question on GSL development

question on GSL development

[Francisco Yepes Barrera]

I read in the GSL website that the development of the library is
considered complete, fundamentally. I'm thinking in the development of
functions related to neural networks (backpropagation) and genetic
algorithms routines (genetic operators as mutation, crossover, and so
on). Related to these fields there are also statistical functionality
as PCA (Parallel Components Analysis), necessary for example for
neural networks, but with an interest also as standalone functions.

My question is: are these topics beyond the scope of the library at
the actual level of maturity? What is the policy of GSL regarding the
extension of the library on new fields?

Thanks.

Paco

Re: question on GSL development

[Jordi Burguet Castell]

Hi Paco,

On Tuesday 20 September 2005 10:52, Francisco Yepes Barrera wrote:
> I read in the GSL website that the development of the library is
> considered complete, fundamentally. I'm thinking in the development of
> functions related to neural networks (backpropagation) and genetic
> algorithms routines (genetic operators as mutation, crossover, and so
> on). Related to these fields there are also statistical functionality
> as PCA (Parallel Components Analysis), necessary for example for
> neural networks, but with an interest also as standalone functions.
>
> My question is: are these topics beyond the scope of the library at
> the actual level of maturity? What is the policy of GSL regarding the
> extension of the library on new fields?

The policy is to first create a package that builds outside of GSL. You can 
find the guidelines at http://sources.redhat.com/gsl/devel.html (under the 
section "Packages").

Cheers,
jordi

Re: question on GSL development

[James Bergstra]

Hi, I'm a grad student at U of Montreal studying machine learning.

I understand that GSL has moved toward a system of adding *extensions*
to the base functionality that is complete.  I am working on an
extension to facilitate building and training neural networks (among
other things).

I have some code in cvs at savannah under the project name "Montreal
Scientific Library" for designing neural networks (differentiable
compositions of functions) in order to train them by gradient descent.
In the next few days I was going to commit a little demo of how to role
your own network, train and test it.

Your comments on my approach would be greatly appreciated!

http://www.nongnu.org/libmsl/


I was thinking about an approach for coding genetic algorithms, and I
concluded(IMHO!) that the cleanest way to provide generic tools for
solving GA problems and other problems in combinatorial optimization
would be to establish a framework for optimizing a function on a
*tensor*, the way the gsl_multimin_* routines optimize a function on a
vector space.

Gibbs-sampling would be one algorithm for this, a GA with given
recombination policies would be another, dynamic programming another, 
and gradient-descent algorithms could be used too, when the values of
the tensor elements are highly correlated in neighbourhoods.

Maybe you, or someone else would like to comment on these ideas,
especially if you have some background in combinatorial optimization :)

James

On Tue, Sep 20, 2005 at 10:52:02AM +0200, Francisco Yepes Barrera wrote:
> I read in the GSL website that the development of the library is
> considered complete, fundamentally. I'm thinking in the development of
> functions related to neural networks (backpropagation) and genetic
> algorithms routines (genetic operators as mutation, crossover, and so
> on). Related to these fields there are also statistical functionality
> as PCA (Parallel Components Analysis), necessary for example for
> neural networks, but with an interest also as standalone functions.
> 
> My question is: are these topics beyond the scope of the library at
> the actual level of maturity? What is the policy of GSL regarding the
> extension of the library on new fields?
> 
> Thanks.
> 
> Paco

-- 
james bergstra
http://www-etud.iro.umontreal.ca/~bergstrj

Re: question on GSL development

[Francisco Yepes Barrera]

James Bergstra <james.bergstra@umontreal.ca> writes:

> I am working on an extension to facilitate building and training
> neural networks (among other things).
> 
> I have some code in cvs at savannah under the project name "Montreal
> Scientific Library" for designing neural networks
> 
> Your comments on my approach would be greatly appreciated!

Thanks, James. I've downloaded the code.

> I was thinking about an approach for coding genetic algorithms, and I
> concluded(IMHO!) that the cleanest way to provide generic tools for
> solving GA problems and other problems in combinatorial optimization
> would be to establish a framework for optimizing a function on a
> *tensor*, the way the gsl_multimin_* routines optimize a function on a
> vector space.
> 
> Gibbs-sampling would be one algorithm for this, a GA with given
> recombination policies would be another, dynamic programming another, 
> and gradient-descent algorithms could be used too, when the values of
> the tensor elements are highly correlated in neighbourhoods.
> 
> Maybe you, or someone else would like to comment on these ideas,
> especially if you have some background in combinatorial optimization :)
> 
> James

Can you give some more detail on this approach? I've worked
fundamentally on steady-state GA, used to solve combinatorial
problems in computational chemistry.

Paco