Re: [OctDev] quick release

[Michael C. <mic...@ua...>]

On Friday 03 September 2004 16:09, David Bateman wrote:
> If this causes a core-dump, the bulletproofing should be in bfgsmin itself
> rather than in the objective function. Nothing the user can do should cause
> a core-dump, otherwise it is a bug..
>
> D.
>
> According to Michael Creel <mic...@ua...> (on 09/03/04):
> > On Friday 03 September 2004 14:24, Teemu Ikonen wrote:
> > > bfgsmin consistently makes octave dump core on at least one fitting
> > > problem I have. Unfortunately I don't have the time to send a more
> > > detailed bug report just now.
> > >
> > > Teemu
> >

So far I haven't seen use of bfgsmin or the other routines that are documented 
as being intended for end users cause core dumps, but that's not to say that 
it can't happen. It's true that Teemu's example for calling newtonstep 
without arguments dumps core, but as the source code says, newtonstep is not 
intended for use by end users. For the moment, I'd point people to the 
documentation in main/optim/doc, but in the medium term I'll try to add 
proper argument checking to all functions.

The reason bsgsmin doesn't do bulletproofing internally is because that leads 
to users minimizing a different function than they think they are, and the 
end point the algorithm converges to may depend upon the specific 
implementation. Difficulties in minimization can lead to learning about where 
the problematic spots are in the function. Different users may like to 
bulletproof in different ways. The documentation does suggest how 
bulletproofing might be done, but I'm reluctant to use a one-size-fits-all 
approach internally to the minimization algorithms.

Suggestions are welcome, however.

Michael