pytoulbar2 1.3.0

ToulBar2 Python package

toulbar2

Exact optimization for cost function networks and additive graphical models

What is toulbar2?

toulbar2 is an open-source black-box C++ optimizer for cost function networks and discrete additive graphical models. This also covers Max-SAT, Max-Cut, QUBO (and constrained variants), among others. It can read a variety of formats. The optimized criteria and feasibility should be provided factorized in local cost functions on discrete variables. Constraints are represented as functions that produce costs that exceed a user-provided primal bound. toulbar2 looks for a non-forbidden assignment of all variables that optimizes the sum of all functions (a decision NP-complete problem).

toulbar2 won several competitions on deterministic and probabilistic graphical models:

• Max-CSP 2008 Competition CPAI08 (winner on 2-ARY-EXT and N-ARY-EXT)
• Probabilistic Inference Evaluation UAI 2008 (winner on several MPE tasks, inra entries)
• 2010 UAI APPROXIMATE INFERENCE CHALLENGE UAI 2010 (winner on 1200-second MPE task)
• The Probabilistic Inference Challenge PIC 2011 (second place by ficolofo on 1-hour MAP task)
• UAI 2014 Inference Competition UAI 2014 (winner on all MAP task categories, see Proteus, Robin, and IncTb entries)
• XCSP3 Competitions (first place on Mini COP in 2023 and 2025, second place on Mini COP and Parallel COP tracks in 2022, third place in 2024)
• UAI 2022 Inference Competition UAI 2022 (winner on all MPE and MMAP task categories)
• Pseudo-Boolean Competition 2025 PB25 (OPT-LIN ranking 39/46 ; PARTIAL-LIN ranking 6/9, but it gave the best known answer from an incomplete solver point of view in 185 instances among 208)

toulbar2 is now also able to collaborate with ML code that can learn an additive graphical model (with constraints) from data (see the associated paper, slides and video where it is shown how it can learn user preferences or how to play the Sudoku without knowing the rules). The current CFN learning code is available on GitHub.

Installation from binaries

You can install toulbar2 directly using the package manager in Debian and Debian derived Linux distributions (Ubuntu, Mint,...):

sudo apt-get update
sudo apt-get install toulbar2 toulbar2-doc

For the most recent binary or the Python API, compile from source.

Python interface

An alpha-release Python interface can be tested through pip on Linux and MacOS:

python3 -m pip install --upgrade pip
python3 -m pip install pytoulbar2

The first line is only useful for Linux distributions that ship "old" versions of pip.

Commands for compiling the Python API on Linux/MacOS with cmake (Python module in lib/*/pytb2.cpython*.so):

pip3 install pybind11
mkdir build
cd build
cmake -DPYTB2=ON ..
make

Move the cpython library and the experimental pytoulbar2.py python class wrapper in the folder of the python script that does "import pytoulbar2".

Download

Download the latest release from GitHub (https://github.com/toulbar2/toulbar2) or similarly use tag versions, e.g.:

git clone --branch 1.2.2 https://github.com/toulbar2/toulbar2.git

Installation from sources

Compilation requires git, cmake and a C++-17 capable compiler (in C++17 mode).

Required library:

• libgmp-dev
• bc (used during cmake)

Recommended libraries (default use):

• libboost-graph-dev
• libboost-iostreams-dev
• libboost-serialization-dev
• zlib1g-dev
• liblzma-dev
• libbz2-dev

Optional libraries:

• libjemalloc-dev
• pybind11-dev
• libopenmpi-dev
• libboost-mpi-dev
• libicuuc
• libicui18n
• libicudata
• libxml2-dev
• libxcsp3parser
• libeigen3-dev

On MacOS, run ./misc/script/MacOS-requirements-install.sh to install the recommended libraries. For Mac with ARM64, add option -DBoost=OFF to cmake.

Commands for compiling toulbar2 on Linux/MacOS with cmake (binary in build/bin/*/toulbar2):

mkdir build
cd build
cmake ..
make

Commands for statically compiling toulbar2 on Linux in directory toulbar2/src without cmake:

bash
cd src
echo '#define Toulbar_VERSION "1.2.2"' > ToulbarVersion.hpp
echo -n "" > tb2config.hpp
g++ -o toulbar2 -std=c++17 -O3 -DNDEBUG -march=native -flto -static -static-libgcc -static-libstdc++ -DBOOST -DLONGDOUBLE_PROB -DLONGLONG_COST -DWCSPFORMATONLY \
 -I. -I./pils/src tb2*.cpp applis/*.cpp convex/*.cpp core/*.cpp globals/*.cpp incop/*.cpp mcriteria/*.cpp pils/src/exe/*.cpp search/*.cpp utils/*.cpp vns/*.cpp ToulbarVersion.cpp \
 -lboost_graph -lboost_iostreams -lboost_serialization -lgmp -lz -lbz2 -llzma

Use OPENMPI flag and MPI compiler for a parallel version of toulbar2 (must be run with mpirun, use mpirun -n 1 for the sequential version of HBFS or VNS):

bash
cd src
echo '#define Toulbar_VERSION "1.2.2"' > ToulbarVersion.hpp
echo -n "" > tb2config.hpp
mpicxx -o toulbar2 -std=c++17 -O3 -DNDEBUG -march=native -flto -DBOOST -DLONGDOUBLE_PROB -DLONGLONG_COST -DWCSPFORMATONLY -DOPENMPI \
 -I. -I./pils/src tb2*.cpp applis/*.cpp convex/*.cpp core/*.cpp globals/*.cpp incop/*.cpp mcriteria/*.cpp pils/src/exe/*.cpp search/*.cpp utils/*.cpp vns/*.cpp ToulbarVersion.cpp \
 -lboost_graph -lboost_iostreams -lboost_serialization -lboost_mpi -lgmp -lz -lbz2 -llzma

Replace LONGLONG_COST by INT_COST to reduce memory usage by two and reduced cost range (costs must be smaller than 10^8).

Replace WCSPFORMATONLY by XMLFLAG3 and add libxcsp3parser.a from xcsp.org in your current directory for reading XCSP3 files:

bash
cd src
echo '#define Toulbar_VERSION "1.2.2"' > ToulbarVersion.hpp
echo -n "" > tb2config.hpp
mpicxx -o toulbar2 -std=c++17 -O3 -DNDEBUG -march=native -flto -DBOOST -DLONGDOUBLE_PROB -DLONGLONG_COST -DXMLFLAG3 -DOPENMPI \
 -I/usr/include/libxml2 -I. -I./pils/src -I./xmlcsp3 tb2*.cpp applis/*.cpp convex/*.cpp core/*.cpp globals/*.cpp incop/*.cpp mcriteria/*.cpp pils/src/exe/*.cpp search/*.cpp utils/*.cpp vns/*.cpp ToulbarVersion.cpp \
 -lboost_graph -lboost_iostreams -lboost_serialization -lboost_mpi -lxml2 -licuuc -licui18n -licudata libxcsp3parser.a -lgmp -lz -lbz2 -llzma -lm -lpthread -ldl

Copyright (C) 2006-2026, toulbar2 team. toulbar2 is currently maintained by Simon de Givry, INRAE - MIAT, Toulouse, France ([e-mail])