数字孪生架构下基于 GAN 增强的多智能体深度强化学习边缘推理与异构资源协同优化

The deployment of large models at the edge has emerged as a pivotal enabler for the intelligent and digital transformation across various domains, including smart healthcare and urban systems. However, the heterogeneity of massive intelligent tasks and the unpredictability of high-dynamic networks pose significant challenges in the limited computational resources of edge devices to meet the demands of complex inference tasks for efficient and reliable Quality of Service (QoS). Therefore, this paper proposes an edge inference and heterogeneous resource collaborative optimization method based on Generative Adversarial Network (GAN) -enhanced Multi-Agent Deep Reinforcement Learning (MADRL), aiming to achieve dynamic load balancing of heterogeneous resources in Digital Twin (DT) -driven edge large model-enabled systems, ensuring the efficiency and reliability of inference tasks. First, this paper analyzes the physical network layer and twin network layer of the DT-driven edge large model system, and leverages GAN for twin mapping, enabling distributed processing, generation, and optimization of massive heterogeneous data. Next, the MADRL algorithm is applied to the comprehensive quantification and collaborative optimization of heterogeneous resources, with the edge inference data being fed back into the MADRL algorithm to reduce the data communication overhead during centralized training. Meanwhile, the framework utilizes federated learning, allowing for multi-party knowledge sharing, effectively improving model training speed and performance. Finally, simulation results demonstrate that the proposed algorithm reduces inference task delay and energy consumption, optimally utilizes system resources, and improves intelligent service quality in dynamic edge environments.