Message Board

Dear readers, authors and reviewers,you can add a message on this page. We will reply to you as soon as possible!

2020 Volume 42 Issue 9
Article Contents

Ji-ying HUANG, Jue-you LI. Distributed Push-Sum Subgradient Algorithm with Quantized Information[J]. Journal of Southwest University Natural Science Edition, 2020, 42(9): 106-114. doi: 10.13718/j.cnki.xdzk.2020.09.013
Citation: Ji-ying HUANG, Jue-you LI. Distributed Push-Sum Subgradient Algorithm with Quantized Information[J]. Journal of Southwest University Natural Science Edition, 2020, 42(9): 106-114. doi: 10.13718/j.cnki.xdzk.2020.09.013

Distributed Push-Sum Subgradient Algorithm with Quantized Information

More Information
  • Corresponding author: Jue-you LI
  • Received Date: 11/04/2018
    Available Online: 20/09/2020
  • MSC: O224

  • In a time-varying directed graph, this paper studies the problem for minimizing the sum of several local convex objective functions, which are known to each agent accordingly. In the real communication network, due to the bandwidth limitation of communication, each agent can only exchangequantized information with its neighbors. Combining the push-sum communication mechanism and distributed subgradient algorithm, a distributed push-sum subgradient algorithm with deterministic quantization is proposedin this paper. It is proved that ifthe step size satisfies a certain condition, the state of each agent converges to the neighborhood of the optimal solution of the problem. A numerical experiment shows that the higher the quantization precision, the closer it isto the optima.
  • 加载中
  • [1] NEDIĆ A, OLSHEVSKY A, OZDAGLAR A, et al. Distributed Subgradient Methods and Quantization Effects[C] //47th IEEE Conference on Decision and Control, New York: IEEE Computer Society Press, 2008.

    Google Scholar

    [2] NEDIĆ A, OLSHEVSKY A. Distributed Optimization Over Time-Varying Directed Graphs[J]. IEEE Transactions on Automatic Control, 2015, 60(3): 601-615. doi: 10.1109/TAC.2014.2364096

    CrossRef Google Scholar

    [3] NEDIĆ A, OLSHEVSKY A. Stochastic Gradient-Push for Strongly Convex Functions on Time-Varying Directed Graphs[J]. IEEE Transactions on Automatic Control, 2016, 61(12): 3936-3947. doi: 10.1109/TAC.2016.2529285

    CrossRef Google Scholar

    [4] 袁德明, 徐盛元, 赵环宇, 等.分布式多自主体优化问题中的概率量化影响研究[J].南京理工大学学报(自然科学版), 2011, 35(2): 209-212.

    Google Scholar

    [5] YUAN D M, XU S Y, ZHAO H Y, et al. Distributde Dual Averaging Method for Multi-Agent Optimization with Quantized Communication[J]. Systems and Control Letters, 2012, 61(11): 1053-1061. doi: 10.1016/j.sysconle.2012.06.004

    CrossRef Google Scholar

    [6] YUAN D M, MA Q, WANG Z. Dual Averaging Method for Solving Multi-Agent Saddle-Point Problems with Quantized Information[J]. Transactions of The Institute of Measurement and Control, 2014, 36(1): 38-46. doi: 10.1177/0142331213487545

    CrossRef Google Scholar

    [7] 刘琳, 杨丽芳.基于随机加速对偶下降算法的分布式网络流量优化[J].重庆邮电大学学报(自然科学版), 2014, 26(6): 838-844.

    Google Scholar

    [8] LI J Y, CHEN G, WU Z Y, et al. Distributed Subgradient Method for Multi-Agent Optimization with Quantized Communication[J]. Mathematical Methods in The Applied Sciences, 2017, 40(4): 1201-1213. doi: 10.1002/mma.4044

    CrossRef Google Scholar

    [9] LI J Y, WU C Z, WU Z Y, et al. Gradient-Free Method for Nonsmooth Distributed Optimization[J]. Journal of Global Optimization, 2015 (61): 325-340.

    Google Scholar

    [10] 叶李.分布式网络中CKNN查询的BORA优化汇聚[J].重庆邮电大学学报(自然科学版), 2016, 28(3): 435-442.

    Google Scholar

    [11] YI P, HONG Y G. Quantized Subgradient Algorithm and Data-Rate Analysis for Distributed Optimization[J]. IEEE Transactions on Control of Network Systems. 2014, 1(4): 380-392. doi: 10.1109/TCNS.2014.2357513

    CrossRef Google Scholar

  • 加载中
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Figures(1)

Article Metrics

Article views(1168) PDF downloads(10) Cited by(0)

Access History

Other Articles By Authors

Distributed Push-Sum Subgradient Algorithm with Quantized Information

    Corresponding author: Jue-you LI

Abstract: In a time-varying directed graph, this paper studies the problem for minimizing the sum of several local convex objective functions, which are known to each agent accordingly. In the real communication network, due to the bandwidth limitation of communication, each agent can only exchangequantized information with its neighbors. Combining the push-sum communication mechanism and distributed subgradient algorithm, a distributed push-sum subgradient algorithm with deterministic quantization is proposedin this paper. It is proved that ifthe step size satisfies a certain condition, the state of each agent converges to the neighborhood of the optimal solution of the problem. A numerical experiment shows that the higher the quantization precision, the closer it isto the optima.

  • 分布式优化在数字通信领域的理论及应用引起了广大学者的关注[1-11].

    受文献[3, 5]的启发,本文进一步考虑了在有向网络下的带确定型量化的分布式push-sum次梯度算法,并讨论了确定型量化对算法的收敛性影响.对于强凸分布式优化问题,在确定型通讯量化的情况下,我们提出的算法能收敛到最优解的附近,其收敛率为$O\left( {\frac{{\ln \tau }}{\tau }} \right)$.最后,数值实验表明量化精度越高,越接近最优.

1.   问题描述与量化
  • 考虑在一个具有n个节点的多个体网络中,每个个体仅知道其自身的目标函数,且仅与其邻居进行局部信息交换,使所有个体的目标函数之和达到最优,即如下的一个多个体网络分布式优化问题[2]

    其中每个个体i仅知道其局部目标函数fi(z):ℝd→ℝ,z是全局决策变量.

    定义1  如果对任意xy∈ℝd,有

    则称函数f是ℝdμ强凸函数,其中常数μ≥0,g(y)是f在点y的次梯度,这里‖·‖是欧几里得范数.

    定义2  如果对任意xy∈ℝd,有‖f(x)-f(y)‖≤Lx-y‖称函数fL-利普希茨连续.

    假设1  对每一个ii=1,…,n,函数fiμi-强凸函数;并且它是Li-利普希茨连续的,常数μi≥0,Li≥0.记$L = \sum\nolimits_{i = 1}^n {{L_i}} $.

    如果假设1成立,则问题(1)存在唯一全局最优解,记为z*,即

    若将网络中每个个体看作是一个节点,则多个体网络中个体间的信息通信可以表述为一个时变的有向图G(t)=(VE(t)),其中V={1,2,…,n}为顶点的集合,E(t)⊆V×V为边的集合表示个体间的信息交流过程.分别用Niin(t)和Niout(t)表示个体it时刻的内邻居点和外邻居点集,即

    di(t)表示个体it时刻的外度,即di(t)=|Niout(t)|.假设个体it时刻仅知道它的外度di(t).

    假设2  有向图序列{G(t)}是B强连通的,即存在正整数B,对∀k≥1使得Ek(t)∪Ek+1(t)∪…∪Ek+B-1(t)是连通的.

  • 假设所有的个体仅能向与之相邻的节点传递量化过后的信息,但是所有节点能存储实值的数据.本文采用确定型量化规则[4-5]如下:假设待量化的值x∈[-MM],M充分大;有l个采样点将区间[-MM]均匀分为l-1个子区间,采样点集合记为m={m1m2,…,ml},并且m1=-Mml=M,则每个子区间的长度为$\frac{1}{\Delta }$,即$\frac{1}{\Delta }$=mi+1-mi,定义

    这里$\left\lfloor x \right\rfloor $表示x向下取整到某个最近的采样点mi(i=1,…,l).令v=Q[x]-x表示量化误差,根据上述的量化规则,有如下关系式成立

2.   带通讯量化的push-sum分布式算法
  • 下面将研究经过量化后的分布式push-sum次梯度算法对优化问题(1)的影响.本文提出如下带量化的分布式push-sum次梯度算法(Q-PSDSG):

    第0步    初始化,对i=1,…,n,取xi(0)∈ℝd满足‖xi(0)‖有界;yi(0)=1.

    第1步    对每个i=1,…,n,更新

    第2步    t=t+1,转到第1步.

    这里步长α(t)>0是单调递减的,gi(t)是fizi(t)的随机次梯度.

3.   量化情形下收敛性分析
  • 为了简便,记所有个体的状态平均值为$\mathit{\boldsymbol{\bar x}}(t) = \frac{1}{n}\sum\limits_{j = 1}^n {{\mathit{\boldsymbol{x}}_j}} (t)$,个体i经过量化后的误差为vi(t)=Q[xi(t)]-xi(t).定义通讯权矩阵A(t)如下:

    引理1  若假设1和2成立,那么对∀y∈ℝdt≥0,有以下不等式成立:

    这里$q = \sum\limits_{i = 1}^n {L_i^2} $.

      由x(t)的定义和(6)式可得,

    y∈ℝd为任意变量,通过(7)式得:

    则有

    因为fi是强凸函数且是次梯度有界的,由文献[3]中引理2的证明类似可知对∀y∈ℝd

    注意到

    并结合(8),(9)式可以得到

    其中$q = \sum\limits_{i = 1}^n {{{\left( {{L_i}} \right)}^2}} $,从而引理1得证.

    下面给出一个关于算法Q-PSDSG的收敛性结果.

    定理1  假设1和假设2成立并假设步长α(t)单调递减且满足$\mathop {\lim }\limits_{t \to \infty } \alpha \left( t \right) = 0$,则对任意的i=1,…,nt≥0有

    其中

      在引理1中,取y=z*

    接下来估计F(x(t))-F(z*).一方面,根据假设1知F是强凸函数,从而有

    另一方面,F是利普希茨连续的,从而有

    由(11)和(12)式得到

    将(13)式代入(10)式中可以得到

    由初始条件和算法Q-PSDSG的第二步,可以得到,对任意i=1,…,nt≥0,‖zj(t+1)‖是有界的[2],从而‖zi(t+1)-z*‖有界,故令

    在(14)式中同时除以$\frac{{a\left( {t + 1} \right)}}{n}$,有

    将(15)式两边乘以t再对t递归求和,同时乘以$\frac{2}{{\tau \left( {\tau - 1} \right)}}$可得

    利用F的凸性,有

    结合(16)和(17)式即可得出结论.

    注1  定理1表明F(i(τ))接近最优F(z*)的程度主要由利普希茨常数、步长、概率量化精度和$\sum\limits_{t = 1}^{\tau - 1} {\left\| {{\mathit{\boldsymbol{z}}_j}\left( {t + 1} \right) - \mathit{\boldsymbol{\bar x}}\left( t \right)} \right\|} $所确定.

    引理2  若假设1和假设2成立,则对于算法中的序列{zi(t)},i=1,…,n和时间τ≥1有:

    其中:常数δ≥0,λ∈(0,1),D>0满足$\delta \ge \frac{1}{{{n^{nB}}}}$$\lambda \le {\left( {1 - \frac{1}{{{n^{nB}}}}} \right)^{\frac{1}{B}}}$.

    引理2的证明类似于文献[2]中引理1的证明.同时,在引理2的条件下还可以得到:

    定理2  在定理1的条件下,特别取α(t)=$\frac{p}{t}$t≥1,其中常数p满足

    则对每一个i

    这里

      将α(t)=$\frac{p}{t}$代入(14)式有

    两边同乘t(t+1),并对t递归求和,可以得到

    运用$\mathop {\max }\limits_i \left\| {{\mathit{\boldsymbol{z}}_i}\left( {t + 1} \right) - {\mathit{\boldsymbol{z}}^*}} \right\| = R$后,约去$\frac{p}{n}$再除以τ(τ-1)即得:

    因为F为凸函数,则有

    结合(20)式和引理2即可推导出

    简化处理(21)式以后定理2得证.

    注2  由定理2可以看出,不等式(19)的前3项是优化误差项,后3项是量化误差项.当迭代次数τ趋于无穷时,优化误差项趋于零.对每个个体i,量化误差项$\frac{1}{\Delta }{Q_1} + \frac{{\ln \tau }}{\Delta }{Q_2} + \frac{\tau }{{{\Delta ^2}}}{Q_3}$,取决于量化精度$\frac{1}{\Delta }$和常数Q1Q2Q3,故当量化精度足够高,即$\frac{1}{\Delta }$→0时,定理2中量化误差项也趋于零,从而F(i(τ))接近最优F(z*).

4.   数值实验
  • 考虑如下的一个多个体分布式强凸问题[8]

    这里aibiλ是已知的.记

    为个体i所持有的局部强凸目标函数.

    假设随机生成n个点对(aibi),取λ=0.1,按照定理2的要求取步长α(t).设置n=100,d=1和n=100,d=2,其中nd分别为节点个数、决策变量维数.在MATLAB中运行带量化的分布式push-sum次梯度算法,取Δ=2k,考虑k=6,8,10 3种不同的量化水平.考查最大误差

    图 1可以看出3种不同的量化水平对算法Q-PSDSG的影响,特别当量化精度越高,最大误差就越小,从而越接近最优,随着问题维数的增大,量化的误差也增大.

5.   结语
  • 本文研究了在有向网络下的带确定型量化的分布式push-sum次梯度算法,讨论了确定型量化对算法的收敛性影响.数值实验表明量化精度越高,越接近最优.

Figure (1)  Reference (11)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return