%% BP 动量 批量学习实例
clc
close all
clear all
%% sample data
X=[0.7853 1.57];
T=[0.707 1.0];
%% initial weights and bias 三层网络结构1-1-1，隐层和输出层都采用log-sigmoid 函数
W1=0.2;
W2=0.4;
B1=0.3;
B2=-0.1;
input_num=1;
hidden_num=1;
output_num=1;
%% 参数值设定
max_epoch=1000;
epoch=0;
goal=10^(-2);
MSE=10^3;
lr=0.1;
momentum=0.5; % 动量值
old_dW1=0;
old_dB1=0;
old_dW2=0;
old_dB2=0;
k=size(X);
sample_num=k(2);
ErrHistory=[];
%% 网络训练
while epoch~=max_epoch
% 网络前向计算
hiddenOut=logsig(W1*X+repmat(B1,1,sample_num)); % 隐含层输出
networkOut=logsig(W2*hiddenOut+repmat(B2,1,sample_num)); % 网络输出层输出
err=T-networkOut;
MSE=0.5*mean(err); % 均方差采用的公式是 err^2/(2*sample_num)
ErrHistory=[ErrHistory MSE];
if MSE<=goal
break;
end
% 权值更新，误差反传
deltaW2=1*err.*networkOut.*(1-networkOut).*hiddenOut;
deltaB2=1*err.*networkOut.*(1-networkOut);
% 加入动量项
dW2=momentum*old_dW2+(1-momentum)*lr*sum(deltaW2);
dB2=momentum*old_dB2+(1-momentum)*lr*sum(deltaB2);
old_dW2=dW2;
old_dB2=dB2;
%%%对输出层与隐含层之间的权值和阈值进行修正
W2=W2+dW2;
B2=B2+dB2;
%隐含层与输入层的取值更新
deltaW1=deltaB2.*W2.*hiddenOut.*(1-hiddenOut).*X;
deltaB1=deltaB2.*W2.*hiddenOut.*(1-hiddenOut);
dW1=momentum*old_dW1+(1-momentum)*lr*sum(deltaW1);
dB1=momentum*old_dB1+(1-momentum)*lr*sum(deltaB1);
old_dW1=dW1;
old_dB1=dB1;
%%%对输入层与隐含层之间的权值和阈值进行修正
W1=W1+dW1;
B1=B1+dB1;
epoch=epoch+1;
end
fprintf('final weights and biases are: ')
W1
B1
W2
B2
epoch
MSE
