SVM的参数优化——遗传算法参数寻优

支持向量机（Support Vector Machine，SVM）是一种广泛应用的监督式机器学习算法。它主要用于分类任务，但也适用于回归任务。在本文中，我们将深入探讨支持向量机的两个重要参数：C和gamma。

%% 子函数 gaSVMcgForClass.m

function [BestCVaccuracy,Bestc,Bestg,ga_option] = gaSVMcgForClass(train_label,train_data,ga_option)
% gaSVMcgForClass
%
% by faruto
%Email:patrick.lee@foxmail.com QQ:516667408 http://blog.sina.com.cn/faruto BNU
%last modified 2010.01.17
%Super Moderator @ www.ilovematlab.cn

% 若转载请注明：
% faruto and liyang , LIBSVM-farutoUltimateVersion 
% a toolbox with implements for support vector machines based on libsvm, 2009. 
% Software available at http://www.ilovematlab.cn
% 
% Chih-Chung Chang and Chih-Jen Lin, LIBSVM : a library for
% support vector machines, 2001. Software available at
% http://www.csie.ntu.edu.tw/~cjlin/libsvm

% 参数初始化
if nargin == 2
    ga_option = struct('maxgen',200,'sizepop',20,'ggap',0.9,...
        'cbound',[0,100],'gbound',[0,1000],'v',5);
end
% maxgen:最大的进化代数,默认为200,一般取值范围为[100,500]
% sizepop:种群最大数量,默认为20,一般取值范围为[20,100]
% cbound = [cmin,cmax],参数c的变化范围,默认为(0,100]
% gbound = [gmin,gmax],参数g的变化范围,默认为[0,1000]
% v:SVM Cross Validation参数,默认为5

%
MAXGEN = ga_option.maxgen;
NIND = ga_option.sizepop;
NVAR = 2;
PRECI = 20;
GGAP = ga_option.ggap;
trace = zeros(MAXGEN,2);

FieldID = ...
[repmat([PRECI],[1,NVAR]);[ga_option.cbound(1),ga_option.gbound(1);ga_option.cbound(2),ga_option.gbound(2)]; ...
 [1,1;0,0;0,1;1,1]];

Chrom = crtbp(NIND,NVAR*PRECI);

gen = 1;
v = ga_option.v;
BestCVaccuracy = 0;
Bestc = 0;
Bestg = 0;
%
cg = bs2rv(Chrom,FieldID);

for nind = 1:NIND
    cmd = ['-v ',num2str(v),' -c ',num2str(cg(nind,1)),' -g ',num2str(cg(nind,2))];
    ObjV(nind,1) = svmtrain(train_label,train_data,cmd);
end
[BestCVaccuracy,I] = max(ObjV);
Bestc = cg(I,1);
Bestg = cg(I,2);

for gen = 1:MAXGEN
    FitnV = ranking(-ObjV);
    
    SelCh = select('sus',Chrom,FitnV,GGAP);
    SelCh = recombin('xovsp',SelCh,0.7);
    SelCh = mut(SelCh);
    
    cg = bs2rv(SelCh,FieldID);
    for nind = 1:size(SelCh,1)
        cmd = ['-v ',num2str(v),' -c ',num2str(cg(nind,1)),' -g ',num2str(cg(nind,2))];
        ObjVSel(nind,1) = svmtrain(train_label,train_data,cmd);
    end
    
    [Chrom,ObjV] = reins(Chrom,SelCh,1,1,ObjV,ObjVSel);
    
    if max(ObjV) <= 50
        continue;
    end
    
    [NewBestCVaccuracy,I] = max(ObjV);
    cg_temp = bs2rv(Chrom,FieldID);
    temp_NewBestCVaccuracy = NewBestCVaccuracy;
    
    if NewBestCVaccuracy > BestCVaccuracy
       BestCVaccuracy = NewBestCVaccuracy;
       Bestc = cg_temp(I,1);
       Bestg = cg_temp(I,2);
    end
    
    if abs( NewBestCVaccuracy-BestCVaccuracy ) <= 10^(-2) && ...
        cg_temp(I,1) < Bestc
       BestCVaccuracy = NewBestCVaccuracy;
       Bestc = cg_temp(I,1);
       Bestg = cg_temp(I,2);
    end    
    
    trace(gen,1) = max(ObjV);
    trace(gen,2) = sum(ObjV)/length(ObjV);
  
end
%
figure;
hold on;
trace = round(trace*10000)/10000;
plot(trace(1:gen,1),'r*-','LineWidth',1.5);
plot(trace(1:gen,2),'o-','LineWidth',1.5);
legend('最佳适应度','平均适应度','Location','SouthWest');
xlabel('进化代数','FontSize',12);
ylabel('适应度','FontSize',12);
axis([0 gen 0 100]);
grid on;
axis auto;

line1 = '适应度曲线Accuracy[GAmethod]';
line2 = ['(终止代数=', ...
    num2str(gen),',种群数量pop=', ...
    num2str(NIND),')'];
line3 = ['Best c=',num2str(Bestc),' g=',num2str(Bestg), ...
    ' CVAccuracy=',num2str(BestCVaccuracy),'%'];
title({line1;line2;line3},'FontSize',12);

%% 选择GA最佳的SVM参数c&g

% GA的参数选项初始化

ga_option.maxgen = 200;
ga_option.sizepop = 20; 
ga_option.cbound = [0,100];
ga_option.gbound = [0,100];
ga_option.v = 5;
ga_option.ggap = 0.9;
[bestacc,bestc,bestg] = gaSVMcgForClass(train_wine_labels,train_wine,ga_option);

% 打印选择结果
disp('打印选择结果');

str = sprintf( 'Best Cross Validation Accuracy = %g%% Best c = %g Best g = %g',bestacc,bestc,bestg);
disp(str);

%% 利用最佳的参数进行SVM网络训练

cmd = ['-c ',num2str(bestc),' -g ',num2str(bestg)];

 源自：《Matlab神经网络43个案例分析》。