A graph reinforcement LearningPowered Online-Computational task offloading and latency minimization framework for wireless mobile edge computing networks

Abstract

Data processing capability of lower power networks can be improved by Mobile Edge Computing (MEC) extending to the wireless sensor networks and IoT. Creating a replication of MEC network with an offloading policy where a choice is made in the Wireless devices (WDs) for each computation task is the focus of this study. Deciding whether the task execution proceeds locally in the same environment or can be handed over to a remote MEC server, an optimized algorithm is needed which adopts task offloading decisions and wireless resource allocation in real time. But adopting this is a challenging solution to the real time fast combinatorial optimization problems, and impossible with the available traditional approaches. As a solution, heuristic algorithms encompassing Deep reinforcement learning (DRL) are emerging; however, it doesn’t make fair use of connection data like device-to-device interaction in MEC network. Moreover, heuristic algorithms rely on precise mathematical models for MEC systems which brought a new theory to the stage. This study revolves around this emerging technique relying on Graph neural networks (GNNs) learns from graph data while forwarding messages in the network. Utilizing GNN benefits, a Graph reinforcement learning-based online offloading framework (GROO) is proposed in this research, where the offloading policy is visualized as a graph state migration and MEC as an acyclic graph. The GROO achieves the lowest weighted task response latency (0.96 s) as compared to the existing DRL method (1.32 s) whereas on unseen circumstances and complex network topologies, GROO achieved lowest average latency up to 25 %.