bpnn.m

追隨sin函數,輸出有包含物插曲線圖,到傳遞神經網路

clc;clear
% L-K-J-I ANN
L=3;K=3;J=5;I=1;
% eta, learning rate
eta=1.e-02;
% No. of epochs
N=1000;
% initial conditions of weight and bias
W_kl=rands(K,L);
W_jk=rands(J,K);
W_ij=rands(I,J);
th_k=rands(K,1);
th_j=rands(J,1);
th_i=rands(I,1);
%input x; output o_i; target T
T=2.5;
x=[0.1;0.7;-0.3];
for epoch=1:N       % 
% 1. forward path
h_k=W_kl*x-th_k;
o_k=tansig(h_k);     
%
h_j=W_jk*o_k-th_j;
o_j=tansig(h_j);
%
h_i=W_ij*o_j-th_i;
o_i=h_i;            % linear
% 2. error signal delta
delta_i=(T-o_i)*1;
delta_j=(W_ij'*delta_i).*dtansig(h_j,o_j);
delta_k=(W_jk'*delta_j).*dtansig(h_k,o_k);
% 3. weight  updating
for k=1:K
   for l=1:L
      W_kl(k,l)=W_kl(k,l)+eta*delta_k(k)*x(l);
   end
end
%
for j=1:J
   for k=1:K
      W_jk(j,k)=W_jk(j,k)+eta*delta_j(j)*o_k(k);
   end
end
%
for i=1:I
   for j=1:J
      W_ij(i,j)=W_ij(i,j)+eta*delta_i(i)*o_j(j);
   end
end
% bias updating
th_k=th_k+eta*delta_k*(-1);
th_j=th_j+eta*delta_j*(-1);
th_i=th_i+eta*delta_i*(-1);
E(epoch)=0.5*delta_i*delta_i;  % error function

y(epoch)=o_i;  %
y1(epoch)=T;   %
end %   

plot(E),figure

z=1:epoch;      %
plot(z,y,z,y1,'r') %