scikit-glpk 0.1.1

Python linprog interface for GLPK

sckit-glpk

Proof of concept Python wrappers for GLPK.

Installation

Should be an easy pip installation:

pip install scikit-glpk

A Python-compatible C compiler is required to build GLPK from source.

Usage

There are a few things in this package:

• glpk() : the wrappers over the solvers (basically acts like Python-friendly glpsol)

• mpsread() : convert an MPS file to some matrices

from glpk import glpk, GLPK

res = glpk(
    c, A_ub, b_ub, A_eq, b_eq, bounds, solver, sense, maxit, timeout,
    basis_fac, message_level, disp, simplex_options, ip_options,
    mip_options, libpath)

c, A_ub, b_ub, A_eq, b_eq, bounds = mpsread(
    filename, fmt=GLPK.GLP_MPS_FILE, ret_glp_prob=False, libpath=None)

There’s lots of information in the docstrings for these functions, please check there for a complete listing and explanation.

Notice that glpk is the wrapper and GLPK acts as a namespace that holds constants.

bounds is also behaves a little differently for both of these:

• as an input to glpk, bounds is a list of triplets in the style of GLPK (probably should be converted to linprog conventions)

• as an output of mpsread, bounds is a list of tuples in the style of linprog

GLPK stuffs

GLPK is installed with the module and a linprog-like wrapper is provided with a ctypes backend. A pared-down version of glpk-4.65 is vendored from here and compile instructions are scraped from the makefiles. I’ll try my best to support cross-platform pip installations.

Background

The GNU Linear Programming Kit (GLPK) has simplex, interior-point, and MIP solvers all callable from a C library. We would like to be able to use these from within Python and be potentially included as a backend for scipy’s linprog function.

Note that there are several projects that aim for something like this, but which don’t match up for what I’m looking for:

• python-glpk : no longer maintained (dead as of ~2013)

• PyGLPK : GPL licensed

• PyMathProg : GPL licensed, uses different conventions than that of linprog

• Pyomo : Big, uses different conventions than that of linprog

• CVXOPT : Big, GPL licensed

• Sage : Big, GPL licensed

• pulp : Calls glpsol from command line (writes problems, solutions to file instead of shared memory model – this is actually easy to do)

• yaposib : seems dead? OSI-centric

• ecyglpki : GPL licensed, dead?

• swiglpk : GPL licensed, low level

• optlang : sympy-like, cool project otherwise

Most existing projects lean to GPL licenses. Not a bad thing, but would hinder adoption into scipy (not GLPK itself, but a wrapper to use from within scipy if the user has chosen to install GLPK herself).

Why do we want this?

GLPK has a lot of options that the current scipy solvers lack as well as robust MIP support (only basic in HiGHS). It is also a standard, well known solver in the optimization community. The only thing that I want that it lacks on an API level is robust support for column generation. Easy access to GLPK as a backend to linprog would be very welcome (to me at least).

Approach

Since the underlying API is quite simple and written in C and only C, ctypes is a good fit for this.

GLPK is packaged but I may want to make it so the user can optionally specify where the installation is on a user’s computer (i.e., path to the shared library) so GLPK is not packaged with scikit-glpk and/or scipy. linprog could then presumably route the problem to the GLPK backend instead of HiGHS or the existing native python solvers.

The ctypes wrapper is required for integrating GLPK into the Python runtime. Instead of using MPS files to communicate problems and reading solutions from files, scipy.sparse.coo_matrix and numpy arrays can be passed directly to the library. More information can be extracted from GLPK this way as well (For example, there is no way to get iteration count except by reading directly from the underlying structs. It is only ever printed to stdout, no other way to get it).

TODO

• Several GLPK solver options (notably tolerances) not wrapped yet