分布式流言push-sum无梯度算法
发布时间：2018-11-18 19:45
【摘要】：研究多个体网络中所有个体目标函数之和最小值问题,其中每个个体仅知其自身目标函数且仅可与其邻居个体交互信息。鉴于个体目标函数通常非光滑,同时个体间单变量信息通信有一定局限性,本文提出一种分布式流言push-sum无梯度算法求解此优化问题。假设每个个体都具有一个服从泊松分布的控制时钟,时钟的每次转动表示随机选择的个体之间进行信息更新。进一步地,在网络连通条件下证明了所提算法的收敛性。数值仿真结果表明,与现有的分布式流言无梯度优化算法相比,本文算法具有更快的收敛速度。
[Abstract]:In this paper, we study the minimum value of the sum of all individual objective functions in a multi-agent network, where each individual only knows its own objective function and can only interact with its neighbors. In view of the fact that the individual objective function is usually non-smooth and the single variable information communication between individuals has some limitations, this paper presents a distributed gossip push-sum algorithm without gradient to solve this optimization problem. Assuming that each individual has a control clock with a Poisson distribution, each rotation of the clock represents the updating of information between randomly selected individuals. Furthermore, the convergence of the proposed algorithm is proved under the condition of network connectivity. The numerical simulation results show that the proposed algorithm has a faster convergence speed than the existing distributed gossip without gradient optimization algorithm.
【作者单位】： 安徽理工大学数学与大数据学院;
【基金】：国家自然科学基金资助项目(61472003) 高校学科(专业)拔尖人才学术资助重点项目(gxbjZD2016049) 安徽省学术和技术带头人及后备人选科研活动经费资助项目(2016H076)
【分类号】：TP18
本文编号：2341008
[Abstract]:In this paper, we study the minimum value of the sum of all individual objective functions in a multi-agent network, where each individual only knows its own objective function and can only interact with its neighbors. In view of the fact that the individual objective function is usually non-smooth and the single variable information communication between individuals has some limitations, this paper presents a distributed gossip push-sum algorithm without gradient to solve this optimization problem. Assuming that each individual has a control clock with a Poisson distribution, each rotation of the clock represents the updating of information between randomly selected individuals. Furthermore, the convergence of the proposed algorithm is proved under the condition of network connectivity. The numerical simulation results show that the proposed algorithm has a faster convergence speed than the existing distributed gossip without gradient optimization algorithm.
【作者单位】： 安徽理工大学数学与大数据学院;
【基金】：国家自然科学基金资助项目(61472003) 高校学科(专业)拔尖人才学术资助重点项目(gxbjZD2016049) 安徽省学术和技术带头人及后备人选科研活动经费资助项目(2016H076)
【分类号】：TP18
【相似文献】
相关期刊论文 前2条
1 李国丽;盛大宁;王俊椋;景佳;王超;闫冰;;IMRT逆向计划中的混合多目标梯度算法(英文)[J];Transactions of Nanjing University of Aeronautics & Astronautics;2010年01期
2 陈云浩,郭达志,周廷刚,陶康华;热图像的梯度噪声剔除法及其温度场的标定[J];遥感信息;1999年02期
相关会议论文 前1条
1 蔡文澜;王俊生;陶军;徐惠斌;马宏绪;;一种PEGASUS策略梯度算法的理论及应用[A];中国仪器仪表学会第九届青年学术会议论文集[C];2007年
，本文编号：2341008
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2341008.html
