Dynamic Scheduling Algorithm for Edge-cloud Collaborative Computing Resources Based on 5G NR-U and Distributed Large Model
With the deployment of 5G NR-U (5th Generation New Radio in Unlicensed Spectrum) in unlicensed spectrum and the rise of sparse large-scale models (exemplified by Switch Transformers) in edge computing, edge-cloud collaborative training faces the dual challenges of dynamic spectrum contention and heterogeneous load distribution. This paper addresses the coupled uncertainty of time-varying 5G NR-U channels and the random arrival of expert tasks in MoE (Mixture of Experts) by introducing a dynamic scheduling algorithm for edge-cloud collaboration based on MADDPG (Multi-Agent Deep Deterministic Policy Gradient). This algorithm models each edge server as an agent and, based on local observations, jointly decides on task offloading targets (terminal, edge, or cloud), expert cache replacement, and 5G NR-U contention window parameters through a continuous action space, achieving cross-layer collaborative optimization of communication and computing cache resources. A centralized critic network is used for global training, and the reward function integrates average iterative latency, communication energy consumption, system throughput, and load balancing. On a simulation platform comprising eight heterogeneous edge nodes and one cloud center, three scenarios—baseline, high-interference, and burst load—were implemented for validation. Experimental results show that the proposed MADDPG algorithm outperforms benchmark algorithms such as CO (Cloud-Only), GO (Greedy Offloading), SEP (Static Expert Partitioning), DDPG (Deep Deterministic Policy Gradient), COMA (Counterfactual Multi-Agent Policy Gradients), and MAPPO (Multi-Agent Proximal Policy Optimization) in terms of average iteration latency (minimum 78 ms), system throughput (maximum 185 K tokens/s), expert load balancing (minimum standard deviation 15.3), and communication energy consumption (minimum 52 J). It demonstrates superior adaptability, robustness, and collaborative efficiency under scenarios of severe channel fluctuations and burst load, providing a feasible intelligent scheduling solution for efficient training of sparse large models in dynamic edge environments.