This is the mail archive of the
gsl-discuss@sourceware.cygnus.com
mailing list for the GSL project.
Re: multidimensional optimization
Cool, that sounds great. Diffs against the current CVS tree would be
the cleanest method. If you could put them on a web page and send the
url that will avoid overloading the list. Any problems/questions - let
me know. Thanks.
Fabrice Rossi writes:
> Hi.
> I've finished a preliminary, buggy, alpha (etc.) version of the
> multidimensional optimization part of GSL I've proposed to code
> last month. I'm not at all asking for inclusion, but for
> review. The question is how should I make the code available? I've
> added a new directory (multmin) which can be released as a tar
> file, but I've also a patch to the min directory (basically a
> bracketing algorithm).
> The implementation follows the general idea discussed in the
> list. Basically, I've targeted descent algorithms, i.e., algorithms
> in which one performs a one dimensional minimization along a
> direction given by the algorithm (the gradient for steepest descent
> for instance). As suggested by Brian Gough, the line search (one
> dimensional minimization) is not at all embedded into the global
> algorithm. In fact, the algorithm produces a direction and updates
> a structure that allows to view the multidimensional function as a
> one dimensional one. The pseudo loop is therefore:
> init do
> compute next direction (d)
> find an initial bracketing for line search
> init 1d minimization
> do
> loop for 1d min
> perform the step (i.e. replace x by x+ed, where e is the
> minimum of g(e)=f(x+ed)
> check for convergence
> I've tried to follow the structure of the multiroots package check
> in by Brian. As a side effect, I've used gsl_vector everywhere.
> Fabrice Rossi
>