Originally posted by: EXCT_RALF_GOLLMER

Hi again,

here another example showing a different behaviour of 12.6.1.0. vs. 12.6.0.1 - resulting in different integer optimal solution values.

May the changes have led to a grown sensitivity against condition numbers?

Attached you find a MILP, a parameter file, a cplex.input with interactive optimizer commands and two logs exhibiting a difference between 12.6.1.0 and 12.6.0.1.

While the previous version produces the same optimal value with and without non-default parameters, the current version does not.

Best regards

Ralf Gollmer, University of Duisburg-Essen

It looks like you may have bumped into a known issue. Does it help to disable the FP heuristic, i.e., set CPX_PARAM_FPHEUR to -1?

Originally posted by: EXCT_RALF_GOLLMER

You are right, just forbidding the feasibility pump resolves the problem.

Btw., when I looked at the fixcentral today I found the selection 12.6.2 for the installed version. Does that indicate it will shortly be released? 

Will the issue causing the behaviour in my post be addressed in that release?

Best regards and many thanks for looking into the problem

Ralf Gollmer

Version 12.6.2 was just announced, see for example here and here. The issue with the feasibility pump heuristic is fixed in 12.6.2.

Originally posted by: Mark L. Stone

Daneil,

Per https://www.ibm.com/developerworks/community/blogs/jfp/entry/CPLEX_Optimization_Studio_12_6_2?lang=en

Our CPLEX Optimizer performance has been significantly improved for quadratic problems of the following types:

• Second Order Cone problems (SOCP)
• Mixed Integer SOCP (MISOCP)
• Non Convex Quadratic problems (QP)

Does the improvement in Non Convex QPs pertain to solutiontarget = 2 (local optimization) , solutiontarget = 3 (global optimization) , or both?

Thanks.

Originally posted by: PierreBonami

Hi Mark,

The notable improvements are for the global solver.

Best,

Pierre

Originally posted by: Mark L. Stone

Hi Pierre,

Do you have any figures on the improvement (run time, memory usage)  in the global Non Convex QP optimizer for 12.6.2 vs. 12.6.1 for various problems?  Or is there perhaps a white paper or briefing on the algorithm improvements, along the lines of  what you produced when the global QP optimizer was first introduced in 12.6.0?

Thanks.

Originally posted by: PierreBonami

Hi Mark,

Thanks for your interest. This was just announced yesterday and is not yet available. At this point we don't have such material.

If you happen to be next week in Montreal for CORS-INFORMS, I will give a tutorial on Wednesday where I will talk about it (among other things). If not I will certainly post the slides somewhere in any case.

There might be a paper but this will take longer.

Best regards,

Pierre

Originally posted by: Mark L. Stone

Pierre, I won't be in Montreal.  I'll wait for your slides and paper as they become available.  Thanks.