%使用遗传算法进行特征选择，学习算法采用贝叶斯
clear
load(2-Class Problem.mat);
tic
% 产生初始化种群，其中含有50个个体，一个个体含有190个维度，最终选择出40个最优的维度
pop_init=pop_init_genrator(50,190,40);
pop_last=pop_init;
% 开始迭代
for iter=1:20
fprintf(开始第%d次迭代\n,iter);
pop_new=pop_select(pop_last,Training_class1,Training_class2,Testing,Label_Testing);%选择
pop_new=cross(pop_new);%交叉
pop_new=variation(pop_new);%变异
pop_last=pop_new;
end
pop_selected=pop_last;
%计算适应度，选择最优的个体，即选出最终的特征
% fit=fitness_jd(pop_selected,Training_class1,Training_class2);
fit=fitness_bayes(pop_selected,Training_class1,Training_class2,Testing,Label_Testing);
fit_sort=sort(fit);
best_pop=pop_selected(fit==fit_sort(end),:);
feature_selected=find(best_pop==1);
save(feature_selected_GA,feature_selected);
toc
% 产生初始种群的函数 pop_nums:种群数量 precision:精度，这里指特征的总数量
% positive：一个个体中‘1’的个数，这里指最终选择40个特征
function pop_init=pop_init_genrator(pop_nums,precision,positive)
pop_init=zeros(pop_nums,precision);
for i=1:pop_nums
pop_init(i,:)=randerr(1,precision,positive);
end
end
% 选择，使用赌轮算法
function pop_new=pop_select(pop_last,train_class1,train_class2,Testing,Label_Testing)
disk=0;%初始化轮盘
% fit=fitness_jd(pop_last,train_class1,train_class2);%计算上一代每个个体的适应度
fit=fitness_bayes(pop_last,train_class1,train_class2,Testing,Label_Testing);
pop_nums=size(fit,2);
% 定义轮盘，每一个个体的区域正比于个体的适应度，个体适应度越高，被选中的概率越大
for i=1:pop_nums
disk(i)=sum(fit(1:i));
end
rand_=max(fit)*rand(1,pop_nums);%产生一组随机数，随机数的个数等于种群的个体数
%选择
for i=1:pop_nums
temp=find(disk<rand_(i));
index=size(temp,2)+1;%找到随机数落在轮盘的位置
pop_new(i,:)=pop_last(index,:);
end
end
%交叉
function pop_output=cross(pop_input)
pop_nums=size(pop_input,1);%种群数量
pop_dim=size(pop_input,2);%特征数
%将种群随机分成两部分，作为两个父代
select_index=randperm(pop_nums,floor(pop_nums/2));
father1=pop_input(select_index,:);
father2=pop_input;
father2(select_index,:)=[];
for i=1:pop_nums/2
cross_len=randi(pop_dim/2,1);%产生一个随机数，以确定交叉片段的长度
father1_cross=father1(i,1:cross_len);
pos_num1=size(father1_cross(father1_cross==1),2);%统计交叉片段中1的个数
% 在第二个父代中搜寻同样长度的片段，并且1的个数应该与父代1中的待交换片段相等
% 以满足交换后特征数不变，若搜寻不到，则跳过此次交叉过程
for j=1:pop_dim-cross_len+1
temp=father2(i,j:j+cross_len-1);
pos_num2=size(temp(temp==1),2);
if pos_num1==pos_num2
%交换
father1(i,1:cross_len)=temp;
father2(i,j:j+cross_len-1)=father1_cross;
break;
end
end
end
pop_output=[father1;father2];
end
%变异
function pop_output=variation(pop_input)
pop_nums=size(pop_input,1);%种群数量
pop_dim=size(pop_input,2);%特征数
prob=rand(1,pop_nums);%产生一组随机数，决定是否变异的概率，大于0.5则变异
for i=1:pop_nums
if prob(i)<0.5
continue;
else
%产生两个个随机数，确定变异的位置，为确保变异后特征数不变，需要变异的两个位置值不能相等
index1=randi(pop_dim,1);
index2=randi(pop_dim,1);
while pop_input(i,index1)==pop_input(i,index2)
index2=randi(pop_dim,1);
end
pop_input(i,index1)=~pop_input(i,index1);
pop_input(i,index2)=~pop_input(i,index2);
end
end
pop_output=pop_input;
end
% 适应度函数,选择类间距离，距离度量采用欧氏距离
function fit=fitness_jd(pop,train_class1,train_class2)
pop_num=size(pop,1);
class1_num=size(train_class1,2);
class2_num=size(train_class2,2);
for i=1:pop_num
d=0;
for j=1:class1_num
for k=1:class2_num
d=d+(train_class1(pop(i,:)==1,j)-train_class2(pop(i,:)==1,k))*(train_class1(pop(i,:)==1,j)-train_class2(pop(i,:)==1,k));
end
end
jd=d/(class1_num*class2_num);
fit(i)=jd;
end
end
%贝叶斯算法的准确度作为可分性依据
function fit=fitness_bayes(pop,Training_class1,Training_class2,testing_data,Testing_label)
pop_num=size(pop,1);
%获取训练数量
n1=size(Training_class1,2);
n2=size(Training_class1,2);
% 先验概率
pw1=n1/(n1+n2);
pw2=n2/(n1+n2);
for p=1:pop_num
%选择训练数据和测试数据相应位置的特征值
Training_data1=Training_class1(pop(p,:)==1,:);
Training_data2=Training_class2(pop(p,:)==1,:);
Testing_data=testing_data(pop(p,:)==1,:);
%参数估计
[miu1,sigma1]=ParamerEstimation(Training_data1);
[miu2,sigma2]=ParamerEstimation(Training_data2);
%预测结果
predict_label=0;
test_num=size(Testing_data,2);%获取测试数据数量
for i=1:test_num
x=Testing_data(:,i);
pxw1=gaussian(miu1,sigma1,x);
pxw2=gaussian(miu2,sigma2,x);
if pw1*pxw1>pw2*pxw2
predict_label(i)=1;
else
predict_label(i)=2;
end
end
% 计算精度
acc=sum(predict_label==Testing_label)/test_num;
end
fit(p)=acc;
end
