[Bonmin] Assertion in OsiSimpleInteger::createBranch()

[David Wilkinson]

Hi Stefan:

In Bonmin 1.8.6, when I activate ma57 solver, I also see a Visual C++ 
warning (signed/unsigned mismatch) from the code

if( wd_lfact_ > std::numeric_limits<size_t>::max() / sizeof(double) ) {

at Ipopt\src\Algorithm\LinearSolvers\IpMa57TSolverInterface.cpp line 657

This signed/unsigned mismatch warning always makes me very nervous 
(especially in 64-bit builds). Personally I would do something like

assert(wd_lfact_ >= 0);
if ( (size_t)wd_lfact_ > std::numeric_limits<size_t>::max() / 
sizeof(double) ) {
// ...
}

Or better, re-write the code so that a cast (implicit or explicit) is 
not necessary.

Thanks,

David Wilkinson

=========================


On 20.06.2017 12:17, Stefan Vigerske wrote:
> Hi,
> 
> On 06/20/2017 05:36 PM, David Wilkinson wrote:
>> Hi Stefan:
>> 
>> I will certainly upgrade to Bonmin 1.8.6, as I see it has all the 
>> changes regarding FORTRAN_INTEGER_TYPE and NO_CATCH_ALL that you and I 
>> have been discussing. Thanks for that! But are there other changes 
>> relative to Bonmin 1,8.4?
> 
> Bonmin 1.8.4 is based on a Cbc that is similar old as the Bonmin 1.8.4
> release. Bonmin 1.8.5/1.8.6 uses a current Cbc release, which should
> have less bugs.
> 
>> What do you mean by "fixed integer variable"? Do you mean one where 
>> the user has set the upper and lower bounds to be the same? That I 
>> certainly have not done.
> 
> It's more likely that Bonmin fixed some variable during branching.
> It's not wrong that variables get fixed (locally) during B&B, or that
> one has fixed variables in the problem, but attempting to branch on
> fixed variables is wrong.
> 
>> Until recently we were using (the very ancient) Bonmin 1.3.2, and 
>> interestingly this problem did not arise there.
> 
> Yes, that were the times when Pierre, the Bonmin developer, was still 
> active :).
> 
> Stefan
> 
>> 
>> David Wilkinson
>> 
>> ============================
>> 
>> On 20.06.2017 09:29, Stefan Vigerske wrote:
>>> Hi,
>>> 
>>> it sounds as if Bonmin tries to branch on a fixed integer variable.
>>> You might want to trace who created this OsiBranchingInformation and
>>> why this has happened.
>>> 
>>> Also consider upgrading to Bonmin 1.8.5 or 1.8.6 first. Bonmin 1.8.4
>>> is 21 months old.
>>> 
>>> Stefan
>>> 
>>> On 06/20/2017 03:16 PM, David Wilkinson wrote:
>>>> This is a Bonmin question, but I haven't seen any activity in the 
>>>> Bonmin group for a while, so I am sending to the Ipopt group also.
>>>> 
>>>> I am using Bonmin 1.8.4 with "B-BB" algorithm to solve a problem 
>>>> with two integer variables, both with lower bound 0 and upper bound 
>>>> 3, with an initial guess of (1,1). I am seeing an assertion error in 
>>>> OsiSimpleInteger::createBranch() around line 626 of 
>>>> OsiBranchingObject.cpp. The purpose of the assertion is to check 
>>>> that the upper bound for the variable under consideration is greater 
>>>> than the lower bound. When I hit the assertion, the value of 
>>>> columnNumber_ is 0, and both info->lower_[0] and info->upper_[0] are 
>>>> 0.0, so the assertion is correctly triggered.
>>>> 
>>>> Since assertions are not supposed to happen, I initially suspected 
>>>> user (me) error in setting up the problem, but I do not think so, 
>>>> because our driver program repeatedly requests optimizations of a 
>>>> similar type (with different objective and constraint functions, but 
>>>> the same bounds), and several such optimizations have been done 
>>>> correctly before this one. It may be that these optimizations do not 
>>>> fully satisfy the Bonmin convexity requirements, but surely this 
>>>> should not cause an assertion in the code? Our program will tolerate 
>>>> cases where Bonmin fails and emits an exception, and indeed if I 
>>>> throw an exception in OsiSimpleInteger::createBranch() instead of 
>>>> the assertion, our program continues with other successful Bonmin 
>>>> optimizations.
>>>> 
>>>> Any ideas?
>>>> 
>>>> Thanks,
>>>> 
>>>> David Wilkinson
>>>> _______________________________________________
>>>> Bonmin mailing list
>>>> Bonmin at list.coin-or.org
>>>> https://urldefense.proofpoint.com/v2/url?u=https-3A__list.coin-2Dor.org_mailman_listinfo_bonmin&d=DwICAg&c=Ngd-ta5yRYsqeUsEDgxhcqsYYY1Xs5ogLxWPA_2Wlc4&r=BRcuJnQr5NAzU29t80hk2rsLc4vrlRySBDabuq0O1ZI&m=fSexfIOcTv9KTMRQVZe65MHJusguSwsHMzfmO1Q4DoE&s=vOOwrD4g3FsVpmxBP1Zdx0SWV-pW18OHMbJUxShxkVg&e=
>>