[求助] 激活函数purelin和logsig为何差别怎么大

function BBP
clear all;
clc;
disp('former Batch BP:');
%%{
%取得当前时钟，设置随机种子
rand('state',sum(100*clock))
%数据产生机制
     D1(1,:)=normrnd(0,1,1,5000);
D1(2,:)=0;
     D2=normrnd(80,1,1,5000);
D2(2,:)=1;
D0=[D1,D2];
D_length=length(D0);
DR=randperm(D_length);
D=D0(:,DR);
%数据已经随机打乱
%将数据D写入文本，并按格式对齐，且保留四位小数
dlmwrite('data0.txt',D,'delimiter','\t','precision','%.4f');
%}
%从文本中读入数据
%1·数据归一化处理，省去
%2·将数据划分为两部分
testpercent=0.4;
trainpercent=0.6;
trainSamples=D(:,1:(D_length*trainpercent));
trainSamplesP=trainSamples(1,:);
trainSamplesT=trainSamples(2,:);
testSamples=D(:,((D_length*trainpercent)+1):D_length);
testSamplesP=testSamples(1,:);
testSamplesT=testSamples(2,:);
%数据划分完毕
%3·建立网络
firstNodeHidNum=3;%第一层隐藏层节点的数目
net=newff(trainSamplesP,trainSamplesT,[firstNodeHidNum],{'logsig','purelin'},'traingd');
%4·训练原始批梯度下降BP神经网络
net.trainParam.epochs=250;
net.trainParam.goal=0.001;
net.trainParam.lr=0.1;
net.trainParam.show=100;
[net,tr]=traingd(net,trainSamplesP,trainSamplesT);
%5·仿真测试并构造结果矩阵BBP_res写入结果文本
resSim=sim(net,testSamplesP);
res=im2bw(resSim,0.5);
sum_BBP_right=0;%计算正确分类的数目
for j=1:1:length(testSamples)
     if res(j)~=testSamplesT(j)        
         rig_err(j)=-8;
     else
         sum_BBP_right=sum_BBP_right+1;
         rig_err(j)=8;
     end
end
BBP_res(1,:)=testSamplesP;
BBP_res(2,:)=testSamplesT;
BBP_res(3,:)=res;
BBP_res(4,:)=rig_err;
dlmwrite('BBP_res.txt',BBP_res,'delimiter','\t','precision','%.4f');
disp('原批梯度下降BP的正确率:');
sum_BBP_right
length(testSamples)
BBP_right_rate=sum_BBP_right/length(testSamples)
end

8楼: Originally posted by nono2009 at 2013-08-02 12:41:09
没做过分类问题，所以看不了...

2楼: Originally posted by ajitai at 2013-08-02 11:11:35
必须差别大啊，看看这两个函数的公式你就知道了。。。当然还和你写的程序还有关系