Nov 12, 2011 12:19:00 PM, goutam_nalamati@yahoo.com wrote:

>Hello Greg Heath, my name is Goutam. I have seen your >posts on the mathwo=
rks solving problems for many people in >field of neural networks. I am doi=
ng a mini-project on Brest >cancer detection based on Radial basis function=
.. I have a set >of 600 sample. I used dividerand to divide the whole data s=
et >into 60,20,20. I want to train the network with 60% data and >use 20% d=
ata for validation.I want to use the newrb >command for training. Newrb fun=
ction tries to minimize the t>raining error only. I want to use my validati=
on set to create >the early stopping point. Please let me know how can i do=
 >this.

>Thank you in advance

Early-Stopping/Stopped-Training is not a NEWRB option.
I just train NEWRB with MSEgoal =3D mean(var(ttrn'))/200
( --> Rtrn^2 >=3D 0.995) and use the validation set error to
optimize the spread. If you wish, you can get different
results for a given spread by changing the identity of the
first training vector.

An alternative is to use NEWFF with Early-Stopping
and RADBAS activations for the hidden layer.

Hope this helps.

Greg

On Nov 13, 2:52=A0pm, Greg Heath <he...@alumni.brown.edu> wrote:
> Nov 12, 2011 12:19:00 PM, goutam_nalam...@yahoo.com wrote:
>
> >Hello Greg Heath, my name is Goutam. I have seen your >posts on the math=
works solving problems for many people in >field of neural networks. I am d=
oing a mini-project on Brest >cancer detection based on Radial basis functi=
on. I have a set >of 600 sample. I used dividerand to divide the whole data=
 set >into 60,20,20. I want to train the network with 60% data and >use 20%=
 data for validation.I want to use the newrb >command for training. Newrb f=
unction tries to minimize the t>raining error only. I want to use my valida=
tion set to create >the early stopping point. Please let me know how can i =
do >this.
> >Thank you in advance
>
> Early-Stopping/Stopped-Training is not a NEWRB option.
> I just train NEWRB with MSEgoal =3D mean(var(ttrn'))/200
> ( --> Rtrn^2 >=3D 0.995) and use the validation set error to
> optimize the spread. If you wish, you can get different
> results for a given spread by changing the identity of the
> first training vector.
>
> An alternative is to use NEWFF with Early-Stopping
> and RADBAS activations for the hidden layer.
>
> Hope this helps.
>
> Greg

See

http://groups.google.com/group/comp.soft-sys.matlab/msg/a7dbae46b73e7f26