SVM的参数优化——粒子群算法参数寻优

粒子群优化算法的基本思想：是通过群体中个体之间的协作和信息共享来寻找最优解。PSO的优势：在于简单容易实现并且没有许多参数的调节。目前已被广泛应用于函数优化、神经网络训练、模糊系统控制以及其他遗传算法的应用领域。

%% 子函数 psoSVMcgForClass.m

function [bestCVaccuarcy,bestc,bestg,pso_option] = psoSVMcgForClass(train_label,train,pso_option)
% psoSVMcgForClass
%
% 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
    pso_option = struct('c1',1.5,'c2',1.7,'maxgen',200,'sizepop',20, ...
        'k',0.6,'wV',1,'wP',1,'v',5, ...
        'popcmax',10^2,'popcmin',10^(-1),'popgmax',10^3,'popgmin',10^(-2));
end
% c1:初始为1.5,pso参数局部搜索能力
% c2:初始为1.7,pso参数全局搜索能力
% maxgen:初始为200,最大进化数量
% sizepop:初始为20,种群最大数量
% k:初始为0.6(k belongs to [0.1,1.0]),速率和x的关系(V = kX)
% wV:初始为1(wV best belongs to [0.8,1.2]),速率更新公式中速度前面的弹性系数
% wP:初始为1,种群更新公式中速度前面的弹性系数
% v:初始为3,SVM Cross Validation参数
% popcmax:初始为100,SVM 参数c的变化的最大值.
% popcmin:初始为0.1,SVM 参数c的变化的最小值.
% popgmax:初始为1000,SVM 参数g的变化的最大值.
% popgmin:初始为0.01,SVM 参数c的变化的最小值.

Vcmax = pso_option.k*pso_option.popcmax;
Vcmin = -Vcmax ;
Vgmax = pso_option.k*pso_option.popgmax;
Vgmin = -Vgmax ;

eps = 10^(-3);

% 产生初始粒子和速度
for i=1:pso_option.sizepop
    
    % 随机产生种群和速度
    pop(i,1) = (pso_option.popcmax-pso_option.popcmin)*rand+pso_option.popcmin;
    pop(i,2) = (pso_option.popgmax-pso_option.popgmin)*rand+pso_option.popgmin;
    V(i,1)=Vcmax*rands(1,1);
    V(i,2)=Vgmax*rands(1,1);
    
    % 计算初始适应度
    cmd = ['-v ',num2str(pso_option.v),' -c ',num2str( pop(i,1) ),' -g ',num2str( pop(i,2) )];
    fitness(i) = svmtrain(train_label, train, cmd);
    fitness(i) = -fitness(i);
end

% 找极值和极值点
[global_fitness bestindex]=min(fitness); % 全局极值
local_fitness=fitness;   % 个体极值初始化

global_x=pop(bestindex,:);   % 全局极值点
local_x=pop;    % 个体极值点初始化

% 每一代种群的平均适应度
avgfitness_gen = zeros(1,pso_option.maxgen);

% 迭代寻优
for i=1:pso_option.maxgen
    
    for j=1:pso_option.sizepop
        
        %速度更新
        V(j,:) = pso_option.wV*V(j,:) + pso_option.c1*rand*(local_x(j,:) - pop(j,:)) + pso_option.c2*rand*(global_x - pop(j,:));
        if V(j,1) > Vcmax
            V(j,1) = Vcmax;
        end
        if V(j,1) < Vcmin
            V(j,1) = Vcmin;
        end
        if V(j,2) > Vgmax
            V(j,2) = Vgmax;
        end
        if V(j,2) < Vgmin
            V(j,2) = Vgmin;
        end
        
        %种群更新
        pop(j,:)=pop(j,:) + pso_option.wP*V(j,:);
        if pop(j,1) > pso_option.popcmax
            pop(j,1) = pso_option.popcmax;
        end
        if pop(j,1) < pso_option.popcmin
            pop(j,1) = pso_option.popcmin;
        end
        if pop(j,2) > pso_option.popgmax
            pop(j,2) = pso_option.popgmax;
        end
        if pop(j,2) < pso_option.popgmin
            pop(j,2) = pso_option.popgmin;
        end
        
        % 自适应粒子变异
        if rand>0.5
            k=ceil(2*rand);
            if k == 1
                pop(j,k) = (20-1)*rand+1;
            end
            if k == 2
                pop(j,k) = (pso_option.popgmax-pso_option.popgmin)*rand + pso_option.popgmin;
            end
        end
        
        %适应度值
        cmd = ['-v ',num2str(pso_option.v),' -c ',num2str( pop(j,1) ),' -g ',num2str( pop(j,2) )];
        fitness(j) = svmtrain(train_label, train, cmd);
        fitness(j) = -fitness(j);
        
        cmd_temp = ['-c ',num2str( pop(j,1) ),' -g ',num2str( pop(j,2) )];
        model = svmtrain(train_label, train, cmd_temp);
        
        if fitness(j) >= -65
            continue;
        end
        
        %个体最优更新
        if fitness(j) < local_fitness(j)
            local_x(j,:) = pop(j,:);
            local_fitness(j) = fitness(j);
        end
        
        if abs( fitness(j)-local_fitness(j) )<=eps && pop(j,1) < local_x(j,1)
            local_x(j,:) = pop(j,:);
            local_fitness(j) = fitness(j);
        end
        
        %群体最优更新
        if fitness(j) < global_fitness
            global_x = pop(j,:);
            global_fitness = fitness(j);
        end
        
        if abs( fitness(j)-global_fitness )<=eps && pop(j,1) < global_x(1)
            global_x = pop(j,:);
            global_fitness = fitness(j);
        end
        
    end
    
    fit_gen(i) = global_fitness;
    avgfitness_gen(i) = sum(fitness)/pso_option.sizepop;
end

% 结果分析
figure;
hold on;
plot(-fit_gen,'r*-','LineWidth',1.5);
plot(-avgfitness_gen,'o-','LineWidth',1.5);
%legend('最佳适应度','平均适应度',3);
legend('最佳适应度','平均适应度','Location','SouthWest','Orientation','horizontal');
xlabel('进化代数','FontSize',12);
ylabel('适应度','FontSize',12);
grid on;

bestc = global_x(1);
bestg = global_x(2);
bestCVaccuarcy = -fit_gen(pso_option.maxgen);

line1 = '适应度曲线Accuracy[PSOmethod]';
line2 = ['(参数c1=',num2str(pso_option.c1), ...
    ',c2=',num2str(pso_option.c2),',终止代数=', ...
    num2str(pso_option.maxgen),',种群数量pop=', ...
    num2str(pso_option.sizepop),')'];
line3 = ['Best c=',num2str(bestc),' g=',num2str(bestg), ...
    ' CVAccuracy=',num2str(bestCVaccuarcy),'%'];
title({line1;line2;line3},'FontSize',12);

 

%% 选择最佳的SVM参数c&g

[bestacc,bestc,bestg] = psoSVMcgForClass(train_wine_labels,train_wine);

% 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》。