[Coin-lpsolver] behaviour of OsiMaxNumIteration with Clp
Kish Shen
kish.shen at crosscoreop.com
Tue Jul 11 20:54:13 EDT 2006
Hi,
I have set the iteration limit for Clp Simplex solver with Osi's
OsiMaxNumIteration parameter (through the OsiClpSolverInterface).
Looking at the messages printed by Clp, it appears that after reaching the
limit, the Simplex solve is stopped, but then restarted with a resolve (all
within a call to initialSolve()). For example:
Clp0006I 2395 Obj 2.08958e+06 Primal inf 21.6669 (19) Dual inf 1.9129e+12
(2125)
Clp0006I 2544 Obj 2.04513e+06 Dual inf 595108 (2277)
Clp0006I 2696 Obj 1.98142e+06 Dual inf 212071 (2052)
Clp0006I 2841 Obj 1.93202e+06 Dual inf 3.0126e+06 (2128)
Clp0006I 2984 Obj 1.84925e+06 Dual inf 4.31107e+07 (2571)
Clp0003I Stopped - objective value 1.83099e+06
Coin0505I Presolved problem not optimal, resolve after postsolve
Coin0511I After Postsolve, objective 1.83099e+06, infeasibilities - dual
8.89174e+6 (2881), primal 0 (0)
Clp0006I 0 Obj 1.83099e+06 Dual inf 1.5021e+07 (2718)
Clp0029I End of values pass after 10 iterations
Clp0006I 10 Obj 1.82246e+06 Dual inf 1.39314e+07 (2975)
Clp0006I 227 Obj 1.71404e+06 Dual inf 7.8932e+06 (2834)
Clp0006I 441 Obj 1.60609e+06 Dual inf 1.78273e+06 (2034)
Clp0006I 657 Obj 1.4714e+06 Dual inf 366286 (1927)
Clp0006I 887 Obj 1.35025e+06 Dual inf 1.36657e+06 (2135)
Clp0006I 1141 Obj 1.29188e+06 Dual inf 3.15447e+06 (2690)
Clp0006I 1448 Obj 1.14856e+06 Dual inf 2.85256e+06 (2792)
Clp0006I 1791 Obj 1.06044e+06 Dual inf 134040 (2050)
Clp0006I 2191 Obj 1.00322e+06 Dual inf 25244.5 (1590)
Clp0006I 2591 Obj 996369 Dual inf 185603 (2050)
Clp0006I 2991 Obj 988731 Dual inf 3707.67 (554)
Clp0003I Stopped - objective value 988703
In this case, the iteration limit was set to 3000, and Primal Simplex was used
to solve the problem. In this case, the solve finally stop after reaching the
iteration limit for a second time.
Is this expected behaviour? It was a bit surprising to me, but I am not that
familiar with the Simplex method.
Cheers,
Kish Shen
More information about the Clp
mailing list