CoinOptServices.jl

Linux, OSX:

Windows:

The build script for this package has known issues on macOS and Linux that are difficult to address because the latest release of Optimization Services (2.10.1) does not compile on recent compilers (e.g., GCC 6.3 and later). If you are unable to install successfully, we recommend using AmplNLWriter for access to Bonmin and Couenne. Linux and macOS users should compile Bonmin or Couenne locally with ASL support and point AmplNLSolver to the path of the respective solver binary. Windows users should continue to use the binaries installed through this package for now. We hope to have new Julia packages to install Bonmin and Couenne binaries soon. (Help is welcome!). Users may also consider trying out pure-Julia solvers that have similar functionality, e.g., Pajarito, Katana, POD, and Juniper.

This Julia package is an interface between MathProgBase.jl and COIN-OR Optimization Services (OS), translating between the Julia-expression-tree MathProgBase format for nonlinear objective and constraint functions and the Optimization Services instance Language (OSiL) XML-based optimization problem interchange format.

By writing .osil files and using the OSSolverService command-line driver, this package allows Julia optimization modeling languages such as JuMP to access any solver supported by OSSolverService. This includes the COIN-OR solvers Clp (linear programming), Cbc (mixed-integer linear programming), Ipopt (nonlinear programming), Bonmin (evaluation-based mixed-integer nonlinear programming), Couenne (expression-tree-based mixed-integer nonlinear programming), and several others.

Note that Clp, Cbc, and Ipopt already have Julia packages that interface directly with their respective in-memory C API's. Particularly for Clp.jl and Cbc.jl, the existing packages should be faster than the CoinOptServices.jl approach of going through an OSiL file on disk. Initial comparisons show that Ipopt.jl is also substantially faster than CoinOptServices.jl. For nonlinear problems OSSolverService performs automatic differentiation in C++ using CppAD, which has different performance characteristics than the pure-Julia ReverseDiffSparse.jl package used for nonlinear programming in JuMP. TODO: determine why CppAD is slower than expected

Writing of .osil files is implemented using the LightXML.jl Julia bindings to libxml2 to construct XML files from element trees. Reading of .osil files will be done later, to provide a (de-)serialization format for storage, archival, and interchange of optimization problems between various modeling languages.

Installation

You can install the package by running:

julia> Pkg.add("CoinOptServices")

On OS X, this will automatically download precompiled binaries via Homebrew.jl.

On Windows, this will automatically download precompiled binaries via WinRPM.jl. Currently these are packaged in @tkelman's personal project on the openSUSE build service, but these will be submitted to the official default repository eventually.

On Linux, this will compile the COIN OS library and its dependencies from source if they are not found in DL_LOAD_PATH. Note that OS is a large C++ library with many dependencies, and it is not currently packaged for any released Linux distributions. Submit a pull request to support using the library from a system package manager if this changes. It is recommended to set ENV["MAKEFLAGS"] = "-j4" before installing the package so compilation does not take as long.

The current BinDeps setup assumes Ipopt.jl and Cbc.jl have already been successfully installed in order to reuse the binaries for those solvers. You will need to have a Fortran compiler such as gfortran installed in order to compile Ipopt. On Linux, use your system package manager to install gfortran. You will also need to have pkg-config installed.

This package builds the remaining COIN-OR libraries OS, CppAD, Bonmin, Couenne, and a few other solvers (DyLP, Vol, SYMPHONY, Bcp) that do not yet have Julia bindings.

Usage

CoinOptServices is usable as a solver in JuMP as follows.

julia> using JuMP, CoinOptServices
julia> m = Model(solver = OsilSolver())

Then model and solve your optimization problem as usual. See JuMP's documentation for more details. The OsilSolver() constructor takes several optional keyword arguments. You can specify OsilSolver(solver = "couenne") to request a particular sub-solver, OsilSolver(osil = "/path/to/file.osil") or similarly osol or osrl keyword arguments to request non-default paths for writing the OSiL instance file, OSoL options file, or OSrL results file. The default location for writing these files is under Pkg.dir("CoinOptServices", ".osil"). The printLevel keyword argument can be set to an integer from 0 to 5, and corresponds to the -printLevel command line flag for OSSolverService. This only controls the print level of the OS driver, not the solvers themselves.

Note that if you want to solve multiple problems simultaneously, you need to set the osil, osol, and osrl keyword arguments to independent file names for each problem. See issue #1 for details.

All additional inputs to OsilSolver are treated as solver-specific options. These options should be input as Julia Dict objects, with keys corresponding to OSoL <solverOption> properties "name", "value", "solver", "category", "type", or "description". A convenience function OSOption(optname, optval, kwargs...) is provided to automatically set "type" based on the Julia type of optval.

CoinOptServices should also work with any other MathProgBase-compliant linear or nonlinear optimization modeling tools, though this has not been tested. There are features in OSiL for representing conic optimization problems, but these are not currently exposed or connected to the MathProgBase conic interface.