Aiming at the difficult regulation problem of modern computer network experimental environment under the condition of high dynamic, multi-objective and heterogeneity, this paper proposes an automatic regulation and performance optimization method based on Deep Reinforcement Learning (DRL). This method constructs a topology aware GNN-MAPPO framework that integrates graph neural networks (GNN) and multi-agent proximal policy optimization (MAPPO) algorithms to achieve high-dimensional modeling of network states and distributed collaborative decision-making. By designing a dynamic weighted reward mechanism, multi-objective conflicts such as throughput, delay and packet loss rate are balanced, and it is supported to adapt to different SLA (Service Level Agreement) scenarios as needed. The experiment was carried out on the OMNeT++ platform based on the Fat-Tree topology. The results show that compared with OSPF+ECMP, static SPF and traditional DRL methods, the proposed method can improve the throughput by 39%, reduce the delay by 38% and reduce the packet loss rate by 66%, and at the same time, it has stronger fault self-healing and dynamic adaptability. This study provides an intelligent and adaptive new paradigm for the next generation network experimental environment, which significantly reduces the manual configuration cost and improves the verification efficiency of complex network protocols.
C Chunliang Li (Sun,) studied this question.