PySCIPOpt  5.1.1
Python Interface for the SCIP Optimization Suite
PySCIPOpt

This project provides an interface from Python to the SCIP Optimization Suite. Starting from v8.0.3, SCIP uses the Apache2.0 license. If you plan to use an earlier version of SCIP, please review SCIP's license restrictions.

Gitter PySCIPOpt on PyPI Integration test coverage AppVeyor Status

Documentation

Please consult the online documentation or use the help() function directly in Python or ? in IPython/Jupyter.

See CHANGELOG.md for added, removed or fixed functionality.

Installation

Using Conda

Conda version Conda platforms

DO NOT USE THE CONDA BASE ENVIRONMENT TO INSTALL PYSCIPOPT.

Conda will install SCIP automatically, hence everything can be installed in a single command:

conda install --channel conda-forge pyscipopt

Using PyPI and from Source

See INSTALL.md for instructions. Please note that the latest PySCIPOpt version is usually only compatible with the latest major release of the SCIP Optimization Suite. The following table summarizes which version of PySCIPOpt is required for a given SCIP version:

SCIP PySCIPOpt
9.1 5.1+
9.0 5.0.x
8.0 4.x
7.0 3.x
6.0 2.x
5.0 1.4, 1.3
4.0 1.2, 1.1
3.2 1.0

Information which version of PySCIPOpt is required for a given SCIP version can also be found in INSTALL.md.

Building and solving a model

There are several examples and tutorials. These display some functionality of the interface and can serve as an entry point for writing more complex code. Some of the common usecases are also available in the recipes sub-package. You might also want to have a look at this article about PySCIPOpt: https://opus4.kobv.de/opus4-zib/frontdoor/index/index/docId/6045. The following steps are always required when using the interface:

1) It is necessary to import python-scip in your code. This is achieved by including the line

from pyscipopt import Model

2) Create a solver instance.

model = Model("Example") # model name is optional

3) Access the methods in the scip.pxi file using the solver/model instance model, e.g.:

x = model.addVar("x")
y = model.addVar("y", vtype="INTEGER")
model.setObjective(x + y)
model.addCons(2*x - y*y >= 0)
model.optimize()
sol = model.getBestSol()
print("x: {}".format(sol[x]))
print("y: {}".format(sol[y]))

Writing new plugins

The Python interface can be used to define custom plugins to extend the functionality of SCIP. You may write a pricer, heuristic or even constraint handler using pure Python code and SCIP can call their methods using the callback system. Every available plugin has a base class that you need to extend, overwriting the predefined but empty callbacks. Please see test_pricer.py and test_heur.py for two simple examples.

Please notice that in most cases one needs to use a dictionary to specify the return values needed by SCIP.

Extending the interface

PySCIPOpt already covers many of the SCIP callable library methods. You may also extend it to increase the functionality of this interface. The following will provide some directions on how this can be achieved:

The two most important files in PySCIPOpt are the scip.pxd and scip.pxi. These two files specify the public functions of SCIP that can be accessed from your python code.

To make PySCIPOpt aware of the public functions you would like to access, you must add them to scip.pxd. There are two things that must be done in order to properly add the functions:

1) Ensure any enums, structs or SCIP variable types are included in scip.pxd
2) Add the prototype of the public function you wish to access to scip.pxd

After following the previous two steps, it is then possible to create functions in python that reference the SCIP public functions included in scip.pxd. This is achieved by modifying the scip.pxi file to add the functionality you require.

We are always happy to accept pull request containing patches or extensions!

Please have a look at our contribution guidelines.

Gotchas

Ranged constraints

While ranged constraints of the form

lhs <= expression <= rhs

are supported, the Python syntax for chained comparisons can't be hijacked with operator overloading. Instead, parenthesis must be used, e.g.,

lhs <= (expression <= rhs)

Alternatively, you may call model.chgRhs(cons, newrhs) or model.chgLhs(cons, newlhs) after the single-sided constraint has been created.

Variable objects

You can't use Variable objects as elements of sets or as keys of dicts. They are not hashable and comparable. The issue is that comparisons such as x == y will be interpreted as linear constraints, since Variables are also Expr objects.

Dual values

While PySCIPOpt supports access to the dual values of a solution, there are some limitations involved:

  • Can only be used when presolving and propagation is disabled to ensure that the LP solver - which is providing the dual information - actually solves the unmodified problem.
  • Heuristics should also be disabled to avoid that the problem is solved before the LP solver is called.
  • There should be no bound constraints, i.e., constraints with only one variable. This can cause incorrect values as explained in #136

Therefore, you should use the following settings when trying to work with dual information:

model.setPresolve(pyscipopt.SCIP_PARAMSETTING.OFF)
model.setHeuristics(pyscipopt.SCIP_PARAMSETTING.OFF)
model.disablePropagation()

Citing PySCIPOpt

Please cite this paper

@incollection{MaherMiltenbergerPedrosoRehfeldtSchwarzSerrano2016,
author = {Stephen Maher and Matthias Miltenberger and Jo{\~{a}}o Pedro Pedroso and Daniel Rehfeldt and Robert Schwarz and Felipe Serrano},
title = {{PySCIPOpt}: Mathematical Programming in Python with the {SCIP} Optimization Suite},
booktitle = {Mathematical Software {\textendash} {ICMS} 2016},
publisher = {Springer International Publishing},
pages = {301--307},
year = {2016},
doi = {10.1007/978-3-319-42432-3_37},
}

as well as the corresponding SCIP Optimization Suite report when you use this tool for a publication or other scientific work.