Message Board

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

2017 Volume 39 Issue 1
Article Contents

Xi-mei YAN, Yin-kui ZHANG, Yong-gang PEI. A Wide Neighborhood Infeasible-Interior-Point Method for Linear Programming[J]. Journal of Southwest University Natural Science Edition, 2017, 39(1): 92-98. doi: 10.13718/j.cnki.xdzk.2017.01.014
Citation: Xi-mei YAN, Yin-kui ZHANG, Yong-gang PEI. A Wide Neighborhood Infeasible-Interior-Point Method for Linear Programming[J]. Journal of Southwest University Natural Science Edition, 2017, 39(1): 92-98. doi: 10.13718/j.cnki.xdzk.2017.01.014

A Wide Neighborhood Infeasible-Interior-Point Method for Linear Programming

More Information
  • Received Date: 28/08/2015
    Available Online: 20/01/2017
  • MSC: O221.1

  • In this paper, we propose a wide neighborhood infeasible-interior-point method for linear programming.The characteristics of the proposed algorithm are as follows:on the one hand, the proposed algorithm is based on a wide neighborhood and has a good calculation results, which is showed by the experiments.On the other hand, by analyzing, we achieve the iteration complexity O(n1.5L) for the proposed algorithm, which is the best complexity result for a wide neighborhood infeasible-interior-point method.
  • 加载中
  • [1] KARMARKAR N. A New Polynomial-Time Algorithm for Linear Programming [J]. Combinatorica, 1984, 4(2): 302-311.

    Google Scholar

    [2] LUSTIG I J. Feasibility Issues in a Primal-Dual Interior-Method for Linear Programming [J]. Math Program, 1991, 49(1-3): 145-162.

    Google Scholar

    [3] MASAKAZU K, NIMROD M, SHINJI M. A Primal-Dual Infeasible-Interior-Point Algorithm for Linear Programming [J]. Math Program, 1993, 61(1-3): 263-280. doi: 10.1007/BF01582151

    CrossRef Google Scholar

    [4] ZHANG Y. On the Convergence of a Class of Infeasible Interior-Point Method for the Horizontal Linear Complementarity Problem [J]. SIAM J Optim, 1994, 4(1): 208-227. doi: 10.1137/0804012

    CrossRef Google Scholar

    [5] MIZUNO S. Polynomiality of Infeasible-Interior-Point Algorithms for Linear Programming [J]. Math Program, 1994, 67(1-3): 109-119. doi: 10.1007/BF01582216

    CrossRef Google Scholar

    [6] POTRA F A. An Infeasible-Interior-Point Predictor-Corrector Algorithm for Linear Programming [J]. SIAM J Optim, 1996, 6(1): 19-32. doi: 10.1137/0806002

    CrossRef Google Scholar

    [7] POTRA F A. A Quadratically Convergent Predictor-Corrector Method for Solving Linear Programs from Infeasible Starting Points [J]. Math Program, 1994, 67(1-3): 383-406. doi: 10.1007/BF01582228

    CrossRef Google Scholar

    [8] KOJIMA M, NOMA T S, YOSHISE A. Global Convergence in Infeasible-Interior-Point Algorithms [J]. Math Program, 1994, 65(1-3): 43-72. doi: 10.1007/BF01581689

    CrossRef Google Scholar

    [9] WRIGHT S J. Primal-Dual Interior-Point Methods [M]. Philsdephia: SIAM Publications, 1997.

    Google Scholar

    [10] YE Y. Interior Point Algorithms: Theory and Analysis [M]. Hoboken: John Wiley & Sons, 2011.

    Google Scholar

    [11] AI W, ZHANG S. An O(nL)-Iteration Primal-Dual Path-Following Method, Based on Wide Neighborhoods and Large Updates, for Monotone LCP [J]. SIAM J Optim, 2005, 16(2): 400-417. doi: 10.1137/040604492

    CrossRef Google Scholar

    [12] YANG X, ZHANG Y, LIU H. A Wide Neighborhood Infeasible-Interior-Point Method with Arc-Search for Linear Programming [J]. J App Math Comput, 2016, 51(1-2): 209-225. doi: 10.1007/s12190-015-0900-z

    CrossRef Google Scholar

    [13] 杨喜美, 刘红卫, 张因奎.带有新的迭代格式的内点算法[J].应用数学与力学, 2014, 35(9): 1063-1070.

    Google Scholar

    [14] BROWNE S, DONGARRA J, GROSSE E, et al. The Netlib Mathematical Software Repository [CP/OL]. (1995-09-15) [2015-06-10]. http: //www. dlib. org/dlib/september95/netlib/09browne. html.

    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

Tables(1)

Article Metrics

Article views(583) PDF downloads(32) Cited by(0)

Access History

Other Articles By Authors

A Wide Neighborhood Infeasible-Interior-Point Method for Linear Programming

Abstract: In this paper, we propose a wide neighborhood infeasible-interior-point method for linear programming.The characteristics of the proposed algorithm are as follows:on the one hand, the proposed algorithm is based on a wide neighborhood and has a good calculation results, which is showed by the experiments.On the other hand, by analyzing, we achieve the iteration complexity O(n1.5L) for the proposed algorithm, which is the best complexity result for a wide neighborhood infeasible-interior-point method.

  • 从1984年,Karmarkar[1]提出内点算法以来,内点算法的研究一直是一个热门的课题.随着研究的不断深入,研究者提出了许多不同类型的内点算法.从初始点的选取不同可分为:可行内点算法和不可行内点算法.不可行内点算法的优势是:不要求初始点严格可行,仅要求在非负象限的内部.由于这种类型的点容易获得,故在实际计算中显得尤为重要.

    Lustig[2]首次提出了不可行内点算法,随后Masakazu[3]提出了原始-对偶不可行内点算法.这些早期的著作均没有给出不可行内点算法的多项式复杂度,直到Zhang[4]给出了第一个具有多项式复杂度O(n2 L)的不可行算法,然而,这个算法的复杂度较高.为了降低复杂度,众多学者做了大量研究工作[5-10].目前,不可行内点算法的最低复杂度为O(nL),这一结果仅对窄邻域不可行内点算法成立,并且窄邻域不可行内点算法的一般复杂度为O(n1.5L).对宽邻域而言一般算法的复杂度为O(n2L),目前最低复杂度为O(n1.5L).

    本文将给出一个复杂度为O(n1.5L)的宽邻域不可行内点算法,该算法迭代在Ai[11]等提出的宽邻域内.最后,通过数值试验说明该算法是有效的.

    本文中:e表示分量全为1的列向量;‖·‖表示向量的2 -范数.对于向量xs$ {{\mathbb{R}}^\mathit{n}} $xs$ {{\mathbb{R}}^\mathit{n}} $表示对应分量的乘,(xs)min表示xs$ {{\mathbb{R}}^\mathit{n}} $的最小分量;让$ \mathit{\boldsymbol{p}} = {\mathit{\boldsymbol{x}}^{ - \frac{1}{2}}}{\mathit{\boldsymbol{s}}^{\frac{1}{2}}}, \mathit{\boldsymbol{\tilde r}} = {\left( {\mathit{\boldsymbol{xs}}} \right)^{\frac{1}{2}}}\mathit{\boldsymbol{r}} $;另外,对于h$ {{\mathbb{R}}^\mathit{n}} $,我们记:h+=max{0h}和hhh+.

1.   预备知识
  • 本文将研究标准形式的原始-对偶线性规划:

    其中:A$ {{\mathbb{R}}^{m\times n}} $cxs$ {{\mathbb{R}}^\mathit{n}} $by$ {{\mathbb{R}}^\mathit{m}} $.

    为获得(P)和(D)的近似解,一般将原始-对偶线性规划转化为扰动系统:

    如果(P)和(D)的严格可行集

    并且A行满秩,即rank(A)=m,则(1) 式存在唯一解.原始-对偶内点算法的基本思想是用Newton方法求解(1) 式,逐渐减小μ,并使得迭代点列包含在中心路径的某个邻域内,最终得到满足允许精度的最优解.本文假设F0$ \phi $且rank(A)=m.

    下面给出求解迭代方向(ΔxΔyΔs)的方程如下:

    其中:

    α表示沿方向(ΔxΔyΔs)的步长,则新的迭代为:

    直接计算,可得下列常用结果:

    本文将限制中心路径在宽邻域$ \mathscr{N} $(τβ)内,其中$ \beta \le \frac{1}{2}, \tau \le \frac{1}{4} $是邻域参数并且

    另外,由(5) 式易知:

    最后,本文通过下列方法计算最大步长$ \tilde{\alpha } $

2.   算法框架
  • 在前面准备工作的基础上,给出不可行内点算法的一般框架,如下:

    初始化:$ \beta \le \frac{1}{2}, \tau \le \frac{1}{4}, \left( {{\mathit{\boldsymbol{x}}}^{0}}, {{\mathit{\boldsymbol{y}}}^{0}}, {{\mathit{\boldsymbol{s}}}^{0}} \right)\in \mathscr{N}\left( \tau, \beta \right), {{\mu }^{0}}=\frac{1}{n}{{\left( {{\mathit{\boldsymbol{x}}}^{0}} \right)}^{\rm{T}}}{{\mathit{\boldsymbol{s}}}^{0}} $.置k:=0.

    步骤1  如果(xk)Tskε(x0)Ts0,终止.

    步骤2  通过解方程组(2) 获得方向(ΔxΔyΔs),并通过(7) 式计算步长$ {{\tilde{\alpha }}^{k}} $.令

    步骤3  计算$ {\mu ^{k + 1}} = \frac{1}{n}{\left( {{\mathit{\boldsymbol{x}}^{k + 1}}} \right)^{\rm{T}}}{\mathit{\boldsymbol{s}}^{k + 1}} $,并置k:=k+1,转到步骤1.

    给出下面关于本文算法的两点说明对分析算法的复杂度至关重要.

    注1  若{(xkyksk)}是上述算法产生的迭代序列,则对于k≥0,有

    其中

    从注1易知:

    这蕴含着νk可以控制在点(xkyksk)处的不可行性.

    另外,由式(6) 可得:

    这能够保证(xk)Tsk小于收敛精度时,不可行也可以小于相同的精度,等价于:当$ \frac{{{{\left( {{\mathit{\boldsymbol{x}}^k}} \right)}^{\rm{T}}}{\mathit{\boldsymbol{s}}^k}}}{{{{\left( {{\mathit{\boldsymbol{x}}^0}} \right)}^{\rm{T}}}{\mathit{\boldsymbol{s}}^0}}} \le \varepsilon $时,必有

    注2  本文将选取与文献[12]中相同的初始点.

3.   复杂性分析
  • 引理1[11]  若(xys)∈$ \mathscr{N} $(τβ),则

    引理2[13]  若(xys)∈$ \mathscr{N} $(τβ),则

    其中

    由于本文使用了与文献[12]相同的初始点,故使用同样的证明方法可得下面引理.

    引理3[12]  若$ \left( {\mathit{\boldsymbol{x}}, \mathit{\boldsymbol{y}}, \mathit{\boldsymbol{s}}} \right) \in \mathscr{N}\left( {\tau, \beta } \right), \mathit{\boldsymbol{p}} = {\left( {{\mathit{\boldsymbol{x}}^{ - 1}}\mathit{\boldsymbol{s}}} \right)^{\frac{1}{2}}} $,则

    引理4  若$ \left( {\mathit{\boldsymbol{x}}, \mathit{\boldsymbol{y}}, \mathit{\boldsymbol{s}}} \right) \in \mathscr{N}\left( {\tau, \beta } \right), \mathit{\boldsymbol{p}} = {\left( {{\mathit{\boldsymbol{x}}^{ - 1}}\mathit{\boldsymbol{s}}} \right)^{\frac{1}{2}}} $,则

      在方程(2) 的第三个等式两边同时乘以$ {\left( {\mathit{\boldsymbol{xs}}} \right)^{ - \frac{1}{2}}} $,可得:

    进一步,可得:

    其中第二个不等式使用了引理2.

    另外,使用引理1和引理4容易获得下面的引理.

    引理5  若(xys)∈$ \mathscr{N} $(τβ),则μ(α)在区间[0, 1]上是单调递减的.

      对(4) 式两边关于α求导可得:

    这蕴含着μ(α)在区间[0, 1]上是单调递减的.

    引理6  若(xys)∈$ \mathscr{N} $(τβ),则

    其中

      使用(4) 式及引理3,可得:

    另外,使用(4) 式及引理1、引理3可得:

    证毕.

    下面,将寻找满足条件(6) 的αf的下界.

    引理7  若(xys)∈$ \mathscr{N} $(τβ)满足条件(6),则

      要证明该引理,只需要证明在区间$ \left[{0, {{\tilde \alpha }^0}} \right] $上,x(α)Ts(α)≥(1-α)xTs成立.为此,首先证明

    其中最后一个不等式使用了引理3.

    使用上面的不等式,可得:

    证毕.

    为了找到最大迭代步长$ \tilde \alpha $的下界,首先需要给出下列引理.

    引理8[13]  若(xys)∈$ \mathscr{N} $(τβ),则:

    如果$ \alpha \ge \frac{1}{{\sqrt n }} $,则

    如果$ \alpha < \frac{1}{{\sqrt n }} $,则

    引理9  若(xys)∈$ \mathscr{N} $(τβ),$ {\tilde \alpha } $满足条件(7),则$ \tilde \alpha \ge {\tilde \alpha ^0} $.

      为了证明$ \left\| {\;{{\left( {\tau \mu \left( \alpha \right)\mathit{\boldsymbol{e}} - \mathit{\boldsymbol{x}}\left( \alpha \right)\mathit{\boldsymbol{s}}\left( \alpha \right)} \right)}^ + }\;} \right\|\; \le \beta \tau \mu \left( \alpha \right) $,首先,需要对不等式的左边进行放大,可得:

    其中第一个不等式使用了文献[11]的注解3.1,第二、第三个不等式分别使用了引理8、引理3.

    构造一个函数f(α),将该引理的证明转化为证明f(α)≤0即可,其中

    使用引理6以及$ \beta \le \frac{1}{2}, \tau \le \frac{1}{4}, \omega \ge 6 $,可得:

    引理证毕.

    下面给出算法的多项式复杂度.

    定理1  让$ \left( {{\mathit{\boldsymbol{x}}^0}, {\mathit{\boldsymbol{y}}^0}, {\mathit{\boldsymbol{s}}^0}} \right) \in {\mathscr{N}}\left( {\tau, \beta } \right), \tau \le \frac{1}{4}, \beta \le \frac{1}{2} $,则本文的算法至多需要O(n1.5L)次迭代停止,其中

      由算法的终止条件步骤1及引理6,可得:

    其中第三个不等式使用了引理5,

    又由于0=(x0)Ts0,log(1-α)≤-α,容易证明

    另外,当$ k \ge \frac{{{n^{1.5}}}}{\eta }\log \left( {\frac{{{{\left( {{\mathit{\boldsymbol{x}}^0}} \right)}^{\rm{T}}}{\mathit{\boldsymbol{s}}^0}}}{\varepsilon }} \right) $时,使用(8) 式可得:

    所以,算法至多需要O(n1.5L)次迭代停止.

4.   数值试验
  • 用本文算法(算法1) 和文献[6](算法2) 对文献[14]中给出的线性规划进行测试.用MATLAB R2011b编程并在Intel Core i3 PC (3.00GHz),2GB内存下运行.最终的测试结果见表 1.从表 1中容易发现:本文提出的算法在迭代次数和时间上均要优于文献[6]中提出的算法,尤其是对于大规模问题,这种优势更明显.

Table (1) Reference (14)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return