/coco

Numerical Black-Box Optimization Benchmarking Framework

Primary LanguageCOtherNOASSERTION

numbbo/coco: Comparing Continuous Optimizers

Please click here to get started (unless for submitting an issue or for working on the code base)

Caution

We are currently refactoring the coco code base to make it more accessible. Much of the documentation is therefore outdated or in a state of flux. The code has been refactored into several repositories under github/numbbo. We try our best to update everything as soon as possible, if you find something that you think is outdated or needs a better description, don't hestitate to open an issue or a pull request!

Important

This repository contains the source files for the coco framework. If you don't want to extend the framework, you probably don't need this! See instead the new documentation and use the language bindings of your choice from the package repository for your language (e.g. PyPI for Python, crates.io for Rust, ...).

DOI [BibTeX] cite as:

Nikolaus Hansen, Dimo Brockhoff, Olaf Mersmann, Tea Tusar, Dejan Tusar, Ouassim Ait ElHara, Phillipe R. Sampaio, Asma Atamna, Konstantinos Varelas, Umut Batu, Duc Manh Nguyen, Filip Matzner, Anne Auger. COmparing Continuous Optimizers: numbbo/COCO on Github. Zenodo, DOI:10.5281/zenodo.2594848, March 2019.

This code provides a platform to benchmark and compare continuous optimizers, AKA non-linear solvers for numerical optimization. It is fully written in ANSI C and Python (reimplementing the original Comparing Continous Optimizer platform) with other languages calling the C code. Languages currently available to connect a solver to the benchmarks are

  • C/C++
  • Java
  • MATLAB
  • Octave
  • Python
  • Rust

Contributions to link further languages (including a better example in C++) are more than welcome.

The general project structure is shown in the following figure where the black color indicates code or data provided by the platform and the red color indicates either user code or data and graphical output from using the platform:

General COCO Structure

For more general information:

Getting Started

Running Experiments

  1. For running experiments in Python, follow the (new) instructions here.
    Otherwise, download the COCO framework code from github,

    • either by clicking the Download ZIP button and unzip the zip file,
    • or by typing git clone https://github.com/numbbo/coco.git. This way allows to remain up-to-date easily (but needs git to be installed). After cloning, git pull keeps the code up-to-date with the latest release.

    The record of official releases can be found here. The latest release corresponds to the master branch as linked above.

  2. In a system shell, cd into the coco or coco-<version> folder (framework root), where the file do.py can be found. Type, i.e. execute, one of the following commands once

      python do.py run-c
  python do.py run-java
  python do.py run-matlab
  python do.py run-octave
  python do.py run-python

    depending on which language shall be used to run the experiments. run-* will build the respective code and run the example experiment once. The build result and the example experiment code can be found under code-experiments/build/<language> (<language>=matlab for Octave). python do.py lists all available commands.

  3. Copy the folder code-experiments/build/YOUR-FAVORITE-LANGUAGE and its content to another location. In Python it is sufficient to copy the file example_experiment_for_beginners.py or example_experiment2.py. Run the example experiment (it already is compiled). As the details vary, see the respective read-me's and/or example experiment files:

    If the example experiment runs, connect your favorite algorithm to Coco: replace the call to the random search optimizer in the example experiment file by a call to your algorithm (see above). Update the output result_folder, the algorithm_name and algorithm_info of the observer options in the example experiment file.

    Another entry point for your own experiments can be the code-experiments/examples folder.

  4. Now you can run your favorite algorithm on the bbob and bbob-largescale suites (for single-objective algorithms), on the bbob-biobj suite (for multi-objective algorithms), or on the mixed-integer suites (bbob-mixint and bbob-biobj-mixint respectively). Output is automatically generated in the specified data result_folder. By now, more suites might be available, see below.

Post-processing Data

  1. Install the post-processing for displaying data (using Python):

        pip install cocopp

  2. Postprocess the data from the results folder of a locally run experiment by typing

        python -m cocopp [-o OUTPUT_FOLDERNAME] YOURDATAFOLDER [MORE_DATAFOLDERS]

    Any subfolder in the folder arguments will be searched for logged data. That is, experiments from different batches can be in different folders collected under a single "root" YOURDATAFOLDER folder. We can also compare more than one algorithm by specifying several data result folders generated by different algorithms.

  3. We also provide many archived algorithm data sets. For example

      python -m cocopp 'bbob/2009/BFGS_ros' 'bbob/2010/IPOP-ACTCMA'

    processes the referenced archived BFGS data set and an IPOP-CMA data set. The given substring must have a unique match in the archive or must end with ! or * or must be a regular expression containing a * and not ending with ! or *. Otherwise, all matches are listed but none is processed with this call. For more information in how to obtain and display specific archived data, see help(cocopp) or help(cocopp.archives) or the class COCODataArchive.

    Data descriptions can be found for the bbob test suite at coco-algorithms and for the bbob-biobj test suite at coco-algorithms-biobj. For other test suites, please see the COCO data archive.

    Local and archived data can be freely mixed like

      python -m cocopp YOURDATAFOLDER 'bbob/2010/IPOP-ACT'

    which processes the data from YOURDATAFOLDER and the archived IPOP-ACT data set in comparison.

    The output folder, ppdata by default, contains all output from the post-processing. The index.html file is the main entry point to explore the result with a browser. Data from the same foldername as previously processed may be overwritten. If this is not desired, a different output folder name can be chosen with the -o OUTPUT_FOLDERNAME option.

    A summary pdf can be produced via LaTeX. The corresponding templates can be found in the code-postprocessing/latex-templates folder. Basic html output is also available in the result folder of the postprocessing (file templateBBOBarticle.html).

  4. In order to exploit more features of the post-processing module, it is advisable to use the module within a Python or IPython shell or a Jupyter notebook or JupyterLab, where

    import cocopp
help(cocopp)

    provides the documentation entry pointer.

If you detect bugs or other issues, please let us know by opening an issue in our issue tracker at https://github.com/numbbo/coco/issues.

Known Issues / Trouble-Shooting

Post-Processing

Too long paths for postprocessing

It can happen that the postprocessing fails due to too long paths to the algorithm data. Unfortunately, the error you get in this case does not indicate directly to the problem but only tells that a certain file could not be read. Please try to shorten the folder names in such a case.

Font issues in PDFs

We have occasionally observed some font issues in the pdfs, produced by the postprocessing of COCO (see also issue #1335). Changing to another matplotlib version solved the issue at least temporarily.

BibTeX under Mac

Under the Mac operating system, bibtex seems to be messed up a bit with respect to absolute and relative paths which causes problems with the test of the postprocessing via python do.py test-postprocessing. Note that there is typically nothing to fix if you compile the LaTeX templates "by hand" or via your LaTeX IDE. But to make the
python do.py test-postprocessing work, you will have to add a line with openout_any = a to your texmf.cnf file in the local TeX path. Type kpsewhich texmf.cnf to find out where this file actually is.

Algorithm appears twice in the figures

Earlier versions of cocopp have written extracted data to a folder named _extracted_.... If the post-processing is invoked with a * argument, these folders become an argument and are displayed (most likely additionally to the original algorithm data folder). Solution: remove the _extracted_... folders and use the latest version of the post-processing module cocopp (since release 2.1.1).

Implementation Details

  • The C code features an object oriented implementation, where the coco_problem_t is the most central data structure / object. coco.h, example_experiment.c and coco_internal.h are probably the best pointers to start to investigate the code (but see also here). coco_problem_t defines a benchmark function instance (in a given dimension), and is called via coco_evaluate_function.

  • Building, running, and testing of the code is done by merging/amalgamation of all C-code into a single C file, coco.c, and coco.h. (by calling do.py, see above). Like this it becomes very simple to include/use the code in different projects.

  • Cython is used to compile the C to Python interface in build/python/interface.pyx. The Python module installation file setup.py uses the compiled interface.c, if interface.pyx has not changed. For this reason, Cython is not a requirement for the end-user.

Citation

You may cite this work in a scientific context as

N. Hansen, A. Auger, R. Ros, O. Mersmann, T. Tušar, D. Brockhoff. COCO: A Platform for Comparing Continuous Optimizers in a Black-Box Setting, Optimization Methods and Software, 36(1), pp. 114-144, 2021. [pdf, arXiv]

@ARTICLE{hansen2021coco,
author = {Hansen, N. and Auger, A. and Ros, R. and Mersmann, O.
          and Tu{\v s}ar, T. and Brockhoff, D.},
title = {{COCO}: A Platform for Comparing Continuous Optimizers 
          in a Black-Box Setting},
journal = {Optimization Methods and Software},
doi = {https://doi.org/10.1080/10556788.2020.1808977},
pages = {114--144},
issue = {1},
volume = {36},
year = 2021
}

Links About the Workshops and Data

Comprehensive List of Documentations