Benchmarking Large-Scale Optimization Software
Services we provide
1. Guide to Software:
"Decision Tree"
2. Software Archive
3. Software Evaluation: "Benchmarks"
4. Archive of Testproblems
5. Web-based Solvers (1/3 of
NEOS)
Outline of Talk
- Software, Testproblem Archives
- Our NEOS Solvers
- Report on recent benchmarks in
mostly discrete optimization
- MILP, MIQP commercial and free
- MINLP GAMS Performance Tools
- SDP DIMACS, sparse SDP
- Conclusions
Software Archive
- held in ftp area
- DONLP2 (NLP-SQP) and
other software
- some f90 versions
- binaries (source not public)
- codes not available otherwise
Testproblem Archive
- benchmark problems not available elsewhere
- converted formats
- large library of AMPL models
- miscellaneous
- MILP, MIQP
- convex QP
- smooth & nonsmooth NLP
- nonlinear systems
- PDE-constrained NLP
NEOS Solvers
(20 solvers, 30 formats, 3 Kestrel)
LP
MILP
Non-differentiable
NLP
SDP/SOCP
- CirCut, CSDP, MOSEK, PENNON, SDPA, SDP-LR, SDPT3, SeDuMi
Semi-infinite
Global
Some Benchmarks
-
Benchmark Policy
- absolute reproducibility
- datafiles, codes available
- default options
- (some) logfiles posted
Selected Benchmarks
- Integer Programming
- SDP/SOCP Programming
Conclusions
- MILP:
CPLEX, XPRESS competitive, GLPK decent in B&B
- MIQP:
CPLEX seems to have edge, see new XPRESS, MOSEK
- MINLP:
DICOPT, MINLP, SBB competitive
- SDP/SOCP:
CSDP, SDPA improved for large, sparse problems,
MOSEK, PENNON, SDPT3, SeDuMi have clear strenghts