mlp.m

这是在国外某网站上下的一些自适应信号处理的源程序

%MLP Multilayer Perceptron algorithm
%
%   'ifile.mat' - input file containing:
%      Nr - members of ensemble
%      dim - iterations
%      Nneur - number of neurons
%      Sx - standard deviation of input
%      Sn - standard deviation of measurement noise
%      u - convergence factor
%      
%   'ofile.mat' - output file containing:
%      MSE - mean-square error


clear all		% clear memory
load ifile;		% read input variables

for j=1:Nr
   n=Sn*randn(dim,1);           % noise at system output   
   x=Sx*(randn(dim,1));            % input signal 
   xl(1)=0; xl(2:dim)=x(1:dim-1);
   d=zeros(dim,1);
   yl(1)=0; dl(1)=0;
   % initial training vectors
   w1=.2*randn(Nneur,3); w2=.2*randn(Nneur,Nneur); w3=.2*randn(Nneur,1);
   b1=.1*randn(Nneur,1); b2=.1*randn(Nneur,1);  b3=.1*randn(1,1);
   for i=1:dim
      d(i)=-.6*dl(i)+x(i)+.1*x(i)*dl(i)+.2*xl(i)*dl(i)+n(i); % unknown system output sample
      uxl=[x(i) dl(i) xl(i)]';     % new input vector
      aux=w2*sgm(w1*uxl-b1,1,2)-b2;
      dr(i)=sgm(aux,1,2)'*w3-b3;   % neural network output sample
      er_out=d(i)-dr(i);           % error sample
      er_hid2=er_out*w3.*sgd(aux,1,2);
      er_hid=w2'*er_hid2.*sgd(w1*uxl-b1,1,2);
     dw3=2*u*er_out*sgm(aux,1,2);
     db3=-2*u*er_out;
     dw2=2*u*er_hid2*sgm(w1*uxl-b1,1,2)';
     db2=-2*u*er_hid2;
     dw1=2*u*er_hid*uxl';
     db1=-2*u*er_hid;       
      w1=w1+dw1; b1=b1+db1;        % new coefficient vectors for layer 1
      w2=w2+dw2; b2=b2+db2;        % new coefficient vectors for layer 2
      w3=w3+dw3; b3=b3+db3;        % new coefficient vectors for layer 3
      mse(j,i)=(d(i)-dr(i))^2;
      dl(i+1)=d(i);
   end
end  

MSE=mean(mse);
save ofile MSE;