Prev: merging ODE45 results
Next: Parallel Implementation of Code in Matlab
From: Mac Saarman on 20 Apr 2010 12:44
Hello,

I am doing research in another field and as such I am not very good at
NNet. I hope people here can give me small advise and insight on
this.

I need a trained NNet to look at current status of my system and
select one of possible 9 moves/decisions. The network will be trained
by supplying previously recorded decisions of professional humans.
Around 20 parameters are used from which 2-3 are of very higher
importance but others are also compulsory to consider.

1- I am confused on what kind of problem is this? Fitting?
Classification? Pattern Recognition? or what? (I am refering to NNet PDF document)

2- If I need to just look at current state (I guess a static network),
what kind of network is possibly suitable?

3- If I need to also consider previous states, then what?

I appreciate your time and help.

Regards,
Mac
From: Juan Isaza on 20 Apr 2010 17:55
"Mac Saarman" <macdermat-remthis(a)remthisgmail.com> wrote in message <hqklkm$fjv$1(a)fred.mathworks.com>...
> Hello,
>
> I am doing research in another field and as such I am not very good at
> NNet. I hope people here can give me small advise and insight on
> this.
>
> I need a trained NNet to look at current status of my system and
> select one of possible 9 moves/decisions. The network will be trained
> by supplying previously recorded decisions of professional humans.
> Around 20 parameters are used from which 2-3 are of very higher
> importance but others are also compulsory to consider.
>
> 1- I am confused on what kind of problem is this? Fitting?
> Classification? Pattern Recognition? or what? (I am refering to NNet PDF document)
>
> 2- If I need to just look at current state (I guess a static network),
> what kind of network is possibly suitable?
>
> 3- If I need to also consider previous states, then what?
>
> I appreciate your time and help.
>
> Regards,
> Mac

Hi Mac

Actually there are only two ways. One is classification the other one is regression. If the decisions are not comparable, I mean dont obey the basics laws of <=>, then you have a multiple classification problem with 9 instances.

One example of not camparable are 5 different illnesses. On the other hand comparable could be the age of a person.

In the multiple classification problem the NNet wil select between 9 diferent classes. I've never done that before (instead only to classes), but I now its straightforward. Maybe its a better aproach to try with a svm (support vector machine) you can download some libraries at the site of the book learning with kernels. You can also try an algorithm called Adaboost for multiple classification.

Juan Isaza
From: Greg Heath on 20 Apr 2010 18:31
On Apr 20, 12:44 pm, "Mac Saarman" <macdermat-
remt...(a)remthisgmail.com> wrote:
> Hello,
>
> I am doing research in another field and as such I am not very good at
> NNet. I hope people here can give me  small advise and insight on
> this.
>
> I need a trained NNet to look at current status of my system and
> select one of possible 9 moves/decisions. The network will be trained
> by supplying previously recorded decisions of professional humans.
> Around 20 parameters are used from which 2-3 are of very higher
> importance but others are also compulsory to consider.
>
> 1- I am confused on what kind of problem is this? Fitting?
> Classification? Pattern Recognition?  or what? (I am refering to NNet PDF document)
>
> 2- If I need to just look at current state (I guess a static network),
> what kind of network is possibly suitable?
>
> 3- If I need to also consider previous states, then what?
>
> I appreciate your time and help.
>
Please do not sent separate posts to separate groups.

Send 1 post with a list of newsgroups in the "Newsgroups" category.

See my response in comp.ai.neural-nets.

Greg
From: Greg Heath on 20 Apr 2010 18:36
On Apr 20, 6:31 pm, Greg Heath <he...(a)alumni.brown.edu> wrote:
> On Apr 20, 12:44 pm, "Mac Saarman" <macdermat-
>
>
>
> remt...(a)remthisgmail.com> wrote:
> > Hello,
>
> > I am doing research in another field and as such I am not very good at
> > NNet. I hope people here can give me  small advise and insight on
> > this.
>
> > I need a trained NNet to look at current status of my system and
> > select one of possible 9 moves/decisions. The network will be trained
> > by supplying previously recorded decisions of professional humans.
> > Around 20 parameters are used from which 2-3 are of very higher
> > importance but others are also compulsory to consider.
>
> > 1- I am confused on what kind of problem is this? Fitting?
> > Classification? Pattern Recognition?  or what? (I am refering to NNet PDF document)
>
> > 2- If I need to just look at current state (I guess a static network),
> > what kind of network is possibly suitable?
>
> > 3- If I need to also consider previous states, then what?
>
> > I appreciate your time and help.
>
> Please do not sent separate posts to separate groups.
>
> Send 1 post with a list of newsgroups in the "Newsgroups" category.
>
> See my response in comp.ai.neural-nets.
>
> Greg- Hide quoted text -
>
> - Show quoted text -

You may also benefit from the CANN FAQ

ftp://ftp.sas.com/pub/neural/FAQ.html

hope this helps.

Greg
 | 
Pages: 1
Prev: merging ODE45 results
Next: Parallel Implementation of Code in Matlab