My Question About Neural Net

Hello , First.. let me say Thanks for read my question, please answer some if you understand.

Today i was test neural net for predict data but some output result is too far with target. I think i need advanced guide for this, because i just use nntool and there is much settings that i don't understand how to tweak it and even i try to change it the result is still far from target.

for information ,the sample data i use is dynamic change every time, its 20 input and 1 output. i was read some code example and try to modify and still nothing.

My question is :

1. Can i use net for predict future data ? "if no please give me other way to do this"

2. How much sample data for net 'best' ?

3. How i can get report success & fail rate in scale of 0-100% from net training result? i want compare old net settings with new net "please give me example command for this"

4. How i get the best net settings ? what parameters should i change for improvement ?

5. Is there any fast way to test and get best net ?

That is..

Once again, Thank You For Help.

Answer 1

[Greg Heath]

I'm not a fan of first learning with nn or ntstool. I suggest the commandline approach as posted in the HELP and DOC documentation

> 1. Can i use net for predict future data ? "if no please give me other way to do this"

 help/doc timedelaynet % past/current input ==> current/future output
 help/doc narnet   % fedback past output ==> current/future output
 help/doc narxnet  % Both of the above

2. How much sample data for net 'best' ?

Typically you don't want fewer training equations (Ntrneq=Ntrn*O), than there are unknown weights Nw. For a given number, Ntrn, of training "I"nput/"O"utput-target pairs of dimensions I and O, respectively, this leads to Hub, an upper bound on the number of hidden layer nodes OR, for a given number of hidden nodes, H, a lower bound on the number of training sample pairs, Ntrn.

 You can search the NEWSGROUP and ANSWERS for some of my tutorials using

 greg narxnet Hub Ntrneq

for details.

3. How i can get report success & fail rate in scale of 0-100% from net training result? i want compare old net settings with new net "please give me example command for this"

 NMSE = mse(target-output)/mean(var(target',1))
 Rsq  = 1 - NMSE

is the fraction of mean target variance that is modeled by the net. I favor an upper bound of NMSE <= 0.005 (Rsq >= 99.5%) for openloop timeseries and 0.01 (Rsq >= 99%) for closedloop timeseries.

4. How i get the best net settings ? what parameters should i change for improvement ?

See my narxnet tutorials where, given a set of significant input and feedback delays, a double loop search over number of hidden nodes and random initial weights tends to lead to the best solutions.

5. Is there any fast way to test and get best net ?

In general, Nope.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Ashley]

Thank You Greg. I Will try your tutorial.. Have a nice day.

Answer 2

[Walter Roberson]

Use ntstool instead of nntool