新手熟悉bagging和boosting算法在虹膜

function [err,perfm,avg_div]=iris_rs_boosting_NNE(hidden_num,routin_num,N) %routin_num 重采样轮数赋值 %N 待集成的最好的N个成员网络 org_train=load('iris_train.txt'); [h_n,l_n]=size(org_train); cls_n=org_train(h_n,l_n); w_sample=ones(1,h_n)/h_n; %sample weights initization w_component=ones(1,routin_num)/routin_num; % 网络权值初始化. net_num=routin_num; % 网络数=重采样轮数 for i=1:routin_num rand('state',sum(100*clock)); fid=fopen('iris_bagging_train.ros','w+'); fprintf(fid,'A1 A2 A3 A4 DEC\n'); fprintf(fid,'float(2) float(2) float(2) float(2) integer\n'); for j=1:h_n sample_Distrb(j)=sum(w_sample(1:j)); %样本权值初始化 end for j=1:h_n t=find((sample_Distrb>rand)==1); %采样 id=t(1); boosting_train(j,:)=org_train(id,:); % boosting_train 采样样本子集 fprintf(fid,'%4.2f,%4.2f,%4.2f,%4.2f,%2d\n',org_train(id,:)); end fclose(fid); %调用Rosetta命令格式函数,实现离散化,约简求取输出,将约简提取出来后生成训练样本,iris_cmds.txt是rostta的命令脚本文件 !D:\applog\rosetta\clrosetta SerialExecutor "OUTPUT=0; FILENAME.COMMANDS = iris_cmds.txt; FILENAME.LOG= log.txt" fid=fopen('iris_bagging_reduct.txt','r'); while ~isempty(fgetl(fid)) 1; %do nothing end line=fgetl(fid); fclose(fid); k=1; attr_id_str=[]; clear attr_id; c_n=1; while c_n<size(line,2) while line(c_n)~=','&line(c_n)~='}' if ~isempty(str2num(line(c_n))) attr_id_str=[attr_id_str line(c_n)]; end c_n=c_n+1; end attr_id(k)=str2num(attr_id_str); k=k+1; while line(c_n)==',' c_n=c_n+1; end attr_id_str=[]; end attr{i}=attr_id; save iris_attr attr; test=load('iris_test.txt'); test_target=test(:,l_n:(l_n+cls_n-1))'; testtarget = (test_target == repmat(max(test_target),cls_n,1)); v2.P=test(:,attr_id)'; v2.T=test(:,l_n:l_n+cls_n-1)'; %生成训练样本和期望输出样本矩阵 train_input=boosting_train(:,attr_id); for j=1:cls_n idx{j}=find(boosting_train(:,l_n)==j); end train_target=ones(size(train_input,1),cls_n)/10; for j=1:cls_n train_target(idx{j},j)=0.9; end org_train_input=org_train(:,attr_id); for j=1:cls_n idx{j}=find(org_train(:,l_n)==j); end org_train_target=ones(size(org_train_input,1),cls_n)/10; for j=1:cls_n org_train_target(idx{j},j)=0.9; end org_train_target =org_train_target'; org_train_target = (org_train_target == repmat(max(org_train_target),cls_n,1)); % train the component neural networks net=newff(MinMax(train_input'),[hidden_num cls_n],{'tansig' 'purelin'},'trainlm'); init(net); net.trainParam.epochs=1000; net.trainParam.goal=0.0; net=train(net,train_input',train_target',[],[],v2); % save the component neural networks netfile = strcat('net',dec2base(i,10)); save(netfile,'net'); output=sim(net, org_train_input'); output=(output==repmat(max(output),cls_n,1)); err_id= sum(xor(output,org_train_target))~=0; err = sum(w_sample(err_id)); if err>0.5 net_num=i; i=routin_num; break else w_component(i)=log((1-err)/err); end err_id=find(sum(xor(output,org_train_target))~=0); w_sample(err_id)=w_sample(err_id)/(2*err); cor_id=find(sum(xor(output,org_train_target))==0); w_sample(cor_id)=w_sample(cor_id)/(2*(1-err)); w_sample=w_sample/sum(w_sample); end enoutput=zeros(cls_n,size(test,1)); if nargin<3 %used all network ensemble output fprintf('all network ensemble result:'); for i=1:net_num netfile = strcat('net',dec2base(i,10)); load(netfile); test_input=test(:,attr{i}); output=sim(net, test_input'); output=(output==repmat(max(output),cls_n,1)); enoutput = enoutput + w_component(i)*output; end enoutput = (enoutput == repmat(max(enoutput),cls_n,1)); err = sum(any(xor(enoutput,testtarget))) / size(test,1); perfm=1-err; avg_div=iris_div_measure(net_num); else % used best N neural network to ensemble fprintf('the best %d network ensemble result:',N); for i = 1:net_num netfile = strcat('net',dec2base(i,10)); load(netfile); test_input=test(:,attr{i}); output=sim(net, test_input'); output=(output==repmat(max(output),cls_n,1)); %%？？？？？？？？？？？？？？？ err_member(i) = sum(any(xor(output,testtarget))) / size(test,1); end [err_seq,id]=sort(err_member); % 各个网络误差排序 id_min=id(1:N); % 取前N个网络 for i = 1:size(id_min,2) netfile = strcat('net',dec2base(id_min(i),10)); load(netfile); test_input=test(:,attr{id_min(i)}); output=sim(net, test_input'); output=(output==repmat(max(output),cls_n,1)); enoutput = enoutput + output; end enoutput = (enoutput == repmat(max(enoutput),cls_n,1)); err = sum(any(xor(enoutput,testtarget))) / size(test,1) ; perfm=1-err; avg_div=iris_div_measure(net_num,id_min); end