qp() in octave-3.0.0 gets stuck on wrong answer to convex problem

[Joshua Redstone]

I put a few printfs in the code and it looks like, after 18 iterations, it
repeatedly enters the clause on line 443 with the comment    // There are no
blocking constraints.  Adding the whole step.
In this clause, alpha = 1, and p alternates in value between +k and -k where
k is a vector:
-0.000997019 -0.00297671 0.00197098 0.00200275 0.0335145 0.0424274 0.0343524
3.8161e-18 -6.72938e-18 0.0362161 0.0458195 0.0291282 0.00380536 0.0184418
0.0161607 -0.0338612 -0 -0.14143 -1.1093e-17 -0 -2.18072e-18 -0 -0
-0.000252608 -0 -0.0215625 -5.90882e-18 -0 -0.0543663 -0.034514 -0.0082697
-0 -0 -0.0727281 1.74693e-17 -2.13228e-18 -0.071406 0.0612498 0.0454769
1.58984e-17 1.37872e-17 0.0720697 2.35337e-17 0.166084 -4.04256e-17
0.0261761 3.58841e-18 -4.63691e-18 -0 -0.00228516 0.0265861 -0.0443588
1.21727e-17 -0.0594975 0.0186493 0.00538351 1.13094e-17 -0 -0 -0 -0 0.166227
2.79509e-17 -0 -1.92503e-17 -0 -0 0.165617 -2.82539e-18 0.132712 0.132843 -0
0.0804525 0.0710416 -0.000252608 -0.89828 -0 0.0126304

So it looks like it is bouncing back and forth between two points, which is
consistent with the observation that it stops making progress.
Also, when it gets stuck as above, p is computed using the 'negative
curvature' section of code earlier in the function.
Though I've no idea why it is doing this....
Hope this helps,
Josh

On Tue, Jun 10, 2008 at 7:35 AM, Joshua Redstone <redstone at gmail.com> wrote:

> Just checked, and the file I sent reproduces the problem on my Ubuntu 8.04
> machine as well.
> Josh
>
>
> On Fri, Jun 6, 2008 at 4:03 PM, Joshua Redstone <redstone at gmail.com>
> wrote:
>
>> Hi Gabriele, Thanks for looking into this!
>> Yes, you're right, thanks for the clarification - I meant 'converge' in
>> the sense that the value of OBJ does
>> not seem to change as the maximum number of iterations increases.
>> If the data file I attached to the bug post doesn't work for you, I'll be
>> back in front of a linux box next Tuesday and can
>> generate a more linux-friendly example for you.
>> Thanks,
>> Josh
>>
>> On Fri, Jun 6, 2008 at 2:43 PM, Gabriele Pannocchia <
>> g.pannocchia at ing.unipi.it> wrote:
>>
>>> Hi all,
>>>
>>> I will look into Josh's example ASAP to see if I can reproduce the
>>> problem and possibly understand why that happens.
>>>
>>> > I just tried octave 3.0.1 (osx) - the behavior is identical (i.e., it
>>> > converges to the wrong answer).
>>>
>>> Just a clarification: if you get INFO=3, it means that the maximum
>>> number of iterations has been reached. Thus, it is not correct to say
>>> that it converges to the wrong answer.
>>>
>>> Gabriele
>>>
>>>
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://www.cae.wisc.edu/pipermail/bug-octave/attachments/20080615/4f815c18/attachment.html