Neural Network tool nftool

1. Once a network is trained how can it be used to find new outputs from known inputs?

2. There is a problem ....When I use nftool with some data set (having 5-dimensional inputs and 4-dimensional outputs) It fit very well. nftool takes 1000 iteration for it. and performance is very well around e-02..... But When I use advanced script generated by nftool, or When I write my own code for the same. It takes 12-20 iteration and performance is as high as e+06 to e+07.

Where is the problem?

Please help !

[Greg Heath]

Not enough info.

It would help if you tried your code on one of the nndatasets

help nndatasets

and post relevant parts of code and all error messages.

[sandeep]

When I tried the code with nndatasets....There is no problem. But When I used the same code with my datasets. This problem arises. Even the code generated by the nftool (Advanced scripts) is not working. My code is this. inputs=[] (Here I have given my input matrix) outputs=[] (Here I have given my output matrix) net=feedforwardnet; net=configure(net,inputs,Outputs); view(net) [net,tr]=train(net,inputs,Outputs);

P.S. There is no error messages, program is running but takes very few iteration and performance is around e+06 or e+07.

Waiting your reply.

Answer 1

[Greg Heath]

Not enough information.

 1. Your spelling of outputs is inconsistent

 2. [ I N ] = size(inputs) =?   

 3  [ O N ] = size(outputs) = ?

 4. MSE00 = mean(var(outputs'),1) = ?   % Reference MSE 

 5. Ntrn = N - 2*(0.15*N)   % Default No. of training examples 

 6. Ntrneq = Ntrn*O         % No. of training equations

 7. H = 10 % default No. of hidden nodes

 8. Nw = (I+1)*H+(H+1)*O % No. of unknown weights to be estimated from Ntrneq

 9. In order to avoid a poor set of initial random weights, obtain Ntrials = 10 
separate designs by looping over i = 1:Ntrials

 10. Pick the best design with the lowest normalized VALIDATION MSE

     NMSEval = tr.best_vperf/MSE00   % Desire 0 < NMSEval << 1

Hope this helps.

Thank you for formally accepting my answer

Greg

P.S. Search on greg i=1:Ntrials