qp() in Octave 3.0.0 returns result egregiously violating input constraints

[A. Scottedward Hodel]

Ages ago I wrote (in C) an interior point method for QP problems.   
I've used it for some simulation work with control allocation.  If  
there's interest, I could try to port this for Octave use.   It's a  
first-implementation based on Boyd's book on Convex Programming.

A. Scottedward Hodel hodelas at mac.com
http://homepage.mac.com/hodelas/tar


On Apr 5, 2008, at 5:58 PM, Joshua Redstone wrote:
> I'm sorry I don't have an instance of the problem on a smaller data  
> set.
> RE debugging - I think it's good news it fails with such outrageous  
> values and that the first element of X is one such value.  If I  
> could get octave to build on my mac,  I'd be tempted to put  
> temporary checks in the __qp__() routine that fail if there are any  
> intermediate values greater than 1e+10.  Hopefully that would  
> narrow it down a bit.
> I may have some time to play with this a bit this weekend.
> Josh
>
> On Sat, Apr 5, 2008 at 3:17 PM, Ben Abbott <bpabbott at mac.com> wrote:
>
> On Apr 5, 2008, at 5:58 PM, Thomas Treichl wrote:
> Ben Abbott schrieb:
> I've built octave using Apple's lapack/blas (i.e. vecLib), so the  
> problem isn't directly related to the lapack/blas, correct?
>
> True. Then what about the difference between your PPC and our IAs?  
> Just an idea...
>
> I get exactly the same solution on my PPC (G4) using a build from  
> Feb 5.
>
>
>
> However, when I tried the release-3-0-x branch which I build  
> yesterday (and passed all tests), it fails (also built with vecLib).
> octave:8> result = 0.5*X'*H*X + X'*Q
> result =  5.4445e+58
> octave:9> tol = 10*eps;
> octave:10> all ( X > LB-tol & X < UB+tol)
> ans = 0
> octave:11> all ( A_LB-tol < A_IN*X & A_UB+tol > A_IN*X)
> ans = 0
>
> Hhhmmm...
>
> Thomas, when did you build last? The sources for one I'm running  
> were pulled from John's archive archive 2-3 days ago.
>
> I've just pulled a new snapshot from John's repository and rebuilt  
> everything. The same qr() results than before. However, somehow I  
> think I still don't know what's going on. Is it possible to reduce  
> that kind of problem to a much easier exercise (that also looks  
> like a fail)?
>
> I agree this would make isolating the problem easier.  As I have  
> not studied qr() I don't know if I can be successful ... but I'll  
> kill a bit of time on it this weekend. If I'm successful I'll post  
> to post immediately.
>
> Ben
>
> _______________________________________________
> Bug-octave mailing list
> Bug-octave at octave.org
> https://www.cae.wisc.edu/mailman/listinfo/bug-octave

-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://www.cae.wisc.edu/pipermail/bug-octave/attachments/20080405/91d9f5d3/attachment-0001.html