huigui-svm.m

利用PSO优化SVM

%%####################################################################
%%#### Particle swarm optimization
%%#### With linkage operator
%%#### Deepak devicharan july 2003
%%####################################################################

%%## to apply this to different equations do the following
%%## generate initial particles in a search space close to actual soln
%%## fool around with no of iterations, no of particles, learning rates

%%## for a truly generic PSO do the following
%%## increase the number of particles , increase the variance
%%## i.e let the particles cover a larger area of the search space
%%## then fool around as always with the above thins


%declare the parameters of the optimization
clc
clear
close all
tic

cancer_input = [-0.927 -0.919 -0.919 -0.968 -1.013 
    -0.919 -0.919 -0.968 -1.013 -1.035 
    -0.919 -0.968 -1.013 -1.035 -0.985 
    -0.968 -1.013 -1.035 -0.985 -1.003 
    -1.013 -1.035 -0.985 -1.003 -0.998 
    -1.035 -0.985 -1.003 -0.998 -1.007 
    -0.985 -1.003 -0.998 -1.007 -1.000 
    -1.003 -0.998 -1.007 -1.000 -0.992 
    -0.998 -1.007 -1.000 -0.992 -0.988 
    -1.007 -1.000 -0.992 -0.988 0.870 
    -1.000 -0.992 -0.988 -0.870 -0.874 
    -0.992 -0.988 -0.870 -0.874 -0.879 
    -0.988 -0.870 -0.874 -0.879 -0.873
    -0.870 -0.874 -0.879 -0.873 -0.929
    -0.874 -0.879 -0.873 -0.929 -0.922 
    -0.879 -0.873 -0.929 -0.922 -0.869 
    -0.873 -0.929 -0.922 -0.869 -0.958
    -0.929 -0.922 -0.869 -0.958 -0.899 
    -0.922 -0.869 -0.958 -0.899 -0.893 
    -0.869 -0.958 -0.899 -0.893 -0.944 
    -0.958 -0.899 -0.893 -0.944 -0.936
    -0.899 -0.893 -0.944 -0.936 -0.934 
    -0.893 -0.944 -0.936 -0.934 -0.887 
    -0.944 -0.936 -0.934 -0.887 -0.943 
    -0.936 -0.934 -0.887 -0.943 -0.940 
    -0.934 -0.887 -0.943 -0.940 -0.939
    -0.887 -0.943 -0.940 -0.939 -0.940 
    -0.943 -0.940 -0.939 -0.940 -0.937 
];
cancer_output = [-1.035 
-0.985 
-1.003 
-0.998 
-1.007 
-1.000 
-0.992 
-0.988 
-0.870 
-0.874 
-0.879 
-0.873
-0.929
-0.922 
-0.869 
-0.958 
-0.899 
-0.893 
-0.944 
-0.936
-0.934 
-0.887 
-0.943 
-0.940 
-0.939
-0.940 
-0.937 
-0.954
];

max_iterations  = 20;
no_of_particles = 20;
dimensions      = 2;

%delta_min       = -0.003;
%delta_max       = 0.003;

c1 = 1.3;
c2 = 1.3;

%initialise the particles and teir velocity components

for count_x = 1:no_of_particles
    for count_y = 1:dimensions
        particle_position(count_x,count_y) = rand*10;
        particle_velocity(count_x,count_y) = rand*1000;
        p_best(count_x,count_y) = particle_position(count_x,count_y);
    end
end

%initialize the p_best_fitness array
for count = 1:no_of_particles
    p_best_fitness(count) = -1000;
end


%particle_position
%particle_velocity

%main particle swrm routine
for count = 1:max_iterations
    
    %find the fitness of each particle
    %change fitness function as per equation requiresd and dimensions
    for count_x = 1:no_of_particles
        x = particle_position(count_x,1);
        y = particle_position(count_x,2);
%---------------------------------------------------
% 产生训练样本与测试样本
% 特别注意：此工具箱用于分类时，只能处理2类分类，且目标值必须为 1 或 -1。

        %n1 = [rand(3,5),rand(3,5)+1];
        %x1 = [1*ones(1,5),-1*ones(1,5)];
        for i = 1:20
            for j = 1:5
                n1(j,i) = cancer_input(i,j);        
            end
        end
        for i = 1:20
            x1(i) = cancer_output(i);
        end
    
        %n2 = [rand(3,5),rand(3,5)+1];
        %x2 = [1*ones(1,5),-1*ones(1,5)];
        for i = 1:8
            for j = 1:5
                n2(j,i) = cancer_input(20+i,j);        
            end
        end
        for i = 1:8
            x2(i) = cancer_output(20+i);
        end

        xn_train = n1;          % 训练样本，每一列为一个样本
        dn_train = x1;          % 训练目标，行向量

        xn_test = n2;           % 测试样本，每一列为一个样本
        dn_test = x2;           % 测试目标，行向量
        

%---------------------------------------------------
% 参数设置

        trnX = xn_train';
        trnY = dn_train';
        tstX = xn_test';
        tstY = dn_test';

        ker = 'rbf';        % 核函数 k = exp(-(u-v)*(u-v)'/(2*p1^2))
        global p1 ;
        p1 = x;             % p1 is width of rbfs (sigma)
        C = y;             % 折衷系数

%---------------------------------------------------
% 训练与测试

        %[nsv,alpha,bias] = svc(trnX,trnY,ker,C);                        % 训练

        %actfunc = 0;                                                    % 1 为实际输出，0 为取sign输出 
        %predictedY = svcoutput(trnX,trnY,tstX,ker,alpha,bias,actfunc);  % 测试
[nsv,beta,bias] = svr(trnX,trnY,ker,C);         % 训练
tstY1 = svroutput(trnX,tstX,ker,beta,bias);     % 测试

tst = tstY - tstY1;
sum = 0;
for i = 1:100
    sum = tst(i)*tst(i) + sum;
end
soln = (sum/100)^(1/2);
%---------------------------------------------------
% 结果统计

        %Result = ~abs(predictedY-tstY)               % 正确分类显示为1
        %Percent = sum(Result)/length(Result)   % 正确分类率
        Percent = 1 - soln;
   
        
        %x = particle_position(count_x,1);
        %y = particle_position(count_x,2);
        %z = particle_position(count_x,3);
        %soln = x^2 - 3*y*x + z;
        
        %x = particle_position(count_x); 
        %soln = x^2-2*x+1;
        
       % x = particle_position(count_x);
   % soln = x-7;
        
        if soln~=0 
            current_fitness(count_x) = 1/abs(soln)+0.0001;
        else
            current_fitness(count_x) =1000;
        end
    end
    
    %decide on p_best etc for each particle
    for count_x = 1:no_of_particles
        if current_fitness(count_x) > p_best_fitness(count_x)
            p_best_fitness(count_x) = current_fitness(count_x);
            for count_y = 1:dimensions
                p_best(count_x,count_y) = particle_position(count_x,count_y);
            end
        end
    end
        
    %decide on the global best among all the particles
    [g_best_val,g_best_index] = max(current_fitness);
    
    %g_best contains the position of teh global best
    for count_y = 1:dimensions
        g_best(count_y) = particle_position(g_best_index,count_y);      
    end

    %update the position and velocity compponents
    for count_x = 1:no_of_particles
        for count_y = 1:dimensions
            p_current(count_y) = particle_position(count_x,count_y);
        end

        for count_y = 1:dimensions
            particle_velocity(count_y) = particle_velocity(count_y) +  c1*rand*(p_best(count_y)-p_current(count_y)) + c2*rand*(g_best(count_y)-p_current(count_y));
            particle_positon(count_x,count_y) = p_current(count_y) +particle_velocity(count_y);
        end
    end
                                 
              
end

g_best        
current_fitness(g_best_index)

t=toc;