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

[Olaf Till]

On Tue, Feb 24, 2015 at 10:28:57PM +0000, Nir Krakauer wrote:
> Olaf,
> 
> Mostly on general principles -- Octave aims to run code written for Matlab,
> and optimization is a capability used across many application areas. If the
> underlying routines are already there, it could be just a matter of
> providing wrappers.

In optimization (and maybe also in other fields, e.g. numeric solving
of differential equations) it is practically impossible always to
produce the same results as Matlab. Even if Matlab provides references
for its applied optimization algorithms, the details of the
algorithm-implementations have impact on the results and are too
individual to be reproduced without using their code.

So making the interface identical to Matlabs interface would only
yield a superficial similarity. In particular there will probably
always be certain optimization problems which Matlabs optimization
algorithms can solve well enough, but the "respective" Octave
algorithms can't (and vice versa, of course).

In fact I'd think that people earnestly involved with optimization are
aware of this. They probably also are aware of the potential problems
of optimization and are prepared to spend some effort coping with
them. Compared to this effort, the effort of getting familiar with an
interface slightly different to Matlabs is minor. Correspondingly,
AFAICR we never had a complaint at the mailing lists that the
optimization interface is not identical to Matlabs from people seeking
help with an actual optimization problem.

OTOH, people unaware of the potential difficulties in optimization may
be tempted to expect that an interface identical to Matlabs should
produce exactly the same result as in Matlab.

Taken together I'd think that in optimization, since we have a
slightly different interface which (hopefully) offers some advantages,
it's reasonable to expect users to use it instead of Matlabs.

To address the issue of visibility by former Matlab users, we could
have e.g. an lsqcurvefit function which just throws an error with a
text informing the user about the existence of our interfaces.

Olaf

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

signature.asc
Description: Digital signature