negative value for non-negative variable and optimal with unscaled infeasibilities

[zm guan]

Hi,
 
Several minutes ago I sent a question on getting negative value for a non-negative variable. I have just found that the solver message says
 
CPLEX 12.2.0.0: optimal with unscaled infeasibilities; objective 18372739870
7909 dual simplex iterations (2041 in phase I) on the dual problem
solve_result_num = 110
 
I wonder what unscaled infeasibilities means? Also, I have made an if-then statement to terminate the algorithm (break the loop) if the solve_result_num is larger than 99, but the algorithm doesn't stop. I wonder why? Thank you.
 
Regards,
Ziming Guan (John)

[Robert Fourer]

CPLEX considers a solution to be feasible if the amount of infeasibility is less than a certain feasibility tolerance, by default 10^-6.  Thus a nonnegative variable can be slightly negative in the solution that CPLEX returns, as long as its magnitude is less than this tolerance.  For nonzero bounds I believe CPLEX uses a relative tolerance, so that the variable can be outside of its bound by at most 10^-6 times the bound value.

 

CPLEX also by default scales the problem before solving it, then unscales the solution to give a solution to the unscaled problem.  It can happen that the solution to the scaled problem is within all feasibility tolerances, but the solution to the unscaled problem is outside some tolerances -- in which case you get the "unscaled infeasibilities" message.  This most often happens when there is a large range of coefficient magnitudes, or there are very large coefficients, which you may be able to fix by more carefully formulating your problem.  (The objective of 18,372,739,870 suggests that something like this may be the cause.)

 

Bob Fourer

4...@ampl.com

[zm guan]

Thank you, and also for the reply on the previous question.
 
Regards,
Ziming Guan (John)
 
 

---

From: 4...@ampl.com
To: am...@googlegroups.com
Subject: RE: [AMPL 5612] negative value for non-negative variable and optimal with unscaled infeasibilities
Date: Fri, 23 Mar 2012 00:14:47 -0500

--
You received this message because you are subscribed to the Google Groups "AMPL Modeling Language" group.
To post to this group, send email to am...@googlegroups.com.
To unsubscribe from this group, send email to ampl+uns...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/ampl?hl=en.