pymhlib - a toolbox for metaheuristics and hybrid optimization methods
pymhlib - A Toolbox for Metaheuristics and Hybrid Optimization Methods
This project is still in early development, any feedback is much appreciated!
pymhlib is a collection of modules supporting the efficient implementation of metaheuristics
and certain hybrid optimization approaches for solving primarily combinatorial optimization
problems in Python 3.7+.
mhlib version emerged from the
mhlib to which it has certain similarities
but also many differences.
The main purpose of the library is to support rapid prototyping and teaching. While ultimately efficient implementations of such algorithms in compiled languages like C++ will likely be faster, the expected advantage of the Python implementation lies in the expected faster implementation.
pymhlib is developed primarily by the
Algorithms and Complexity Group of TU Wien,
Vienna, Austria, since 2019.
- Günther Raidl (primarily responsible)
- Nikolaus Frohner
- Thomas Jatschka
- Daniel Obszelka
- Andreas Windbichler
Major versions of
pymhlib can be installed from
python3 -m pip install -U pymhlib
and development versions are available at https://github.com/ac-tuwien/pymhlib.
An abstract base class
Solutionthat represents a candidate solution to an optimization problem and derived classes
SetSolutionfor solutions which are represented bei general fixed-length vectors, boolean vectors or sets of arbitrary elements.
A more specific solution class
BinaryVectorSolutionfor problems in which solutions are represented by fixed-length binary vectors.
A more specific solution class
SubsetVectorSolutionfor problems in which solutions are subsets of a larger set. The set is realized by an efficient numpy array which is split into two parts, the one with the included elements in sorted order and the one with the remaining elements.
A more specific solution class
PermutationSolutionfor problems in which solutions are permutations of a set of elements.
- scheduler.py: A an abstract framework for single metaheuristics that rely on iteratively applying certain methods to a current solution. Modules like gvns.py and alns.py extend this abstract class towards more specific metaheuristics.
- gvns.py: A framework for local search, iterated local search, (general) variable neighborhood search, GRASP, etc.
- alns.py: A framework for adaptive large neighborhood search (ALNS).
- par_alns.py: A multi-process implementation of the ALNS where destroy+repair operations are parallelized.
- population.py A population class for population-based metaheuristics.
- pbig.py: A population based iterated greedy (PBIG) algorithm.
- ssga.py: A steady-state genetic algorithm (SSGA).
- sa.py: A simulated annealing (SA) algorithm with geometric cooling.
- decision_diag.py: A generic class for (relaxed) decision diagrams for combinatorial optimization.
Provides two logger objects, one for writing out general log information, which is typically
written into a
*.outfile, and one for iteration-wise information, which is typically written into a
*.logfile. The latter is buffered in order to work also efficiently, e.g., on network drives and massive detailed log information. A class
LogLevelis provided for indented writing of log information according to a current level, which might be used for hierarchically embedded components of a larger optimization framework, such as a local search that is embedded in a population-based approach.
Allows for defining module-specific parameters directly in each module in an independent distributed
way, while values for these parameters can be provided as program arguments or in
configuration files. Most
pyhmlibmodules rely on this mechanism for their external parameters.
Modules/scripts for analyzing results of many runs:
multi_run_summary.py: Collects essential information from multiple
pymhlibalgorithm runs found in the respective out and log files and returns a corresponding pandas dataframe if used as a module or as a plain ASCII table when used as independent script. The module can be easily configured to extract also arbitrary application-specific data.
aggregate_results.py: Calculate grouped basic statistics for one or two dataframes/TSV files obtained e.g. from
multi-run-summary.py. In particular, two test series with different algorithms or different settings can be statistically compared, including Wilcoxon signed rank tests. The module can be used as standalone script as well as module called, e.g., from a jupyter notebook.
For demonstration purposes, simple metaheuristic approaches are provided in the
demo subdirectory for the following
well-known combinatorial optimization problems. They can be startet by
python3 -m pymhlib.demos.<problem> ...
<problem> is one of the following and
... represents further parameters that can be seen by providing
It is recommended to take such a demo as template
for solving your own problem.
maxsat: maximum satisfiability problem based on
tsp: traveling salesperson problem based on
qap: quadratic assignment problem based on
vertex_cover: minimum vertex cover problem based on
graph_coloring: graph coloring problem based on
misp: maximum (weighted) independent set problem based on
mkp: multidimensional 0-1 knapsack problem based on
Shared code of these demos is found in the submodules
test instance data in
julia-maxsat.jl demonstrate the integration with the Julia programming language.
Implementing time-critical parts of an application in Julia may accelerate the code substantially.
To run this demo, Julia must be set up correctly and Python's
julia package must be installed.
While this demo derives a whole solution class in Julia,
julia-maxsat2.py is a variant where only two functions
are realized in Julia.
Major changes over major releases:
- directory renamed to pymhlib to correspond to module name
- bug fix in Metropolis criterion of ALNS
- boolean arguments must now be specified in the command line as any other parameter
- basic functionality test
tests/test_all.pyfor all problems and algorithms added
- polishing, minor fixes
- ALNS and parallel ALNS added
- graph coloring, TSP, and minimum vertex cover demos added
- population based iterated greedy and steady state genetic algorithms added
- SA with geometric cooling added
- demos.graphs introduced
- mhlib renamed to pymhlib
- demo for interfacing with Julia added
- many smaller improvements, bug fixes, improvements in documentation
- Initial version
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size pymhlib-0.1.2-py3-none-any.whl (222.2 kB)||File type Wheel||Python version py3||Upload date||Hashes View|
|Filename, size pymhlib-0.1.2.tar.gz (233.6 kB)||File type Source||Python version None||Upload date||Hashes View|