bpprogram.m

两层BP神经网络程序

%两层BP网络程序


inputsamplematrix=finalout1327;
%本神经网络将采用sigmoid函数，因而输出层各神经元的理想输出值只能趋近于1或0，而不能达到1或0，
%因此将理想输出设置为0.9或0.1是较为适宜的。
targetvalue=targetvalue327;
trainingtime=0;%训练次数
abs_error=0;
square_error=0;
inputweightmatrix=abs(randn(4,13));%输入层至隐含层的连接权
hiddenlayerthreshold1=randn(4,1);%隐含层阈值
hiddenlayerthreshold=zeros(4,27);%共有27个样本
for ii=1:27
    hiddenlayerthreshold(:,ii)=hiddenlayerthreshold1;
end
hiddenlayerin=zeros(4,27);%隐含层净输入
hiddenlayerout=zeros(4,27);%隐含层输出

%隐含层到输出层
layerweightmatrix=abs(randn(3,4));%隐含层到输出层的连接权
outputlayerthreshold1=randn(3,1);%输出层阈值
outputlayerthreshold=zeros(3,27);%共有27个样本
for ii=1:27
    outputlayerthreshold(:,ii)=outputlayerthreshold1;
end
outputlayerin=zeros(3,27);%输出层的净输入
outputlayerout=zeros(3,27);%输出层输出，即此BP网络的输出

outputlayer_err=zeros(3,27);%输出层个神经元的一般化误差
hiddenlayer_err=zeros(4,27);%隐含层个神经元的一般化误差


% 迭代第一次
hiddenlayerin=inputweightmatrix*inputsamplematrix-hiddenlayerthreshold;
hiddenlayerout=1./(1+exp(-hiddenlayerin));%The training function in hidden layer is sigmoid.
%隐含层到输出层
outputlayerin=layerweightmatrix*hiddenlayerout-outputlayerthreshold;
outputlayerout=1./(1+exp(-outputlayerin));%The training function in output layer is sigmoid.
%计算误差
abs_error=targetvalue-outputlayerout;
outputlayer_err=abs_error.*outputlayerout.*(ones(3,27)-outputlayerout);
hiddenlayer_err=layerweightmatrix'*outputlayer_err.*hiddenlayerout.*(ones(4,27)-hiddenlayerout);
learningrate=1;%步幅
layerweightmatrix=layerweightmatrix+learningrate*outputlayer_err*hiddenlayerout';
outputlayerthreshold=outputlayerthreshold+learningrate*outputlayer_err;
inputweightmatrix=inputweightmatrix+learningrate*hiddenlayer_err*inputsamplematrix';
hiddenlayerthreshold=hiddenlayerthreshold+learningrate*hiddenlayer_err;
    
square_error=abs_error.*abs_error;
errorfunction(1)=ones(1,3)*square_error*ones(27,1)/27/2;


for iternum=2:1000%从第二次接着迭代   
    hiddenlayerin=inputweightmatrix*inputsamplematrix-hiddenlayerthreshold;
    hiddenlayerout=1./(1+exp(-hiddenlayerin));%The training function in hidden layer is sigmoid.
    %隐含层到输出层
    outputlayerin=layerweightmatrix*hiddenlayerout-outputlayerthreshold;
    outputlayerout=1./(1+exp(-outputlayerin));%The training function in output layer is sigmoid.
    %计算误差
    abs_error=targetvalue-outputlayerout;
    outputlayer_err=abs_error.*outputlayerout.*(ones(3,27)-outputlayerout);
    hiddenlayer_err=layerweightmatrix'*outputlayer_err.*hiddenlayerout.*(ones(4,27)-hiddenlayerout);
    square_error=abs_error.*abs_error;
    errorfunction(iternum)=ones(1,3)*square_error*ones(27,1)/27/2;
    if errorfunction(iternum) < 1e-4%中途停止迭代的条件
        break;
    end
    %调整learningrate (Variable Learning Rate)
    %If the error function over the entrie training set has decreased,
    %increase the learning rate by multiplying it by a number typically
    %1.05
    if errorfunction(iternum) < errorfunction(iternum-1)
        learningrate=1.05*learningrate;
    %If the error function has increased more than some set percentage
    %typically a few percent(4%), decrease the learning rate by multiplying
    %it by a number typically 0.7
    elseif errorfunction(iternum) > 1.04*errorfunction(iternum-1)
        learningrate=0.7*learningrate;
    %If the error function is increased less than the percentage(%4), the learning rate ramins unchanged.    
    else
        learningrate=learningrate;
    end
    
    layerweightmatrix=layerweightmatrix+learningrate*outputlayer_err*hiddenlayerout';
    outputlayerthreshold=outputlayerthreshold+learningrate*outputlayer_err;
    inputweightmatrix=inputweightmatrix+learningrate*hiddenlayer_err*inputsamplematrix';
    hiddenlayerthreshold=hiddenlayerthreshold+learningrate*hiddenlayer_err;
end

plot(errorfunction)