How can I improve my neural network output?

Hello,

I have a prediction problem with 6x365 inputs( 6 represent, for example, the temperature for last six hours since now of every day from my database registrations) and 1x365 targets(I want to predict the next hour temperature). I don't how to choose the number of hidden layers, the number of nodes, the training algorithm, the transfer function.

That's my code now and the output is not very desirable:

net1 = newff(minmax(Input_temp),[20 1],{'tansig' 'purelin'},'traincgb','learngd');
net1.trainParam.epochs = 5000;
[net1,pr] = train(net1,Input_temp,Target_temp);
Output_temp = net1(Sample_temp)

Thank you for your time, Have a good day!

Answer 1

[Greg Heath]

"Not very desirable" is not very helpful. Need a numerical measure. For example

        NMSE = mse(target-output)/var(target,1) % Normalized mean-square-error

NEWFF with the (minmax ... syntax has been obsolete for 10 yrs or so.

NEWFF with the (input,target,... syntax has been obsolete for 5 yrs or so.

What version of the NNToolbox are you using? Do you have FITNET?

 [I N ] = size(input)  % [ 6 365 ]
 [O N ] = size(target) % [ 1 365 ] 

 Ntrn   = N-2*round(0.15*N) % 255 (default 70% training examples)
 Ntrneq = Ntrn*O            % 255 training equations

 % Nw = (I+1)*H + (H+1)*O    % Number of unknown weights
 % Ntrneq > Nw <==> H <= Hub % "u"pper "b"ound

 Hub = -1+ceil((Ntrneq-O)/(I+O+1))  % 31

Try 10 or more multiple designs with H as small as possible (H << Hub is highly desirable but not necessary). For each value of H create 10 designs with different random initial weights and data divisions.

For many, many examples search the NEWSGROUP and ANSWERS using

 greg Hmin:dH:Hmax Ntrials

Hope this helps.

Thank you for formally accepting my answer

Greg

[Alexandru Vasile]

Thank you! I'll try.