Re: GSoC 2015: Optimization Package: Non-linear and constrained least squares lsqcurvefit, lsqlin, lsqnonlin

[Olaf Till]

On Fri, Feb 27, 2015 at 10:30:48PM +0000, Nir Krakauer wrote:
> On Fri, Feb 27, 2015 at 8:25 PM, Olaf Till <address@hidden> wrote:
> 
> > On Fri, Feb 27, 2015 at 03:01:14PM +0000, Nir Krakauer wrote:
> > > Also, what do you think about adding as a possible component of the
> > project
> > > to benchmark the existing and new optimization functions
> >
> > What do you mean by "existing and new"? I thought the project is now
> > decided not to create new algorithms?
> >
> 
> There would be new functions, like lsq* -- not necessarily new algorithms
> 
> Have you looked at "optim_problems.m" in the optim package?
> >
> 
>  It might be a good starting point for benchmarking. Looks like it has a
> dozen or so problems.
> 
> > How should tests for interface limitations be performed?
> 
> Basically I was just thinking of a case where the existing interface
> doesn't allow the form of a given test-suite problem to be posed in Octave,
> which is something that would be discovered by going through the suite.

I'd say each test-suite problem can be applied to each optimization
interface with a suitable "adapter" (wrapper) (of course the optimizer
must be applicable at all and be able to deal with the posed
constraints).

And if an optimization interface does not directly match a test-suite
interface, I wouldn't say this is necessarily a shortcoming of the
optimization interface.

So I still couldn't say how it should be possible to apply a test
suite to the properties of interfaces. Have I misunderstood something?

Olaf

-- 
public key id EAFE0591, e.g. on x-hkp://pool.sks-keyservers.net

signature.asc
Description: Digital signature