[Cgl] Re: Numerical precision in cut generators

[Cynthia Phillips]

Good idea to set a default for silence :)

I agree that standardization is a good idea, though I don¹t have a strong
opinion about what the value should be.  This value for EPS seems fine.  I
will ask some of my more numerically savvy collaborators for their opinions.

I just reread Francois¹ original message (sorry, I didn¹t join the mailing
list till September).  I think it makes sense to have a set of computational
precision/tolerance values defined and set to defaults in the base class.
Individual cut generators can derive other tolerances from these (if too
generator-specific to belong in the base class) and combinatorial ones can
ignore them.

I¹d like to suggest that while we are considering standardized numerical
precision, that we also make these values parameters within the cut
generators.  That is, for the CGs that don¹t already do this, replace all
hard-coded tolerance/precision values with parameters (derived from a set in
the base class if appropriate) and make those parameters adjustable both at
the base class level and at the CG level if reasonable.  CglRedSplit does
this through its setEPS method, for example.

Thanks!

    - Cindy


On 9/20/06 7:45 AM, "Robin Lougee-Heimer" <robinlh at us.ibm.com> wrote:

> 
> In July,  Francois proposed using standard variables for handling numerican
> precision, e.g., 
>    EPS: Precision of double computations. Default value: 1e-7.
> 
> I agree - we should have a standard. It makes a lot of sense.
> 
> The list of standards he proposed was based on CglRedSplit, and CglGomory.  It
> seemed a bit of overkill to me, but if  they were necessary/useful for Gomory
> and RedSplit then it seems a reasonable list to begin.
> 
> ----> Does anyone else have any thoughts/comments?
> 
> Silence will be taken as thunderous approval, and the next step would be
> asking each CG-maintainer to update their CG.
> 
> Robin 
> 
> ------------------------------------------------------------------------------
> ----
> Robin Lougee-Heimer
> IBM TJ Watson Research Center
> 1101 Kitchawan Road, Yorktown Heights, NY 10598
> ph: 914-945-3032   fax: 914-945-3434
> robinlh at us.ibm.com
> http://www.coin-or.org
> 
> 
> 
> _______________________________________________
> Cgl mailing list
> Cgl at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/cgl


-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://list.coin-or.org/pipermail/cgl/attachments/20060920/bf93699d/attachment.html