|
|
|
|
|
| Programming Language |
Python |
Python |
Julia |
C++, Python, Java, C# |
| Official Documentation Link |
Link |
Link |
Link |
Link |
| Open-source vs License requirement |
Requires Gurobi solver license for model generation. |
open-source |
Open-source |
Open-source |
|
|
|
|
|
| However, for basic API testing restricted, size-limited license can be obtained freely that can handle model of size: 2000 variables, 2000 linear constraints, and 200 variables if quadratic terms. |
|
|
|
|
| Maintainers |
|
|
|
|
| Gurobi |
Open-source |
JuMP-dev |
Google |
|
| Supported Problem Types |
LP, MIP, MILP, NLP, MINLP |
LP, MIP, MILP, NLP, MINLP |
LP, MIP, MILP, NLP, MINLP Link |
LP, MIP, MILP |
|
|
|
|
|
| What hack to do when you have nonlinear functions? Piecewise? |
|
|
|
|
| Solver independence vs Lock in |
Lock in. Gurobi solver is the only supported solver. |
Independent |
Independent |
Independent |
| Supported Solvers |
Gurobi |
20+ solvers supported including major commercial and open-source ex. CPLEX, Gurobi, BARON, IPOPT, GLPK etc. |
Incredibly broad solver support. 40+ solvers supported (Link). |
Half a dozen solvers supported: commercial solvers such as Gurobi or CPLEX, or open-source solvers such as SCIP, GLPK, or Google's GLOP, CP-SAT. |
|
|
|
|
|
| However, Pyomo is designed to be extensible and custom solver interfaces are supported. |
Includes nearly all major commercial and opens-source solvers. |
|
|
|
| NEOS Server Compatible |
No. |
Yes |
Yes (Link) |
No |
|
|
|
|
|
| Available solver interfaces: Link |
Solvers available: CPLEX, FICO-Xpress, Ipopt, Knitro, MOSEK, OCTERACT, SNOPT. |
|
|
|
| ** Ease of use and Learning Curve |
Quite Pythonic. |
Quite Pythonic. |
Has a few week learning curve for Julia if coming from Python. |
Relatively less steep for existing Python developers. |
| ** Expressiveness / Readability |
Considered Good. |
Clear Python based syntax |
Considered good. |
Complex models can be difficult to read. |
| ** Vendor / Community Support |
Developing. |
Strong community support. |
Moderate |
Developing |
|
|
|
|
|
| Strong technical support from Gurobi through dedicated support as well as through forum (Link) |
Active discourse community platform (Link) |
|
|
|
| Library Actively Maintained |
|
|
|
|
| Yes |
|
|
|
|
|
|
|
|
|
| Stability / Number of releases |
11 releases since April 2021. |
77 since Dec 2014. |
105 since Oct 2013 with a release every 2-4 weeks |
58 since Jan 2014. |
| Data Transformation and Data Visualization |
|
|
|
|
| Yes. |
|
|
|
|
|
|
|
|
|
| Variable values, bounds, objective function value, constraints slacks can all be transformed to Python data structures and conversely, they can be used to initialize variable values, bounds, constraint limits etc. |
|
|
|
|
|
|
|
|
|
| Notebooks: Interactive Graphical Environment |
|
|
|
|
| Yes |
Yes |
Yes. Through IJulia. |
Yes |
|
| ** Automatic Differentiation? |
No |
yes |
Yes |
There is no code doing automatic-differentiation. That is one reason why NLP, MINLP models are not supported. |
| Model Generation times |
|
|
|
|
| ** Memory Consumption |
|
|
|
|
| Solver Callback |
Supports Gurobi Callback |
No |
For selected solvers: Gurobi, CPLEX, GLPK |
Is Possible with some solvers. |
|
|
|
|
|
| JuMP's internal representation of the problem as close as possible to the solver's internal representation. This allows us to make efficient in-memory modifications to models in a loop |
|
|
|
|
| ** Support for Abstract Models |
No |
Abstract – Yes |
No. Default is concrete. |
Yes |
| Concrete – Yes |
Abstract models need a way around. Link |
|
|
|
|
|
|
|
|
| Presolve/Reformulate |
|
No |
|
|
| Data Structures |
Tuplelist |
Leverages basic data structures of Python like list, dict, tuple. |
Arrays, DenseAxisArrays, and SparseAxisArrays |
|
| Tupledict |
|
|
|
|
| These data structures lend additional methods for efficient sub-listing and expression building. |
|
|
|
|
| Ex. |
|
|
|
|
| Other Features? |
Model updates are possible, but you have to use model.update() again and again. |
Stochastic programming |
|
|
|
|
|
|
|
| Gurobipy pandas |
Bilevel programming |
|
|
|
|
|
|
|
|
| Using machine learning predictors in Gurobipy |
Generalized disjunctive programming |
|
|
|
|
|
|
|
|
| clear separation between declarative model and data |
|
|
|
|
| MPS format vs NL format |
|
|
|
|
| Parallel Solver Support? |
yes |
yes |
Yes |
yes |
|
|
|
|
|
| JuMP is single threaded (JuMP model must be built sequentially on a single thread), but supports parallel solvers. |
|
|
|
|
| Multiple Objectives |
Yes |
No |
No. |
No |
|
|
|
|
|
| Model.getMultiobjEnv() |
Pyomo does not have special facilities to handle multi-objective models. The best method is to self-implement augmented epsilon-constraint method |
With the standard JuMP you can have only one goal function at a time. Running another @objective macro just overwrites the previous goal function. |
But can be achieved with custom epsilon-constraint method. |
|
|
|
|
|
|
|
|
Check MultiObjectiveAlgorithms.jl package |
|
|
| ** Piecewise Linear |
yes |
yes |
No |
yes |
| Infeasibility Log |
Model.computeIIS() |
log_infeasible_constraints(model) |
Not as easy as in Pyomo, though sometimes solver functionality can be leveraged. |
Not straightforward again as in Gurobipy and Pyomo. |
| Non-contiguous Intvar ?? |
|
|
|
|
| Solution hint/Warm Start? |
|
|
|
|
| IIS – Irreducible Infeasible Subset. |
|
|
|
|
|
|
|
|
|
|
|
|
|
|