function [bestindividual] = ga(train, label, x, d, popsize)
% load sonar.mat
% label = [ones(97,1);2*ones(111,1)];
% train = sonar;
%
% % load iris.mat
% % label = [ones(50,1);2*ones(50,1);3*ones(50,1)];
% % train = iris;
% %样本类别数
% x = 2; %sonar上
% % x = 3; %iris上
% %要选出的特征数
% d = 20; %sonar上
% % d = 2; %iris上
% %种群大小
% popsize = 100; %sonar上
% % popsize = 4; %iris上
% %二进制编码长度
chromlength = length(train(1,:));
%交叉概率
pc = 0.5;
%变异概率
pm = 0.05;
%初始种群
temp = [ones(1,d),zeros(1,chromlength - d)];
pop = zeros(popsize,chromlength);
newpop = pop;
for i = 1:popsize
pop(i,:) = temp(randperm(chromlength));
end
%初始化适应度函数值
fitvalue = zeros(popsize,1);
%最优适应度初值
bestf = -inf;
%初始化每代最优适应度函数值
Bestf = zeros(1,200);
%开始迭代
for i = 1:300
%计算适应度值（函数值）
for j = 1:popsize
index = find(pop(j,:)==1);
fitvalue(j) = fitness(train(:,index)',label,x);
end
[fmax,max_index] = max(fitvalue);
p = fitvalue./sum(fitvalue);
q = cumsum(p);
if fmax > bestf
bestf = fmax;
bestindividual = pop(max_index,:);
end
Bestf(i) = bestf;
%选择操作
for k = 1:popsize-1
r = rand;
tmp = find(r<=q);
newpop(k,:) = pop(tmp(1),:);
end
newpop(popsize,:) = bestindividual;
%交叉操作
child = zeros(pc*popsize,chromlength);
t = randperm(popsize);
mather = newpop(t(1:2:pc*popsize),:);
father = newpop(t(2:2:pc*popsize),:);
for m = 1:pc*popsize/2
y = mather(m,:)&father(m,:);
adv = find(y == 1);
a = length(adv);
e = d - a;
disadv = find(mather(m,:) ~= father(m,:));
s = randperm(length(disadv));
child(m,adv) = father(m,adv);
child(m,disadv(s(1:e))) = 1;
child(m+pc*popsize/2,adv) = mather(m,adv);
child(m+pc*popsize/2,disadv(s(e+1:2*e))) = 1;
end
newpop(t(1:pc*popsize),:) = child;
%变异操作
for n = 1:popsize
if rand < pm
inx3 = find(newpop(n,:)==1);
inx4 = find(newpop(n,:)==0);
tmp3 = randperm(length(inx3));
tmp4 = randperm(length(inx4));
newpop(n,inx3(tmp3(1))) = 0;
newpop(n,inx4(tmp4(1))) = 1;
end
end
%更新种群
pop = newpop;
end
% figure;
% plot(Bestf);
end
评论0