基于自注意力深度强化学习的特定流路由选择算法  

Specific flow routing selection algorithm based on Self-Attention deep reinforcement learning

作　　者：袁帅 张慧 蔡安亮 沈建华[1] YUAN Shuai;ZHANG Hui;CAI Anliang;SHEN Jianhua(School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;CypressTel Shenzhen Communication Technology Company,Shenzhen Guangdong 518000,China)

机构地区：[1]南京邮电大学通信与信息工程学院,南京210003 [2]深圳赛柏特通信技术有限公司,广东深圳518000

出　　处：《光通信技术》2024年第3期7-12,共6页Optical Communication Technology

基　　金：国家自然科学青年基金项目(62301284)资助;南京邮电大学企业委托研发重点课题(KH0020322072)资助。

摘　　要：为了有效降低传统流量工程机制中重路由对网络带来的负面影响,基于软件定义网络的全局网络视角和管理能力,提出一种基于自注意力深度强化学习的特定流路由选择算法,以重新路由少量流量达到接近最优的性能。通过多尺度融合注意力机制的神经网络模型来提取流量的特征,并采用集中式训练-分布式执行架构,根据观测网络状态做出实时决策。理论研究和实验结果表明,与传统深度强化学习算法与启发式算法相比,所提算法在平均负载和端到端延迟性能方面均有显著改进。To effectively mitigate the negative impact of rerouting on the network in traditional traffic engineering mechanisms,this paper proposes a specific flow routing selection algorithm based on Self-Attention deep reinforcement learning,leveraging the global network perspective and management capabilities of software-defined networking,to reroute a small amount of traffic and achieve near-optimal performance.A neural network model with multi-scale fusion attention mechanism is used to extract features of traffic,and a centralized training distributed execution architecture is adopted to make real-time decisions based on the observed network state.The theoretical research and experimental results show that compared with traditional deep reinforcement learning algorithms and heuristic algorithms,the proposed algorithm has significant improvements in average load and end-to-end delay performance.

关 键 词：软件定义网络 多智能体深度强化学习 流量工程 负载均衡 

分 类 号：TN91[电子电信—通信与信息系统]