i have problem with ffnn on matlab

this is my code

clear
clc
input=[1 2 3];
target=[5 8 4];
n=2; 
net=feedforwardnet(n);
net=configure(net,input,target);
net.layers{1}.transferFcn = 'purelin';
net.layers{2}.transferFcn = 'purelin';
ukuran1=size(input);
ukuran2=size(target);
m=ukuran1(1,1); %ini banyak input
o=ukuran2(1,1); %ini banyak output
kk=((m+1+o)*n)+1; %ini banyak bobot
net.iw{1}=[0;0]; 
net.lw{2}=[0 0]; 
net.b{1}=[0;0]; 
net.b{2}=[0];
y=net(input)

why the output (y) always different with my manual computation, can someone help me? maybe there is somethings i didnt know about FFNN matlab that

[Greg Heath]

 clear, clc
input=[1 2 3]; target=[5 8 4];
n=2; net=feedforwardnet(n);
net=configure(net,input,target);

% 1. EXPLICIT CONFIGURATION IS UNNECCESSARY IF YOU TRAIN THE NET

net.layers{1}.transferFcn = 'purelin';

% 2. NON OUTPUT LAYERS SHOULD HAVE NONLINEAR TRANSFER FUNCTIONS

net.layers{2}.transferFcn = 'purelin'; ukuran1=size(input); ukuran2=size(target); m=ukuran1(1,1); %ini banyak input o=ukuran2(1,1); %ini banyak output kk=((m+1+o)*n)+1; %ini banyak bobot net.iw{1}=[0;0]; net.lw{2}=[0 0]; net.b{1}=[0;0]; net.b{2}=[0];

% 3. INITIAL WEIGHTS SHOULD BE RANDOM. YOU CAN ASSIGN RANDOM VALUES, OR USE CONFIGURE OR JUST TRAIN THE NET WHICH WILL AUTOMATICALLY INITIALIZE WEIGHTS.

y=net(input)

% THE NET HAS NO WEIGHTS! YOU NEED TO TRAIN THE NET.

SEE MY ANSWER

GREG

[affan hanafie]

thanks for your answaer

Answer 1

[Greg Heath]

 close all, clear, clc 
% N      No. of input and output-target vectors
% I      Input dimension
% O      Output dimension
% H      No. of hidden units
% I-H-O  Typical network node topology
% Nw = (I+1)*H+(H+1)*O     No. of weights
%
% Ntst = round(0.15*N) Default No. of test vectors
% Nval = Ntst          Default No. of validation design vectors
% Ntrn = N-Nval-Ntst   Default No. of training design vectors (~0.7*N)
% 
% Training vectors: Design vectors used to determine network weight values
% Validation vectors: Design vectors used to stop training when val errors 
%         increase continually for 6 (default) continuous training epochs
% Test vectors: Nondesign vectors used to obtain UNBIASED performance 
% estimates

 input=[1 2 3]; target=[5 8 4];
[I N ] = size(input) % [1 3]
[O N ] = size(target)%[ 1 3 ]
plot(input,target)
rng(0)    % Allows duplication of results
n=2; net=feedforwardnet(n);
[net tr output error ] = train(net,input,target);
% Normalized Mean Square Error 0<= NMSE <= 1

   NMSE = mse(error)/var(target,1)     % 4.3684e-26

[Greg Heath]

1. See the documentation examples in

 help feedforwardnet

and

 doc feedforwardnet

2. For more examples, search ANSWERS using

 greg feedforwardnet

Hope this helps.

Greg

[affan hanafie]

thanks man