| 摘要: |
| 针对现有路由方案不适合学习图形结构信息，对陌生拓扑适应性不佳的问题，提出了一种基于图神经网络的软件定义网络(Software Defined Network，SDN)路由算法G-PPO。引入近端策略优化(Proximal Policy Optimization,PPO)强化学习算法实现模型训练，利用消息传递神经网络(Massage Passing Neural Network,MPNN)对网络拓扑进行学习，通过调整链路权重完成路由路径的调整。G-PPO将图神经网络对网络拓扑信息的感知能力和深度强化学习的自主学习能力有效结合，提升路由策略的性能。实验结果表明，与相关算法比较，所提算法的平均时延和丢包率、网络链路利用率和吞吐量指标均为最优。在3种不同拓扑上，该算法较其他算法最少提升10.5%吞吐量，最多提升95.6%丢包率，表明所提算法具有更好的适应不同网络拓扑的能力。 |
| 关键词: 软件定义网络 路由优化 图神经网络 深度强化学习 近端策略优化 |
| DOI：10.20079/j.issn.1001-893x.231114002 |
|
| 基金项目:国家自然科学基金青年科学基金项目(61901358) |
|
| Optimization of SDN Routing Algorithm Based on Graph Neural Network |
| ZHANG Xiaoli,TANG Yingqi,SONG Wanying |
| (School of Communication and Information Engineering,Xi’an University of Science and Technology,Xi’an 710600,China) |
| Abstract: |
| For the problems that existing routing schemes are not suitable for learning graph structure information and have poor adaptability to unfamiliar topologies,a software defined network(SDN) routing algorithm based on graph neural network called G-PPO is proposed.Proximal policy optimization(PPO) reinforcement learning algorithm is introduced to realize model training,massage passing neural network(MPNN) is used to learn network topology,and route adjustment is completed by adjusting link weights.G-PPO effectively combines the perception ability of graph neural network to network topology information with the autonomous learning ability of deep reinforcement learning to improve the performance of routing strategies.Experimental results show that compared with related algorithms,the proposed algorithm has the best average delay,packet loss rate,higher network link utilization rate and throughput.In three different topologies,the throughput and packet loss rate of proposed algorithm are improved by at least 10.5% and at most 95.6% respectively compared with those of other algorithms,indicating that the algorithm has better ability to adapt to different network topologies. |
| Key words: software defined network routing optimization graph neural network deep reinforcement learning proximal policy optimization |
