摘要
针对算力网络中算网资源协同不足、任务需求适配性差的问题,将算力路由问题建模为序列决策问题,提出了基于深度强化学习的算网协同动态路由调度算法。该算法借鉴混合专家模型思想,针对时延敏感型、普通型以及计算密集型3类典型场景,设计了基于编码器-解码器结构的差异化专家网络进行专项优化,并通过动作屏蔽机制约束路由选择空间,实现高效的逐跳决策,输出包含最优计算节点的路径。仿真实验结果表明,相较于其他路由调度算法,该算法在服务成功率上提升约17%,降低了端到端时延,优化了节点间的负载均衡度,展现出良好的网络拓扑适应性,能够有效满足多样化计算任务的差异化需求。
To address the issues of insufficient collaboration among computing resources and poor adaptability to task requirements in computing power networks,the computing power routing problem was modeled as a sequential decision problem.A deep reinforcement learning-based computing-aware routing algorithm was proposed for dynamic routing scheduling of computing network collaboration.The idea of hybrid expert models was drawn on and a differentiated expert network was designed based on an encoder-decoder structure for specialized optimization in three typical scenarios:delay-sensitive,ordinary,and computationally intensive.The routing selection space was constrained through an action masking mechanism to achieve efficient hop-by-hop decision-making and output a path containing the optimal computing node.The simulation experiment results show that compared with other routing scheduling algorithms,the proposed algorithm improves service success rate by about 17%,reduces end-to-end latency,optimizes load balancing between nodes,demonstrates good network topology adaptability,and can effectively meet the differentiated needs of diverse computing tasks.
作者
越奇强
田乐
魏帅
胡宇翔
冯旭
董永吉
陈博
YUE Qiqiang;TIAN Le;WEI Shuai;HU Yuxiang;FENG Xu;DONG Yongji;CHEN Bo(Information Engineering University,Zhengzhou 450002,China;National Key Laboratory of Advanced Communication Networks,Zhengzhou 450002,China;Key Laboratory of Cyberspace Security,Ministry of Education,Zhengzhou 450002,China)
出处
《电信科学》
北大核心
2025年第8期33-50,共18页
Telecommunications Science
基金
国家重点研发计划项目(No.2024YFB2906704)
河南省重大专项课题项目(No.22110021090003)。
关键词
算力路由
算网融合
多场景优化
序列决策
深度强化学习
computing-aware routing
computing-network integration
multi-scenario optimization
sequential decision-making
deep reinforcement learning
