Models: 38 convex MINLPs from the GAMS MINLPLib
Note that this is a quite small testset, so that conclusions based on the results presented here should be made with care.
Solvers:
- BARON 8.1.5 (as in GAMS 22.8.1) with optionfile
epsr 0.01
contol 1e-6
maxtime 3600
PEnd 10
PDo 100
(the latter two options reduce the amount of probing done by BARON and
have shown to give slightly improved results)
- LINDOGLOBAL 5.0.1.292 (as in GAMS 22.8.1) with optionfile
SOLVER_FEASTOL 1e-6
- AlphaECP 1.63 (as in GAMS 22.8.1) with optionfile
epsilon_g 1e-6
- BONMIN 0.99 with optionfile
bonmin.algorithm B-OA
using Bonmin/stable/0.99 rev. 1325 (26.08.08), CoinUtils/stable/2.3 rev. 1070, Clp/stable/1.8 rev. 1265,
Cgl/stable/0.53 rev. 680, Osi/stable/0.99 rev. 1288, Cbc/stable/2.2 rev. 1052, Ipopt/stable/3.5 rev. 1314,
MA27, MA28, MC19, Netlib Blas, Netlib Lapack
with extra compiler flags: -DCLP_FAST_CODE -DCOIN_FAST_CODE -DCOIN_USE_RESTRICT -DNEW_STYLE_SOLVER=0
- DICOPT 2x-C (as in GAMS 22.8.1) with optionfile
mipsolver cplex
nlpsolver conopt
maxcycles 10000
stop 1
- SBB (as in GAMS 22.8.1) with optionfile
rootsolver conopt.1
subsolver conopt.1
acceptnonopt 1
memnodes 9999999
and conopt optionfile
Rtmaxv 1E+20
- OQNLP (as in GAMS 22.8.1) with optionfile
FEASIBILITY_TOLERANCE 1e-6
ITERATION_LIMIT 100000000
ITERATION_PRINT_FREQUENCY 100
LOGFILE_ITN_PRINT_FREQUENCY 100
MAX_SOLVER_CALLS 100000000
MAX_SOLVER_CALLS_NOIMPROVEMENT 1000
MAXTIME 3600
NLPSOLVER conopt.1
and conopt optionfile
Rtmaxv 1E+20
Subsolvers:
- CONOPT 3.14S (as in GAMS 22.8.1)
- CPLEX 11.1.1 (as in GAMS 22.8.1)
Timelimit: 3600 seconds
Gap tolerance: 1%
Platform:
- Intel Core 2 Duo CPU T9300, 2.5 GHz, 4 GB RAM
- Linux 2.6.22.18, SuSE Linux 10.3, 32 Bit
- gcc 4.2.1
The following modifications on the generated tracefiles have been made manually:
- LINDOGLOBAL:
The model status had been changed from optimal (1) to integer feasible (8) if the gap (objective value - objective estimate) was nonzero.
The objective estimate was set to NA if it contradicted the optimal value reported by another solver.
- OQNLP:
Reported negative resource usages have been replaced with the wallclock time.
Due to large differences between the reported resource usages (cputime) and the measured wallclock time, all resource usage have been replaced by wallclock times.
- SBB:
Due to large differences between the reported resource usages (cputime) and the measured wallclock time, all resource usage have been replaced by wallclock times.
- ALPHAECP:
For instances where the resource usage differed by more than 10% from the wallclock time, the resource usage was replaced by the wallclock time.
- DICOPT:
For instances where the resource usage differed by more than 10% from the wallclock time, the resource usage was replaced by the wallclock time.
- BONMIN:
For convex models, the reported objective estimate was set to NA if it contradicted the optimal value reported by another solver.
Results
Note: The solvers AlphaECP, DICOPT, and OQNLP do not report objective estimates in their solution records, so that
they do not appear in the performance profiles where only models with 1% or 10% gap are considered.