Search All of the Math Forum:

Views expressed in these public forums are not endorsed by NCTM or The Math Forum.

Topic: ANN_Error Goal
Replies: 3   Last Post: Feb 10, 2013 2:05 PM

 Messages: [ Previous | Next ]
 Greg Heath Posts: 6,325 Registered: 12/7/04
Re: ANN_Error Goal
Posted: Feb 10, 2013 2:05 PM

Subject: ANN_Error Goal
From: Suresh
Date: 10 Feb, 2013 13:34:09
"Suresh" wrote in message <kf87kh\$dta\$1@newscl01ah.mathworks.com>...
> Dear Greg,
> As you told I am following these steps to estimate Error goal for an example set.

> >[I N]=size(Input);[13x178]
> >[O N]=size(Target);[3x178]

> From the size of above matrices
> >I = 13 ; O = 3 ;N = 178 ; H =10

Did H=10 come out of midair?

If you do not use validation stopping or regularization,
(e.g.,msereg) Neq > Nw is required. This results in the
upper bound

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

To mitigate measurement error and noise, Neq >> Nw
is desired. A reasonable first try in a search for a practical
value is

H ~ round(Hub/10) = 3

Design Ntrials ~ 10 random initial weight designs. If unsuccessfull,
increase H. However I would not exceed ~ Hub/2 ~ 15 unless using
valstopping or regularization.

If you use valstopping or regularization, you can even use
H > Hub. However, I find those values esthetically displeasing.

> > Nw = (I+1)*H+(H+1)*O = (13+1)*10 + (10+1)*3 = 173
> > Neq = N*O = 178*3 = 534 % Is this correct ?
> > Ndof = Neq-Nw = 534-173 = 361
> > MSE00a = mean(var(Target')) = 0.2207

> MSEgoal=0.01*361*0.2207/534 = 0.0015;
> Is this procedure correct.

The procedure I would use is described above. However, if you
use valstopping or regularization, H = 10 will probably not be
too large.

When you finish let us know your values for H and R2a.

Hope this helps.

Greg

PS I have seen people use BOTH valstopping AND regularization.