The progress of modern industry has given rise to great requirements for network transmission latency and reliability in domains such as smart grid and intelligent driving.To address these challenges,the concept of Ti...The progress of modern industry has given rise to great requirements for network transmission latency and reliability in domains such as smart grid and intelligent driving.To address these challenges,the concept of Time-sensitive networking(TSN)is proposed by IEEE 802.1TSN working group.In order to achieve low latency,Cyclic queuing and forwarding(CQF)mechanism is introduced to schedule Timetriggered(TT)flows.In this paper,we construct a TSN model based on CQF and formulate the flow scheduling problem as an optimization problem aimed at maximizing the success rate of flow scheduling.The problem is tackled by a novel algorithm that makes full use of the characteristics and the relationship between the flows.Firstly,by K-means algorithm,the flows are initially partitioned into subsets based on their correlations.Subsequently,the flows within each subset are sorted by a new special criteria extracted from multiple features of flow.Finally,a flow offset selecting method based on load balance is used for resource mapping,so as to complete the process of flow scheduling.Experimental results demonstrate that the proposed algorithm exhibits significant advantages in terms of scheduling success rate and time efficiency.展开更多
联邦学习作为解决数据隔离问题的新兴范式,能够在不需要客户端上传原始数据的情况下训练全局模型,有效保护用户隐私。由于客户端数量众多但通信资源有限,只能选择部分客户端参与模型聚合。然而联邦学习系统存在设备异构和数据异质等挑战...联邦学习作为解决数据隔离问题的新兴范式,能够在不需要客户端上传原始数据的情况下训练全局模型,有效保护用户隐私。由于客户端数量众多但通信资源有限,只能选择部分客户端参与模型聚合。然而联邦学习系统存在设备异构和数据异质等挑战,简单的客户端选择策略无法考虑环境的动态特性,会拖慢模型的收敛速度,降低模型性能。考虑到客户端状态的时变,提出了全新的客户端可用性评估指标,建立了多重约束下的联邦学习客户端选择模型,建模为损失最小化问题;将优化问题转化为马尔可夫决策过程,提出了一种基于深度强化学习的联邦学习客户端自适应选择(Adaptive Selection for Clients in Federated Learning based on Deep Reinforcement Learning,ASC-DRL)算法,综合考虑通信延迟、资源消耗及客户端可用性,通过代理服务器与环境之间的持续交互最大化奖励函数,得到最优客户端选择方案。实验结果表明,提出的ASC-DRL算法相比于传统联邦学习算法,在模型精度和训练损失方面有着最高89.2%和99.8%的效果提升,能够有效适应动态环境变化,提升联邦学习整体性能和稳定性。展开更多
针对室内兴趣点(point of interest,POI)显著度评价问题,提出了一种基于数据平衡的随机森林(random forest,RF)模型。鉴于现有模型在处理数据不均衡及模拟显著度与影响指标之间复杂非线性关系方面的局限性,聚焦视觉、语义和结构三大维度...针对室内兴趣点(point of interest,POI)显著度评价问题,提出了一种基于数据平衡的随机森林(random forest,RF)模型。鉴于现有模型在处理数据不均衡及模拟显著度与影响指标之间复杂非线性关系方面的局限性,聚焦视觉、语义和结构三大维度,构建了包含34项特征的指标体系。通过合成少数类过采样技术(synthetic minority over-sampling technique,SMOTE)有效缓解数据不平衡和基于重要性权重的特征优化,形成了随机森林评价模型。实验结果表明,该模型在室内POI数据集上展现出卓越的性能,其准确率、精确度、召回率、加权F1分数和曲线下面积分别达到了0.987、0.984、0.987、0.987和0.999;与未进行数据均衡处理的RF模型相比,性能提升了一倍;与其他模型(如支持向量机、遗传规划算法)相比,性能分别提升了15%和5%,且在测试集上也显示出了良好的泛化性能。展开更多
基金supported by Science and Technology Project of State Grid Corporation Headquarters under Grant 5108-202218280A-2-170-XG(Development and Application of Power Time-Sensitive Network Switching Chip。
文摘The progress of modern industry has given rise to great requirements for network transmission latency and reliability in domains such as smart grid and intelligent driving.To address these challenges,the concept of Time-sensitive networking(TSN)is proposed by IEEE 802.1TSN working group.In order to achieve low latency,Cyclic queuing and forwarding(CQF)mechanism is introduced to schedule Timetriggered(TT)flows.In this paper,we construct a TSN model based on CQF and formulate the flow scheduling problem as an optimization problem aimed at maximizing the success rate of flow scheduling.The problem is tackled by a novel algorithm that makes full use of the characteristics and the relationship between the flows.Firstly,by K-means algorithm,the flows are initially partitioned into subsets based on their correlations.Subsequently,the flows within each subset are sorted by a new special criteria extracted from multiple features of flow.Finally,a flow offset selecting method based on load balance is used for resource mapping,so as to complete the process of flow scheduling.Experimental results demonstrate that the proposed algorithm exhibits significant advantages in terms of scheduling success rate and time efficiency.
文摘联邦学习作为解决数据隔离问题的新兴范式,能够在不需要客户端上传原始数据的情况下训练全局模型,有效保护用户隐私。由于客户端数量众多但通信资源有限,只能选择部分客户端参与模型聚合。然而联邦学习系统存在设备异构和数据异质等挑战,简单的客户端选择策略无法考虑环境的动态特性,会拖慢模型的收敛速度,降低模型性能。考虑到客户端状态的时变,提出了全新的客户端可用性评估指标,建立了多重约束下的联邦学习客户端选择模型,建模为损失最小化问题;将优化问题转化为马尔可夫决策过程,提出了一种基于深度强化学习的联邦学习客户端自适应选择(Adaptive Selection for Clients in Federated Learning based on Deep Reinforcement Learning,ASC-DRL)算法,综合考虑通信延迟、资源消耗及客户端可用性,通过代理服务器与环境之间的持续交互最大化奖励函数,得到最优客户端选择方案。实验结果表明,提出的ASC-DRL算法相比于传统联邦学习算法,在模型精度和训练损失方面有着最高89.2%和99.8%的效果提升,能够有效适应动态环境变化,提升联邦学习整体性能和稳定性。
