communications. The backscatter radios communicate in passive mode by modulating and reflecting the incident RF signals, without the need for power consuming components. Compared to the conventional active radios that operate on self-generated carrier signals, the backscatter radios consume orders of magnitude less power, enabling wireless power transfer a cost-effective and, most importantly, a feasible solution to sustain wireless communications for IoT devices. The integration of active and passive radios in one network opens a new paradigm of radio resource allocation and network optimization, especially when the radios can smoothly switch between two operating modes. The difference between the passive and active radio modes makes it possible for them to complement each other in data transmission. However, such difference also makes the resource allocation more complicated and requires innovative design for transmission scheduling and access control strategies in a hybrid radio network.
The design of future wireless networks needs to meet diverse Quality of Service (QoS) requirements. This calls for the network entities in nature to be cognitive of network environment and autonomous in decision making. Different network entities in the network layer, control layer, and management and orchestration layer, such as mobile devices, base stations, and SDN controllers need to make local and autonomous decisions, including spectrum access, channel allocation, power control, etc. to achieve the goals of different networks, e.g., throughput maximization, delay and energy minimization. As the modern networks have been becoming large-scale and complicated, we face a more decentralized, ad-hoc, and diverse network environment. The network control problems are very challenging as the dimensionality and computational complexity rapidly increase, due to the dynamic and uncertain network status, as well as strong couplings among different wireless users with heterogeneities in, e.g., QoS provisioning, wireless resource, air interface, and mobility. Deep reinforcement learning (DRL) has been developing as a promising solution to address high dimensional and continuous control problems effectively, by the use of deep neural networks (DNNs) as powerful function approximators. The integration of DRL into future wireless networks will revolutionize the conventional model-based network optimization to model-free approaches and meet various application demands. By interacting with the environment, DRL provides an autonomous decision-making mechanism for the network entities to solve non-convex, complex model-free problems, e.g., spectrum access, handover, scheduling, caching, data offloading, and resource allocation. This not only reduces the communication overheads but also improves network security and robustness. Though DRL has shown great potential to address emerging issues in complex wireless networks, there are still domain-specific challenges that require further investigation. These may include the design of proper DNN architectures to capture the characteristics of 5G network optimization problems, the state explosion in dense networks, multi-agent learning in dynamic networks, limited training data and exploration space in practical networks, the inaccessibility and high cost of network information, as well as the balance between information quality and learning performance.