Large Models for Resource Allocation in Edge Computing Power Networks

Computing Power Network (CPN), as an integration of heterogeneous computing resources, has turned to be difficult to cope with the rapid growth of end users and strict task delay constraints due to the possible long-distance transmission between its service facilities and remote users. To overcome this problem, we propose an Edge Computing Power Network (ECPN), which reduces long-distance transmission overhead, optimizes resource utilization, and improves system performance by fully utilizing edge computing resources. In the area of resource scheduling, traditional machine learning is widely used due to its fast problem solving capabilities. However, traditional machine learning faces challenges in ECPN, particularly with handling heterogeneous tasks and adapting to highly dynamic network conditions. To address these challenges, we propose a large model-enabled ECPN framework that enhances the capabilities of ECPN in task offloading and resource allocation while optimizing the deployment of large models in the framework. To demonstrate the advantages of this framework, we introduce a smart transportation application scenario and employ a large language model based graph multi-agent deep reinforcement learning algorithm to determine optimal task matching strategies that balance energy consumption and privacy protection in the ECPN. Experimental results demonstrate that the algorithm applied in this framework reduces energy consumption, decreases the convergence time for task matching, and enhances privacy protection in ECPN.