建立smo.m

% function [alpha,bias] = smo(X, y, C, tol)

function model = smo(X, y, C, tol)

% SMO: SMO algorithm for SVM
%
%Implementation of the Sequential Minimal Optimization (SMO)
%training algorithm for Vapnik's Support Vector Machine (SVM)
%
% This is a modified code from Gavin Cawley's MATLAB Support
% Vector Machine Toolbox
% (c) September 2000.
%
% Diego Andres Alvarez.
%
% USAGE: [alpha,bias] = smo(K, y, C, tol)
%
% INPUT:
%
% K: n x n kernel matrix
% y: 1 x n vector of labels, -1 or 1
% C: a regularization parameter such that 0 <= alpha_i <= C/n
% tol: tolerance for terminating criterion
%
% OUTPUT:
%
% alpha: 1 x n lagrange multiplier coefficient
% bias: scalar bias (offset) term % Input/output arguments modified by JooSeuk Kim and Clayton Scott, 2007 global SMO; y = y'; ntp = size(X,1);
%recompute C
% C = C/ntp; %initialize
ii0 = find(y == -1);
ii1 = find(y == 1); i0 = ii0(1);
i1 = ii1(1); alpha_init = zeros(ntp, 1);
alpha_init(i0) = C;
alpha_init(i1) = C;
bias_init = C*(X(i0,:)*X(i1,:)' -X(i0,:)*X(i1,:)') + 1; %Inicializando las variables
SMO.epsilon = 10^(-6); SMO.tolerance = tol;
SMO.y = y'; SMO.C = C;
SMO.alpha = alpha_init; SMO.bias = bias_init;
SMO.ntp = ntp; %number of training points %CACHES:
SMO.Kcache = X*X'; %kernel evaluations
SMO.error = zeros(SMO.ntp,1); %error numChanged = 0; examineAll = 1; %When all data were examined and no changes done the loop reachs its
%end. Otherwise, loops with all data and likely support vector are
%alternated until all support vector be found.
while ((numChanged > 0) || examineAll)
numChanged = 0;
if examineAll
%Loop sobre todos los puntos
for i = 1:ntp
numChanged = numChanged + examineExample(i);
end;
else
%Loop sobre KKT points
for i = 1:ntp
%Solo los puntos que violan las condiciones KKT
if (SMO.alpha(i)>SMO.epsilon) && (SMO.alpha(i)<(SMO.C-SMO.epsilon))
numChanged = numChanged + examineExample(i);
end;
end;
end; if (examineAll == 1)
examineAll = 0;
elseif (numChanged == 0)
examineAll = 1;
end;
end;
alpha = SMO.alpha';
alpha(alpha < SMO.epsilon) = 0;
alpha(alpha > C-SMO.epsilon) = C;
bias = -SMO.bias; model.w = (y.*alpha)* X; %%%%%%%%%%%%%%%%%%%%%%
model.b = bias; return; function RESULT = fwd(n)
global SMO;
LN = length(n);
RESULT = -SMO.bias + sum(repmat(SMO.y,1,LN) .* repmat(SMO.alpha,1,LN) .* SMO.Kcache(:,n))';
return; function RESULT = examineExample(i2)
%First heuristic selects i2 and asks to examineExample to find a
%second point (i1) in order to do an optimization step with two
%Lagrange multipliers global SMO;
alpha2 = SMO.alpha(i2); y2 = SMO.y(i2); if ((alpha2 > SMO.epsilon) && (alpha2 < (SMO.C-SMO.epsilon)))
e2 = SMO.error(i2);
else
e2 = fwd(i2) - y2;
end; % r2 < 0 if point i2 is placed between margin (-1)-(+1)
% Otherwise r2 is > 0. r2 = f2*y2-1 r2 = e2*y2;
%KKT conditions:
% r2>0 and alpha2==0 (well classified)
% r2==0 and 0% r2<0 and alpha2==C (support vectors between margins)
%
% Test the KKT conditions for the current i2 point.
%
% If a point is well classified its alpha must be 0 or if
% it is out of its margin its alpha must be C. If it is at margin
% its alpha must be between 0%take action only if i2 violates Karush-Kuhn-Tucker conditions
if ((r2 < -SMO.tolerance) && (alpha2 < (SMO.C-SMO.epsilon))) || ...
((r2 > SMO.tolerance) && (alpha2 > SMO.epsilon))
% If it doens't violate KKT conditions then exit, otherwise continue. %Try i2 by three ways; if successful, then immediately return 1;
RESULT = 1;
% First the routine tries to find an i1 lagrange multiplier that
% maximizes the measure |E1-E2|. As large this value is as bigger
% the dual objective function becames.
% In this first test, only support vectors will be tested. POS = find((SMO.alpha > SMO.epsilon) & (SMO.alpha < (SMO.C-SMO.epsilon)));
[MAX,i1] = max(abs(e2 - SMO.error(POS)));
if ~isempty(i1)
if takeStep(i1, i2, e2), return;
end;
end; %The second heuristic choose any Lagrange Multiplier that is a SV and tries to optimize
for i1 = randperm(SMO.ntp)
if (SMO.alpha(i1) > SMO.epsilon) & (SMO.alpha(i1) < (SMO.C-SMO.epsilon))
%if a good i1 is found, optimise
if takeStep(i1, i2, e2), return;
end;
end
end %if both heuristc above fail, iterate over all data set
for i1 = randperm(SMO.ntp)
if ~((SMO.alpha(i1) > SMO.epsilon) & (SMO.alpha(i1) < (SMO.C-SMO.epsilon)))
if takeStep(i1, i2, e2), return;
end;
end
end;
end; %no progress possible
RESULT = 0;
return; function RESULT = takeStep(i1, i2, e2)
% for a pair of alpha indexes, verify if it is possible to execute
% the optimisation described by Platt. global SMO;
RESULT = 0;
if (i1 == i2), return;
end; % compute upper and lower constraints, L and H, on multiplier a2
alpha1 = SMO.alpha(i1); alpha2 = SMO.alpha(i2);
y1 = SMO.y(i1); y2 = SMO.y(i2);
C = SMO.C; K = SMO.Kcache; s = y1*y2;
if (y1 ~= y2)
L = max(0, alpha2-alpha1); H = min(C, alpha2-alpha1+C);
else
L = max(0, alpha1+alpha2-C); H = min(C, alpha1+alpha2);
end; if (L == H), return;
end; if (alpha1 > SMO.epsilon) & (alpha1 < (C-SMO.epsilon))
e1 = SMO.error(i1);
else
e1 = fwd(i1) - y1;
end; %if (alpha2 > SMO.epsilon) & (alpha2 < (C-SMO.epsilon))
% e2 = SMO.error(i2);
%else
% e2 = fwd(i2) - y2;
%end; %compute eta
k11 = K(i1,i1); k12 = K(i1,i2); k22 = K(i2,i2);
eta = 2.0*k12-k11-k22; %recompute Lagrange multiplier for pattern i2
if (eta < 0.0)
a2 = alpha2 - y2*(e1 - e2)/eta; %constrain a2 to lie between L and H
if (a2 < L)
a2 = L;
elseif (a2 > H)
a2 = H;
end;
else
%When eta is not negative, the objective function W should be
%evaluated at each end of the line segment. Only those terms in the
%objective function that depend on alpha2 need be evaluated... ind = find(SMO.alpha>0); aa2 = L; aa1 = alpha1 + s*(alpha2-aa2); Lobj = aa1 + aa2 + sum((-y1*aa1/2).*SMO.y(ind).*K(ind,i1) + (-y2*aa2/2).*SMO.y(ind).*K(ind,i2)); aa2 = H; aa1 = alpha1 + s*(alpha2-aa2);
Hobj = aa1 + aa2 + sum((-y1*aa1/2).*SMO.y(ind).*K(ind,i1) + (-y2*aa2/2).*SMO.y(ind).*K(ind,i2)); if (Lobj>Hobj+SMO.epsilon)
a2 = H;
elseif (Lobj<Hobj-SMO.epsilon)
a2 = L;
else
a2 = alpha2;
end;
end; if (abs(a2-alpha2) < SMO.epsilon*(a2+alpha2+SMO.epsilon))
return;
end; % recompute Lagrange multiplier for pattern i1
a1 = alpha1 + s*(alpha2-a2); w1 = y1*(a1 - alpha1); w2 = y2*(a2 - alpha2); %update threshold to reflect change in Lagrange multipliers
b1 = SMO.bias + e1 + w1*k11 + w2*k12;
bold = SMO.bias; if (a1>SMO.epsilon) & (a1<(C-SMO.epsilon))
SMO.bias = b1;
else
b2 = SMO.bias + e2 + w1*k12 + w2*k22;
if (a2>SMO.epsilon) & (a2<(C-SMO.epsilon))
SMO.bias = b2;
else
SMO.bias = (b1 + b2)/2;
end;
end; % update error cache using new Lagrange multipliers
SMO.error = SMO.error + w1*K(:,i1) + w2*K(:,i2) + bold - SMO.bias;
SMO.error(i1) = 0.0; SMO.error(i2) = 0.0; % update vector of Lagrange multipliers
SMO.alpha(i1) = a1; SMO.alpha(i2) = a2; %report progress made
RESULT = 1;
return;

画图文件:start_SMOforSVM.m(点击自动生成二维两类数据,画图,这里只是线性的,非线性的可以对应修改)

clear
X = []; Y=[];
figure;
% Initialize training data to empty; will get points from user
% Obtain points froom the user:
trainPoints=X;
trainLabels=Y;
clf;
axis([-5 5 -5 5]);
if isempty(trainPoints)
% Define the symbols and colors we'll use in the plots later
symbols = {'o','x'};
classvals = [-1 1];
trainLabels=[];
hold on; % Allow for overwriting existing plots
xlim([-5 5]); ylim([-5 5]); for c = 1:2
title(sprintf('Click to create points from class %d. Press enter when finished.', c));
[x, y] = getpts; plot(x,y,symbols{c},'LineWidth', 2, 'Color', 'black'); % Grow the data and label matrices
trainPoints = vertcat(trainPoints, [x y]);
trainLabels = vertcat(trainLabels, repmat(classvals(c), numel(x), 1));
end end % C = 10;tol = 0.001;
% par = SMOforSVM(trainPoints, trainLabels , C, tol );
% p=length(par.b); m=size(trainPoints,2);
% if m==2
% % for i=1:p
% % plot(X(lc(i)-l(i)+1:lc(i),1),X(lc(i)-l(i)+1:lc(i),2),'bo')
% % hold on
% % end
% k = -par.w(1)/par.w(2);
% b0 = - par.b/par.w(2);
% bdown=(-par.b-1)/par.w(2);
% bup=(-par.b+1)/par.w(2);
% for i=1:p
% hold on
% h = refline(k,b0(i));
% set(h, 'Color', 'r')
% hdown=refline(k,bdown(i));
% set(hdown, 'Color', 'b')
% hup=refline(k,bup(i));
% set(hup, 'Color', 'b')
% end
% end
% xlim([-5 5]); ylim([-5 5]);
%
% pause C = 10;tol = 0.001;
par = smo(trainPoints, trainLabels, C, tol);
p=length(par.b); m=size(trainPoints,2);
if m==2
% for i=1:p
% plot(X(lc(i)-l(i)+1:lc(i),1),X(lc(i)-l(i)+1:lc(i),2),'bo')
% hold on
% end
k = -par.w(1)/par.w(2);
b0 = - par.b/par.w(2);
bdown=(-par.b-1)/par.w(2);
bup=(-par.b+1)/par.w(2);
for i=1:p
hold on
h = refline(k,b0(i));
set(h, 'Color', 'r')
hdown=refline(k,bdown(i));
set(hdown, 'Color', 'b')
hup=refline(k,bup(i));
set(hup, 'Color', 'b')
end
end
xlim([-5 5]); ylim([-5 5]);

  

支持向量机的smo算法(MATLAB code)的更多相关文章

  1. SVM-非线性支持向量机及SMO算法

    SVM-非线性支持向量机及SMO算法 如果您想体验更好的阅读:请戳这里littlefish.top 线性不可分情况 线性可分问题的支持向量机学习方法,对线性不可分训练数据是不适用的,为了满足函数间隔大 ...

  2. smo算法matlab实现

    看完CSDN上结构之法,算法之道的支持向量机通俗导论(理解SVM的三层境界) http://blog.csdn.net/v_july_v/article/details/7624837     参考了 ...

  3. 机器学习算法整理(七)支持向量机以及SMO算法实现

    以下均为自己看视频做的笔记,自用,侵删! 还参考了:http://www.ai-start.com/ml2014/ 在监督学习中,许多学习算法的性能都非常类似,因此,重要的不是你该选择使用学习算法A还 ...

  4. 支持向量机原理(四)SMO算法原理

    支持向量机原理(一) 线性支持向量机 支持向量机原理(二) 线性支持向量机的软间隔最大化模型 支持向量机原理(三)线性不可分支持向量机与核函数 支持向量机原理(四)SMO算法原理 支持向量机原理(五) ...

  5. [笔记]关于支持向量机(SVM)中 SMO算法的学习(一)理论总结

    1. 前言 最近又重新复习了一遍支持向量机(SVM).其实个人感觉SVM整体可以分成三个部分: 1. SVM理论本身:包括最大间隔超平面(Maximum Margin Classifier),拉格朗日 ...

  6. 支持向量机(Support Vector Machine)-----SVM之SMO算法(转)

    此文转自两篇博文 有修改 序列最小优化算法(英语:Sequential minimal optimization, SMO)是一种用于解决支持向量机训练过程中所产生优化问题的算法.SMO由微软研究院的 ...

  7. 机器学习之支持向量机(二):SMO算法

    注:关于支持向量机系列文章是借鉴大神的神作,加以自己的理解写成的:若对原作者有损请告知,我会及时处理.转载请标明来源. 序: 我在支持向量机系列中主要讲支持向量机的公式推导,第一部分讲到推出拉格朗日对 ...

  8. 统计学习方法c++实现之六 支持向量机(SVM)及SMO算法

    前言 支持向量机(SVM)是一种很重要的机器学习分类算法,本身是一种线性分类算法,但是由于加入了核技巧,使得SVM也可以进行非线性数据的分类:SVM本来是一种二分类分类器,但是可以扩展到多分类,本篇不 ...

  9. 支持向量机(SVM)中的 SMO算法

    1. 前言 最近又重新复习了一遍支持向量机(SVM).其实个人感觉SVM整体可以分成三个部分: 1. SVM理论本身:包括最大间隔超平面(Maximum Margin Classifier),拉格朗日 ...

随机推荐

  1. 5.4.1 termios结构,关闭回显功能,一键入字符fgetc立刻返回,不用按下回车键

    Linux提供了一组编程接口,用来控制终端驱动程序的行为.这样我们可以更精细的来控制终端. 例如: 回显:允许控制字符的回显,例如读取密码时. 使用termios结构的密码程序 #include &l ...

  2. log4j的使用(1) —— 简单入门篇

    这里会介绍三种打印日志的方法:控制台Console,文件File,数据库DataBase 1.下载lo4j的jar包并导入project 2.因为要在数据库添加日志,所以先新建一个库,并新建打印日志的 ...

  3. STL区间成员函数及区间算法总结

    STL区间成员函数及区间算法总结 在这里总结下可替代循环的区间成员函数和区间算法: 相比单元素遍历操作,使用区间成员函数的优势在于: 1)更少的函数调用 2)更少的元素移动 3)更少的内存分配 在区间 ...

  4. lazyload懒加载的使用

    1.引用<script src="http://a.tbcdn.cn/apps/baron/js/??lib/tmm/tmm.js,lib/lazyload/lazyload.js?2 ...

  5. 个人博客Week3

    必应软件客户端测评 话不多说 先来一张高大上的客户端界面截图! 首先 映入眼帘的就是四个主要的模块!(1)词典(2)例句(3)翻译(4)应用 (1)词典: 在查询词条的时候,一旦查询了一个词条就不能够 ...

  6. iOS - Swift available 平台判断

    前言 Swift 语言中的 @available 和 #available,Swift 2.0 中,引入了可用性的概念.对于函数,类,协议等,可以使用 @available 声明这些类型的生命周期依赖 ...

  7. RFC总结-SD模块

    1.客户主数据1.1 创建.修改客户主数据*该函数更新所有字段,X表用来新增和修改,Y表为删除  CALL FUNCTION 'SD_CUSTOMER_MAINTAIN_ALL'   EXPORTIN ...

  8. JavaScript基础知识点

    本书目录 第一章:  JavaScript语言基础 第二章:  JavaScript内置对象第三章:  窗口window对象第四章:  文档document对象第五章:  表单form对象第六章:   ...

  9. Android热修复

    https://github.com/WeMobileDev/article/blob/master/%E5%BE%AE%E4%BF%A1Android%E7%83%AD%E8%A1%A5%E4%B8 ...

  10. Android APK混淆

    APK混淆 1 修改project.properties文件 即可实现对项目进行全局混码将proguard.config=${sdk.dir}/tools/proguard/proguard-andr ...