[Ipopt] [ProbableSpam] Re: Why sometimes can solve, but sometimes can't solve with the same problem?

[Andreas Waechter]

Hi Edson,

Yes, your observation is correct.  If you reorder the constraints, 
floating point values appear in the numerical computations in a different 
order; e.g., the linear system is given to the linear solver in a permuted 
way, and due to round-off the result returned by the linear solver will be 
different.  The differences are usually first rather small (last few 
significant digits), but over time can accumulate to a considerable 
deviation (depending on how ill-conditioned the problem is) and even make 
a difference between solving the problem and failure.

However, if the input to Ipopt is identical (same values and order of all 
variables, constraints, order and values of non-zero elements of 
matrices), the algorithm should perform identially in two different runs.

Regards

Andreas

On Fri, 26 Mar 2010, Edson Cordeiro do Valle wrote:

>        Hello All
>           Just a little contribution from my experience.
> As stated some time ago in this list, the order which constraints are
> being written can produce this behavior. I mean, if you have a set of 20
> constraints and write them in a different order, this behavior can
> happen, at least with Mumps solver. I don't know if it can happen with
> MA57.
> This happened to me some time ago where the software we were working
> with. In that software (RecoVR - http://www.vrtech.com.br/en_us/recovr)
> the equality constraints are written according to the order of a block
> diagram drawn by the user.
> I don't know if its your case, but it may give you a clue.
> Regards
>
>                                                  Edson C. do Valle
>                                                edsoncv at enq.ufrgs.br
>                                                   Skype: edson.cv
>
>
>>
>> Message: 4
>> Date: Fri, 26 Mar 2010 10:37:00 -0400 (EDT)
>> From: Andreas Waechter <andreasw at watson.ibm.com>
>> Subject: Re: [Ipopt] Why sometimes can solve, but sometimes can't
>> 	solve with  the same problem?
>> To: Lewis I <lewis369lewis at yahoo.com>
>> Cc: ipopt at list.coin-or.org
>> Message-ID: <Pine.A41.4.61.1003261035200.430144 at bigopt.watson.ibm.com>
>> Content-Type: TEXT/PLAIN; charset=US-ASCII; format=flowed
>>
>> Hi Lewis,
>>
>> The Ipopt code is deterministic.  Potentially, the linear solver is not,
>> but I would think that MA57 is.  Maybe your BLAS library is not
>> deterministic?  Another thing to check is if there is uninitialized memory
>> usage, e.g., in your function evaluations, which might lead to different
>> outcomes (this could easily be checked with a memory checker like
>> valgrind).
>>
>> Andreas
>>
>> On Thu, 25 Mar 2010, Lewis I wrote:
>>
>>
>>> Hi Frank,
>>>
>>> I used the same options, same starting points and same obj function, and all other thing are the same too. (using  linear system solver ma57).
>>> (and  used the hessian approximation). But sometime ipopt can find the
>>> optimal point, but sometime can't.
>>>
>>> I found that ipopt provide different number of iteraitons for the same
>>> problem.
>>> So I guess there are "something" random that existed in the ipopt
>>> method.
>>> Does it relate to my question?
>>> Thx
>>>
>>> Lewis
>>>
>>>
>>>
>>> ________________________________
>>> From: Frank E. Curtis <frank.e.curtis at gmail.com>
>>> To: Lewis I <lewis369lewis at yahoo.com>
>>> Cc: ipopt at list.coin-or.org
>>> Sent: Fri, March 26, 2010 5:08:23 AM
>>> Subject: Re: [Ipopt] Why sometimes can solve, but sometimes can't solve with  the same problem?
>>>
>>>
>>> Hi Lewis,
>>>
>>> I suspect you'll need to answer a few questions before anyone can help you figure out what's really going on here.
>>>
>>> 1.) Naturally, if, say, your function evaluations involve random quantities, than the results can differ from run to run.  (I'm guessing that's not your case, though.)
>>> 2.) If you are changing any of the options, then the results can certainly vary, even for the same problem.
>>> 3.) If you are running different versions of Ipopt, or if you're running even the same version on different platforms, things can change.
>>>
>>> If it's not any of these basic things, we'd probably need to know more about your problem, the options you're using (including the linear system solver), etc.
>>>
>>> On Thu, Mar 25, 2010 at 5:00 PM, Lewis I <lewis369lewis at yahoo.com> wrote:
>>>
>>>
>>>> Dear all,
>>>>
>>>>
>>>>> Why ipopt sometimes can solve, but sometimes can't solve with the same problem?
>>>>> And the number of iteration are different too.why?
>>>>>
>>>>> Regards,
>>>>> Lewis
>>>>>
>>>>
>>>>
>>>>
>>>>
>>>>> _______________________________________________
>>>>> Ipopt mailing list
>>>>>
>>>> Ipopt at list.coin-or.org
>>>> http://list.coin-or.org/mailman/listinfo/ipopt
>>>>
>>>>
>>>
>>>
>>>
>>
>>
>>
>> ------------------------------
>>
>> _______________________________________________
>> Ipopt mailing list
>> Ipopt at list.coin-or.org
>> http://list.coin-or.org/mailman/listinfo/ipopt
>>
>>
>> End of Ipopt Digest, Vol 63, Issue 11
>> *************************************
>>
>
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt
>