BP神经网络

function [W,err]=BPTrain(data,label,hiddenlayers,nodes,type)
%Train the bp artial nueral net work
%input data,label,layers,nodes,type
%data：dim*n
%label:1*n
%layers:m:number of hidden layers
%nodes:num_1;num_2...num_m
%type==1:create and train
%type==0:train
%tanh / 双曲正切： tanh(x) = sinh(x) / cosh(x)=[e^x - e^(-x)] / [e^x + e^(-x)]
%(tanh(x))‘=sech^2(x)
%sech / 双曲正割： sech(x) = 1 / cosh(x) = 2 / [e^x + e^(-x)]
if type==1
   %create the nureal network and train
   nodes=[size(data,1);nodes];
   nodes=[nodes+1;size(label,1)];
   %W{1}=random(,nodes(1));
   layers=hiddenlayers+2;
   for i=1:layers-2
       W{i}=rand(nodes(i),nodes(i+1)-1);
   end
   W{layers-1}=rand(nodes(layers-1),nodes(layers));
else
    %do nothing
end
%train the bp network
%the termination condition
%iteration.error
iter=0;
error=inf;
maxiter=2000;
lr=0.1;
epision=0.1;
tic
while iter<maxiter&&error>epision
    iter=iter+1;
    error=0;
    for k=1:size(data,2)
        %forward process
        y{1}=[data(:,k)];
        v{1}=y{1};
        for i=1:layers-1
            y{i}=[1;y{i}]; 
            v{i+1}=W{i}‘*y{i};
            y{i+1}=tanh(v{i+1});
        end
        %back process
        error=error+abs(label(k)-y{layers});
        delta=(label(k)-y{layers}).*((sech(v{layers}).^2));
        W{layers-1}=W{layers-1}+lr.*(y{layers-1}*delta);
        for i=layers-1:-1:2     
            delta=sech(v{i}).^2.*(W{i}(1:size(W{i},1)-1,:)*delta);     
            W{i-1}=W{i-1}+lr.*(y{i-1}*delta‘);
        end
    end 
    err(iter)=error;
    error
end
toc

function res=BPTest(W,data)
for k=1:size(data,2)
   y=data(:,k); 
   for i=1:length(W)-1
       y=[1;y];
       y=tanh((W{i}‘*y));
   end
   res(k)=tanh(W{i+1}‘*[1;y]);
end

global rbf_sigma;
global rbf_center;
global rbf_weight;
if strcmp(traintype,‘data‘)
    traindist=pdist2(traindata,traindata);
    rbf_sigma=max(max(traindist))/(scale.^2);%/(2*sqrt(sqrt(length(traindata))));
    rbf_center=traindata;
    Phi=exp(-traindist./rbf_sigma);
    rbf_weight=inv(Phi)*trainlabel;
    
else if strcmp(traintype,‘cluster‘)
        [Idx,C,sumD,D]=kmeans(traindata,K,‘emptyaction‘,‘singleton‘);
        traindist=pdist2(traindata,C);
        Cdist=pdist2(C,C);
        rbf_sigma=max(max(Cdist))/(scale.^2);%/(2*sqrt(sqrt(length(traindata))));
        rbf_center=C;
        Phi=exp(-traindist./rbf_sigma);
        rbf_weight=inv(Phi‘*Phi)*Phi‘*trainlabel;
    else if strcmp(traintype,‘descend‘)
            
            
        end
        
    end
end

function predcict=RBFTest(data)

global rbf_sigma;
global rbf_center;
global rbf_weight;

testdist=pdist2(data,rbf_center);

predcict=exp(-testdist./(2*rbf_sigma))*rbf_weight;