Saturday, March 2, 2013

Option optcr

For MIQCPs Mosek seems to ignore the relative gap provided by GAMS (option optcr).

With either option opctr=0.1 (default) or optcr=0 (solve to optimality) I see:

An optimal solution satisfying the relative gap tolerance of 1.00e-002(%) has been located.
The relative gap is 9.36e-003(%).

Strangely enough for linear MIPs the optcr seems to work fine.  Well, almost. It does not seem to like a real zero gap.

So probably we have:


  OPTCR Mosek gap tolerance  
MIP 0 1.00e-002(%) almost ok
MIP 0.1 1.00e+001(%) ok
MIQCP 0 1.00e-002(%) almost ok
MIQCP 0.1 1.00e-002(%) error

Most solvers produce a report in the listing file:

MIQP  Solution:      16305.625000    (12 iterations, 0 nodes)
Final Solve:         16301.875000    (0 iterations)

Best possible:       16222.656250
Absolute gap:           82.968750
Relative gap:            0.005088

which is helpful understanding the quality of the solution. Unfortunately Mosek does not do this (it writes nothing to the listing file except for some advertisement).