Enhancing Computation Offloading In Wireless-Powered Mobile-Edge Computing Networks With Deep Reinforcement Learning For Online Optimization

Edge computing at the mobile frontier, enhanced by the integration of wireless energy, represents a cutting-edge strategy to boost processing performance in networks with limited energy resources, such as wireless sensor networks and the Internet of Things (IoT). This study investigates a mobile edge computing (MEC) framework powered by wireless energy, employing a dual-mode offloading scheme. In this paradigm, tasks from wireless devices (WD) may either be processed on-device or entirely shifted to an MEC server. In this approach, tasks from a wireless device (WD) are either processed locally or completely transferred to an MEC server. The objective is to create an online algorithm that can adjust task offloading and wireless resource allocation adaptively according to the variable conditions of the wireless channel. Conventional numerical optimization methods are inadequate due to the swift variations within the channel's coherence time. Our aim is to develop an algorithm that operates online and can dynamically adjust both offloading and resource distribution in response to the fluctuating state of the wireless channel. Traditional numerical optimization approaches fall short because they cannot swiftly adapt to the rapid changes in the channel's coherence. The solution we propose is a framework based on Deep Reinforcement Learning for Online Offloading that utilizes deep neural networks to incrementally learn from offloading decisions, thereby circumventing the need for complex combinatorial optimization. This leads to a significant reduction in the computational load, particularly in expansive networks. We've further enhanced this system with a method that enables real-time modification of the DROO algorithm's parameters. Our experiments demonstrate that this novel algorithm nearly achieves optimal efficiency and significantly reduces computation times—by more than ten times relative to existing techniques. For example, in a network with 30 users, DROO achieves CPU processing times of less than 0.1 seconds. This allows for optimal offloading decisions in real-time amidst the rapidly changing conditions of wireless environments.