How can i predict data by using neural network from input after fitting the data??

I used Neural Network fitting tool for training my data and got outputs for each target that i supplied to the network. Those outputs are well within the error range and give a good fit for the network. But, now i want to predict output based on input samples not included within the data set that i previously provided to the nnftool for getting the outputs. Please tell me how i can do that? The input samples are withing the training set range.

[sidra muqaddas]

x = [0 1 0; 0 1 0; 1 1 0; 1 1 1; 0 1 1; 1 1 1; 0 1 1; 1 1 0]; % three samples from input training data t = [0 0 1; 1 0 0 ; 0 1 0; 0 0 0;0 0 0;0 0 0]; %three samples from target training data [ni N]=size(x); % ni= no of input neurons [no N]=size(t); %no= no of output neurons nh=8; % no of hidden neurons in hidden layer wih = 0.01*randn(nh,ni+1); %weight matrix (iput to hidden layer) who = 0.01*randn(no,nh+1); %weight matrix (hidden to output layer) c = 0; while(c < 1000) c = c+1;

   for i=1:N
        for j = 1:nh
            netj(j) = wih(j,1:end-1)*x(:,i)+wih(j,end);
            outj(j) = tansig(netj(j));
        end

 for k = 1:no            
           netk(k) = who(k,1:end-1)*outj' + who(k,end);
           outk(k) = 1./(1+exp(-netk(k))); 
           delk(k) = outk(k)*(1-outk(k))*(t(k,i)-outk(k));
 end
 %back propagation
 for j = 1:nh
            s=0; 
            for k = 1:no 
               s = s + who(k,j)*delk(k); 
            end
            delj(j) = outj(j)*(1-outj(j))*s;

 end      
        for k = 1:no
            for l = 1:nh
                who(k,l) = who(k,l)+.5*delk(k)*outj(l);
            end
            who(k,l+1) = who(k,l+1)+1*delk(k)*1;
        end  
        for j = 1:nh
            for ii = 1:ni
                wih(j,ii) = wih(j,ii)+.5*delj(j)*x(ii,i);
            end
            wih(j,ii+1) = wih(j,ii+1)+1*delj(j)*1;
        end    
   end
 end
h = tansig(wih*[x;ones(1,N)]);

y = logsig(who*[h;ones(1,N)]); y=round(y); e = t-y; % new iput to the network csr=[0 1 0 0 0 0 1 0]; % current sensor reading

i need to add more training samples. right now my question is how to predict the corresponding output for csr

[sidra muqaddas]

how to predict output from a new input,after you have done with the training.(using code not nntoolbox variables)

[sidra muqaddas]

x = [0 1 0; 0 1 0; 1 1 0; 1 1 1; 0 1 1; 1 1 1; 0 1 1; 1 1 0]; % three samples from input training data
t = [0 0 1; 1 0 0 ; 0 1 0; 0 0 0;0 0 0;0 0 0]; %three samples from target training data
[ni N]=size(x); % ni= no of input neurons
[no N]=size(t); %no= no of output neurons
nh=8; % no of hidden neurons in hidden layer
wih = 0.01*randn(nh,ni+1); %weight matrix (iput to hidden layer)
who = 0.01*randn(no,nh+1); %weight matrix (hidden to output layer)
c = 0;
while(c < 1000)
   c = c+1;

   for i=1:N
        for j = 1:nh
            netj(j) = wih(j,1:end-1)*x(:,i)+wih(j,end);
            outj(j) = tansig(netj(j));
        end

 for k = 1:no            
           netk(k) = who(k,1:end-1)*outj' + who(k,end);
           outk(k) = 1./(1+exp(-netk(k))); 
           delk(k) = outk(k)*(1-outk(k))*(t(k,i)-outk(k));
 end
 %back propagation
 for j = 1:nh
            s=0; 
            for k = 1:no 
               s = s + who(k,j)*delk(k); 
            end
            delj(j) = outj(j)*(1-outj(j))*s;

 end      
        for k = 1:no
            for l = 1:nh
                who(k,l) = who(k,l)+.5*delk(k)*outj(l);
            end
            who(k,l+1) = who(k,l+1)+1*delk(k)*1;
        end  
        for j = 1:nh
            for ii = 1:ni
                wih(j,ii) = wih(j,ii)+.5*delj(j)*x(ii,i);
            end
            wih(j,ii+1) = wih(j,ii+1)+1*delj(j)*1;
        end    
   end
 end
h = tansig(wih*[x;ones(1,N)]);

y = logsig(who*[h;ones(1,N)]); y=round(y); e = t-y; % new iput to the network csr=[0 1 0 0 0 0 1 0]; % current sensor reading

Answer 1

[Greg Heath]

Incorrect understanding:

Generalization: Ability to perform well on nontraining data

Overfitting: Number of training equations, Ntrneq, not being sufficiently larger than the number of unknown weights, Nw, can be a cause of DECREASED generalization.

Mitigation: Either increase Ndof and/or use validation stopping(default) and/or use regularization (e.g., TRAINBR)

Insufficient information:

 size(input)  = [ I N ] = [ ? ? ]
 size(target) = [ O N ] = [ ? ? ]
 default number of training examples Ntrn = N-2*round(0.15*N) = ?
 number of training equations Ntrneq = Ntrn*O
 reference mean-square errors
       MSEtrn00  = mean(var(trntarget',1)) % Biased
       MSEtrn00a = mean(var(trntarget',0))% DOF adjusted
       MSEval00  = mean(var(valtarget',1)) % Unbiased
       MSEtst00  = mean(var(tsttarget',1)) % Unbiased

 number of hidden nodes, H = ?
 number of unknown weights Nw = (I+1)*H+(H+1)*O = ?
 number of estimation degrees of freedom Ndof = Ntrneq-Nw = ?
normalized-mean-squuare-errors
      SSEtrn  = sse(trntarget-trnoutput)
      MSEtrn  = SSEtrn/Ntrneq   % mse(trntarget-trnoutput)
      MSEtrna = SSEtrn/Ndof
      NMSEtrn =  MSEtrn/MSEtrn00
      NMSEtrna = MSEtrna/MSEtrn00a
      NMSEval  = MSEval/MSEval00
      NMSEtst  = MSEtst/MSEtst00

[Atiyo Banerjee]

Well thanks for the information, really helpful to me..i used LM algorithm with 12 neurons in 1 hidden layer and it simulated fairly well..and could mimic the parabolic behavior of the system. But some of the results are inconsistent based on the experimental data that i had with me. Maybe the values are falling in local minima traps and the learning is not that good for the system. I referred to a research paper on absorption that did the same experimental fitting and used 1200 data sets for training. In my situation, i had about 150 static datasets of 4 different variables. At first i was getting repetitive values in the simulation results, then i corrected it by changing weights and re-training the network.

Answer 2

[Greg Heath]

newoutput = net(newinput)

THank you for formally accepting my answer

Greg

[Atiyo Banerjee]

i used 'simulate' option in the nntool to get the outputs but based on the data they aren't giving good results. Increasing the number of neurons overfits the function and generalizes it.What is the solution to this problem?

[Greg Heath]

You still seem confused. Please reread what I have written about overfitting degrading generalization (the ability to perform well on nontraining data).

It is better to try to minimize the number of hidden nodes subject to the constraint that the MSE is no greater than a specified value like 0.01*Ndof^MSE00a/Ntrneq.

I have posted tens of examples. Most use a double loop approach designing ~10 separate random weight-initialization and data-division designs for each value of ~ 10 values of H that are considered. Try searching

 greg fitnet Hub Ntrials

for details. Note that

Ntrneq ~ 0.7*150 = 105

Nw = (4+1+1)*12+1 = 73 > Ntrneq/2

Hope this helps.

Greg

[Atiyo Banerjee]

Thank you so much for your conceptual reply...i am still searching for deeper concepts on the neural network problem solving ability..and yes, there always remains the notion whether the system is over-defined, under-defined or well defined..the problem in using mse as the performance function is that it gives the mean deviations summed up over the observation but does not tell us about the individual deviations or the number of observations that have deviated from optimum performance..in that case keeping the weights to a necessary value would solve the problem.

[Greg Heath]

You can always superimpose output plots (red) over target plots (blue) to obtain a better understanding of what causes errors.