AI Multi-Agent via Deep Reinforcement Learning for Resource Allocation and Task Offloading in Edge Computing Power Networks

Abstract

Edge computing deploys computing resources at the network edge to shorten the distance to access computing services. However, the limited computing power and heterogeneous distribution of edge servers pose huge challenges for handling the tremendous computing tasks generated by devices. To this end, Edge Computing Power Networks (EdgeCPN) is proposed to leverage wireless communication technology by integrating distributed computing power, to provide users with on-demand computing resources. In this paper, we first propose a comprehensive framework of EdgeCPN that facilitates collaborative computing power among devices, neighboring computing nodes, and edge servers. Then, we formulate task offloading and computing power allocation as a joint optimization problem to support as many devices as possible to complete real-time applications with high computing demands. Finally, we utilize Multi-Agent Deep Reinforcement Learning (MADRL) to explore optimal resource scheduling and task offloading strategies and use broadcasting to mitigate the communication overhead caused by information exchange between agents. Simulation results show that the proposed scheme can effectively improve the computing power allocation and reduce system cost by up to $\mathrm{6 8. 6 \%}$ compared with bench marks.