Forum Discussion
Additional execution after crossover
I am running a problem with the Newton Barrier method using crossover as I need to compute dual variables.
However, after crossover completes in approx 50 seconds it starts another process that takes more than 1000 seconds.
Is this additional process necessary? If not, how could I deactivate it?
Thank you
--------------------------------------------------------------------------------
FICO Xpress v8.8.0, Hyper, solve started 15:29:21, Dec 31, 2022
Heap usage: 87MB (peak 87MB, 1MB system)
Minimizing LP
Original problem has:
331237 rows 139361 cols 806241 elements
Presolved problem has:
163217 rows 89334 cols 500625 elements
Presolve finished in 0 seconds
Heap usage: 105MB (peak 255MB, 1MB system)
Coefficient range original solved
Coefficients [min,max] : [ 1.81e-09, 4.00e+07] / [ 2.38e-09, 2.00e+00]
RHS and bounds [min,max] : [ 2.28e-12, 7.93e+12] / [ 7.25e-10, 9.68e+08]
Objective [min,max] : [ 1.00e+00, 1.00e+00] / [ 3.36e+07, 3.36e+07]
Building dual
Barrier cache sizes : L1=48K L2=12288K
Using AVX support
Cores per CPU (CORESPERCPU): 8
Barrier starts after 0 seconds, with L2 cache 1536K
Matrix ordering - Dense cols.: 325 NZ(L): 2480835 Flops: 211105568
Its P.inf D.inf U.inf Primal obj. Dual obj. Compl.
0 9.63e+06 7.74e+08 0.00e+00 -1.4790928e+16 -0.0000000e+00 2.2e+17
1 9.63e+06 7.74e+08 0.00e+00 -1.5130937e+16 5.4589503e+10 2.2e+17
2 9.53e+06 7.50e+08 0.00e+00 -1.5283823e+16 4.1732226e+11 2.1e+17
3 9.53e+06 7.50e+08 0.00e+00 -1.4528070e+16 4.5861312e+11 2.1e+17
...
26 2.56e+02 4.47e-01 0.00e+00 5.1999331e+15 5.4386219e+12 7.2e+11
27 2.58e+07 4.47e-01 0.00e+00 3.3636207e+20 5.4386819e+12 1.3e+16
Barrier method finished in 7 seconds
Crossover starts
Its Obj Value S Ninf Nneg Sum Inf Time
97921 5.43214E+12 P 0 0 .000000 43
0 5.43214E+12 N 0 0 .000000 43
2000 5.43214E+12 D 0 64554 .000000 43
Crossover successful
Objective function value: 5.43214E+12 time: 36
0 5.43214E+12 P 0 64554 .000000 43
1244 5.43224E+12 P 0 65798 .000000 47
2518 5.43229E+12 P 0 62547 .000000 50
...
156062 5.43251E+12 P 0 29382 .000000 1146
1 Reply
Hi Diego,
the issue seems to be that barrier did not actually manage to solve the problem to optimality but ran into numerical issues and terminated with large infeasibilities and complementarity. In this case also crossover will only find a basic solution with no-worse objective and infeasibilities than that, and the final cleanup of dual infeasibilities has to do the actual work to solve the problem.
If you can provide an mps file (either here or also privately via email) we could have a look why barrier struggles. You could also try BARALG=2 which tends to be more stable when encountering numerical difficulties.
Best,
Tristan
