Research on Dynamic Computing Power Network Transmission Optimization Based on DGNN-PPO and a Variant of the Polya Ball Method

To meet the growing computational demands of expanding intelligence application scenarios, computing networks are becoming increasingly complex. However, operational challenges such as frequent node online/offline events and dynamic link status changes significantly complicate the coordinated scheduling of computational and transmission resources. Furthermore, traditional static optimization methods exhibit stability issues and slow convergence in medium-scale dynamic networks, limiting their ability to satisfy low-latency and high-reliability requirements for complex services. Therefore, we propose a dynamic computing network transmission optimization method that integrates Dynamic Graph Neural Networks (DGNN), dual virtual queues, and Proximal Policy Optimization (PPO) with Nesterov momentum acceleration. This approach aims to optimize network transmission performance during frequent node changes. This approach employs DGNN for real-time modeling of dynamic topology and node load characteristics, utilizes a dual virtual queue mechanism to characterize stability constraints of computational and transmission loads, and incorporates Nesterov momentum during policy optimization to enhance convergence speed and stability of policy updates. Simulation experiments indicate that the proposed method outperforms comparison models in queue stability, convergence performance, and quality of service, thereby improving transmission efficiency in dynamic computing networks.