Re: [Bug-glpk] GLPK bugs report

[Andrew Makhorin]

> I tested the MIPLIB 2010 benchmark files on GLPK 4.51 (with gmp-5.1.2)
> compiled with gcc 4.8.1 on Fedora 18 64 bits.
>  
> The conclusions are :
>  
> The tests files come from MIPLIB 2010 from http://miplib.zib.de/
> --benchmark instances--
> The acc-tight5.mps result with 4.49 comes from
> http://plato.asu.edu/ftp/milpc_res1/benchmark.glpk.res,
> see also http://plato.asu.edu/ --Benchmarks for Optimization
> Software--

Thank you for your bug report.

Please note that most of miplib2010 instances are very hard while glpk
lacks of some important classes of cutting planes.

>  
> triptim1 without GLPK option wrongly declared 
> PROBLEM HAS NO FEASIBLE SOLUTION

Mps file for this mip (as well as for some other mips) has a
non-standard bound indicator LI in the BOUNDS section. Glpk considers
corresponding integer variables as binary ones that leads to incorrect
results.

>  
> acc-tight5 still working after 7200 sec., was resolved fast with 4.49

This may happen because of inefficiency of the clique cut generator. Now
the generator is essentially improved (will appear in a next release).

>  
> neos13 (as a lot of others) with "--proxy --tmlim 3600" generated
> Error: 1986 memory block(s) were lost
> Error detected in file glpapi21.c at line 1350

This is a non-crucial error detected by glpsol on exit. It happens
because of a memory leak bug in the proxy heuristic routine. Hope this
bug will be fixed in a next release.

>  
> glass4 with "--proxy --tmlim 3600" generated
> Assertion failed: piv1 != 0.0
> Error detected in file glpspx02.c at line 3000
> 
> mine-166-5 (as mine-90-10 and satellite1-25) with "--cuts" generated 
> Assertion failed: x >= -0.001
> Error detected in file glpios06.c at line 968
> Abandon (core dumped)
> 

These errors happen because of very large right-hand sides and column
bounds that make the instances badly conditioned. Sorry, glpk is not
perfect.

>  
> pw-myciel with "--cuts" generated
> PROBLEM HAS NO INTEGER FEASIBLE SOLUTION
> although it is resolved fast by CPLEX
> (http://plato.asu.edu/ftp/milpc_res12/benchmark.cplex.res) and others

See comments for triptim1 above.

>  
>  
> I hope this will help you improve GLPK and can send you, on request,
> any information or files from my test session.

Thank you very much.


Andrew Makhorin