Very high ratio of the Targets / inputs

Hi All

I have to design a neural network (fitting ) with one input and 7 outputs , that the order of 100 and the outputs in the order of 1e+8 , the Gradient is too big , something like 1e+5 , so the validation stops so early which training method should I use ? or what modification should I apply on the network ? if the solution is mapminmax , how should I roll back to the real quantities of the output as the target was ?

[farzad]

Actually

without using the mapminmax , the gradient minimum is 1e14 , but the results are more acceptable , by using the mapminmax , the gradient is so smaller , but the results are not really acceptable, and the regression shows it's less accurate

WHY ?!?!?!

[farzad]

Anyone has had the same experience?

[farzad]

Even when I try with one input one output , and there is too much difference in orders , this problem persists

Answer 1

[Greg Heath]

I usually standardize inputs and outputs using ZSCORE (zero mean, unit variance) before training.

Hope this helps.

Thank you for formally accepting my answer

Greg

[farzad]

thank you dear professor

I am just not sure about this :

[ t meanT stdT] = zscore(T)

here ,

Y = meanT + stdT.*y;

I have problem in the multiplication of stdT .*y , cause y does not have the same size of stdT

[farzad]

since definitely stdT and y do now have the same dimension

[farzad]

I did

meanT= mean(tc);
stdT=std(tc);
[ tc meanT stdT] = zscore(tc)

and after training :

inp=[602.1900247744 2014.5917518769  2.4056877954  1.1305517568]';
an = sim(net3,inp)
an = meanT + stdT'* an';

but it doesn't work

the std is 1x1055 and an is 4x1

imposible to give a result

meanT is 1x1055

[farzad]

yet no sucsses

is the mapstd work the same as zcore ?

[farzad]

Is there a major difference applying mapstd instead of Zcore ?