/coco

Numerical Black-Box Optimization Benchmarking Framework

Primary LanguageCOtherNOASSERTION

numbbo/coco: Comparing Continuous Optimizers

CircleCI Appveyor 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 reimplements the original Comparing Continous Optimizer platform, now rewritten fully in ANSI C and Python with other languages calling the C code. As the name suggests, the code provides a platform to benchmark and compare continuous optimizers, AKA non-linear solvers for numerical optimization. Languages currently available are

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

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:

Requirements

  1. For a machine running experiments
  • A C compiler, such as gcc
  • Python >=2.6 with setuptools installed
  • optional: git
  1. For a machine displaying data by running the post-processing

For Ubuntu 16.04+, all the requirements can be installed using the following command:

apt-get install build-essential python-dev python-numpy python-matplotlib \
                python-scipy python-six python-setuptools

Windows Specifics

Under Windows, two alternative compile toolchains can be used:

  1. Cygwin which comes with gcc and make, available in 32- and 64-bit versions.
  2. MinGW's gcc (http://www.mingw.org/ for 32-bit or https://mingw-w64.org for 64-bit machines). Make sure to update the Windows path to MinGW's make.exe and rename/link the gcc.exe to cc.exe.

For using git under Windows (optional), we recommend installing TortoiseGit.

Programming Language Specifics

Additional requirements for running an algorithm in a specific language.

  • C: make, such as GNU make (when using GNU make for Windows, make sure that your CC environment variable is set to gcc by potentially typing set CC=gcc if you see an error).
  • Java: gcc and any Java Development Kit (JDK), such that javac and javah are accessible (i.e. in the system path).
  • MATLAB: at least MATLAB 2008, for details, see here
  • Python on Windows with MinGW: Python 2.7 and the Microsoft compiler package for Python 2.7 containing VC9, available here. These are necessary to build the C extensions for the Python cocoex module for Windows. The package contains 32-bit and 64-bit compilers and the Windows SDK headers.
  • Python on Linux: python-dev must be installed to compile/install the cocoex module.
  • Octave: Octave 4.0.0 or later. On operating systems other than Windows, earlier versions might work. Under Linux the package liboctave-dev might be necessary.

Guaranties (None)

We tested the framework on Mac OSX, Ubuntu linux, Fedora linux, and Windows (XP, 7, 10) in various combinations of 32-bit and 64-bit compilers, python versions etc. Naturally, we cannot guarantee that the framework runs on any combination of operating system and software installed. In case you experience some incompatibilies, check out the Known Issues / Trouble Shooting Section below. Otherwise we will be happy if you can document them in detail on the issue tracker.

Getting Started

  1. Check out the Requirements above.

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

        pip install cocopp
    

    As long as no experiments are meant to be run, the next points 2.-6. can be skipped and continue with points 7. and 8. below.

  3. 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.

  4. 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.

  5. 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.

  6. 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.

  7. Postprocess the data from the results folder 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.

  8. 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.

    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).

  9. 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.

  10. Once your algorithm runs well, increase the budget in your experiment script, if necessary implement randomized independent restarts, and follow the above steps successively until you are happy.

  11. The experiments can be parallelized with any re-distribution of single problem instances to batches (see example_experiment2.py for an example). Each batch must write in a different target folder (this should happen automatically). Results of each batch must be kept under their separate folder as is. These folders then must be moved/copied into a single folder which becomes the input folder to the post-processing. (The post-processing searches in all subfolders and subsub... for .info files to begin with. The folder structure of a single sub-experiment must not be changed, as the .info file relies on it to find the data files.)

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.

Description by Folder

  • the do.py file in the root folder is a tool to build the entire distribution. do.py is a neat and simplifying replacement for make. It has switches for just building some languages etc, type

      python do.py
    

    to see a list of all available commandes.

  • the code-experiments/build folder is to a large extend the output folder of the ./do.py build command.

    • the exampleexperiment.??? files in the build folder are the entry points to understand the usage of the code (as end-user). They are supposed to actually be executable (in case, after compilation, which should be taken care of by do.py and/or make) and run typically random search on (some of) the provided benchmark suites.
  • documentation and examples might not be too meaningful for the time being, even though code-experiments/documentation/onion.py describes a (heavily) used design pattern (namely: inheritance) in a comparatively understandable way (though the implementation in C naturally looks somewhat different). Section Links and Documentation provides a list of pointers.

  • the code-experiments/src folder is where most of the important/interesting things happen. Many files provide comparatively decent documentation at the moment which are translated via doxygen into a more readable web page at http://numbbo.github.io/coco-doc/C/. Generally:

    • coco.h is the public interface, in particular as used in the example_experiment.c file
    • coco_internal.h provides the type definition of coco_problem_t
    • coco_suite.c is code that deals with an entire benchmark suite (i.e. a set of functions, eg. sweeping through them etc...)
    • coco_generics.c is somewhat generic code, e.g. defining a function call via coco_evaluate_function etc
    • coco_problem.c is the implementation of the coco_problem_t type/object (allocation etc).
    • observer / logger files implement data logging (as wrappers around a coco problem inheriting thereby all properties of a coco problem)
    • most other files implement more or less what they say, e.g. the actual benchmark functions, transformations, benchmark suites, etc.
    • currently, the following benchmark suites and corresponding logging facilities are supported:
      • bbob: standard single-objective BBOB benchmark suite with 24 noiseless, scalable test functions
      • bbob-biobj: a bi-objective benchmark suite, combining 10 selected functions from the bbob suite, resulting in 55 noiseless functions
      • bbob-largescale: a version of the bbob benchmark suite with dimensions 20 to 640, employing permuted block-diagonal matrices to reduce the execution time for function evaluations in higher dimension.
      • bbob-mixint: a mixed-integer version of the original bbob and bbob-largescale suites in which 80% of the variables have been discretized
      • bbob-biobj-mixint: a version of the (so far not supported) bbob-biobj-ext test suite with 92 functions with 80% discretized variables
      • toy: a simple, probably easier-to-understand example for reading and testing
  • code-experiments/tools are a few meta-tools, mainly the amalgamate.py to merge all the C code into one file

  • code-experiments/test contains unit- and integration-tests, mainly for internal use

  • code-postprocessing/cocopp contains the postprocessing code, written in python, with which algorithm data sets can be read in and the performance of the algorithms can be displayed in terms of data profiles, aRT vs. dimension plots, or simple tables.

  • code-postprocessing/helper-scripts contains additional, independent python scripts that are not part of the cocopp module but that might use it.

  • code-postprocessing/latex-templates contains LaTeX templates for displaying algorithm performances in publisher-conform PDFs for the GECCO conference.

  • code-preprocessing/archive-update/ contains internal code for combining the archives of algorithms to create/update the hypervolume reference values for the bbob-biobj test suite

  • code-preprocessing/log-reconstruction/ contains internal code for reconstructing output of the bbob-biobj logger from archive files (needed when the hypervolume reference values are updated)

  • howtos contains a few text files with generic howtos.

Known Issues / Trouble-Shooting

Java

javah call fails

If you see something like this when running python do.py run-java or build-java under Linux

COPY    code-experiments/src/coco.h -> code-experiments/build/java/coco.h
WRITE   code-experiments/build/java/REVISION
WRITE   code-experiments/build/java/VERSION
RUN     javac CocoJNI.java in code-experiments/build/java
RUN     javah CocoJNI in code-experiments/build/java
Traceback (most recent call last):
  File "do.py", line 590, in <module>
    main(sys.argv[1:])
  File "do.py", line 563, in main
    elif cmd == 'build-java': build_java()
  File "do.py", line 437, in build_java
    env = os.environ, universal_newlines = True)
  File "/..../code-experiments/tools/cocoutils.py", line 34, in check_output
    raise error
subprocess.CalledProcessError: Command '['locate', 'jni.h']' returned non-zero exit status 1

it means javah is either not installed (see above) or cannot be found in the system path, see this and possibly this for a solution.

Matlab

Path to matlab

If you see something like this when running python do.py build-matlab

AML	['code-experiments/src/coco_generics.c', 'code-experiments/src/coco_random.c', 'code-experiments/src/coco_suite.c', 'code-experiments/src/coco_suites.c', 'code-experiments/src/coco_observer.c', 'code-experiments/src/coco_runtime_c.c'] -> code-experiments/build/matlab/coco.c
COPY	code-experiments/src/coco.h -> code-experiments/build/matlab/coco.h
COPY	code-experiments/src/best_values_hyp.txt -> code-experiments/build/matlab/best_values_hyp.txt
WRITE	code-experiments/build/matlab/REVISION
WRITE	code-experiments/build/matlab/VERSION
RUN	matlab -nodisplay -nosplash -r setup, exit in code-experiments/build/matlab
Traceback (most recent call last):
  File "do.py", line 447, in <module>
    main(sys.argv[1:])
  File "do.py", line 429, in main
    elif cmd == 'build-matlab': build_matlab()
  File "do.py", line 278, in build_matlab
    run('code-experiments/build/matlab', ['matlab', '-nodisplay', '-nosplash', '-r', 'setup, exit'])
  File "/Users/auger/workviasvn/newcoco/numbbo/code-experiments/tools/cocoutils.py", line 68, in run
    universal_newlines=True)
  File "//anaconda/lib/python2.7/subprocess.py", line 566, in check_output
    process = Popen(stdout=PIPE, *popenargs, **kwargs)
  File "//anaconda/lib/python2.7/subprocess.py", line 710, in __init__
    errread, errwrite)
  File "//anaconda/lib/python2.7/subprocess.py", line 1335, in _execute_child
    raise child_exception
OSError: [Errno 2] No such file or directory

it might be because your system does not know the matlab command. To fix this, you should edit the file /etc/paths and add the path to the matlab bin file (Linux/Mac) or add the path to the folder where the matlab.exe lies to your Windows path. For instance, the etc/paths should look like something like this

/usr/local/bin
/usr/bin
/bin
/usr/sbin
/sbin
/Applications/MATLAB_R2012a.app/bin/

SMS-EMOA example does not compile under Mac

With the more complex SMS-EMOA example, the problem is related to the compilation of the external C++ hypervolume calculation in hv.cpp.

A fix for this issue consists in adding to the files hv.cpp and paretofront.c

#define char16_t UINT16_T

just before the line:

#include "mex.h"

Access to mex files denied

If it happens that you get some Access is denied errors during python do.py build-matlab or python do.py run-matlab like this one

C:\Users\dimo\Desktop\numbbo-brockho>python do.py run-matlab
Traceback (most recent call last):
  File "do.py", line 649, in <module>
    main(sys.argv[1:])
  File "do.py", line 630, in main
    elif cmd == 'run-matlab': run_matlab()
  File "do.py", line 312, in run_matlab
    os.remove( filename )
WindowsError: [Error 5] Access is denied: 'code-experiments/build/matlab\\cocoEv
aluateFunction.mexw32'

a reason can be that a previously opened Matlab window still has some file handles open. Simply close all Matlab windows (and all running Matlab processes if there is any) before to run the do.py command again.

Octave

octave-dev under Linux

When running

  python do.py run-octave

or

  python do.py build-octave

and seeing something like

   [...]
   compiling cocoCall.c...error: mkoctfile: please install the Debian package "liboctave-dev" to get the mkoctfile command

then, unsurprisingly, installing liboctave-dev like

  sudo apt-get install liboctave-dev

should do the job.

Python

setuptools is not installed

If you see something like this

$ python do.py run-python  # or build-python
[...]
PYTHON  setup.py install --user in code-experiments/build/python
ERROR: return value=1
Traceback (most recent call last):
 File "setup.py", line 8, in <module>
   import setuptools
ImportError: No module named setuptools

Traceback (most recent call last):
 File "do.py", line 562, in <module>
   main(sys.argv[1:])
 File "do.py", line 539, in main
   elif cmd == 'build-python': build_python()
 File "do.py", line 203, in build_python
   python('code-experiments/build/python', ['setup.py', 'install', '--user'])
 File "/vol2/twagner/numbbo/code-experiments/tools/cocoutils.py", line 92, in p                                         ython
   universal_newlines=True)
 File "/usr/local/lib/python2.7/subprocess.py", line 575, in check_output
   raise CalledProcessError(retcode, cmd, output=output)
subprocess.CalledProcessError: Command '['/usr/local/bin/python', 'setup.py', 'i                                        nstall', '--user']' returned non-zero exit status 1

then setuptools needs to be installed:

    pip install setuptools

or easy_install setuptools should do the job.

Compilation During Install of cocoex Fails (under Linux)

If you see something like this:

$ python do.py run-python  # or build-python
[...]
cython/interface.c -o build/temp.linux-i686-2.6/cython/interface.o
cython/interface.c:4:20: error: Python.h: file not found
cython/interface.c:6:6: error: #error Python headers needed to compile C extensions, please install development version of Python.
error: command 'gcc' failed with exit status 1

or

$ python do.py run-python  # or build-python
[...]
cython/interface.c -o build/temp.linux-x86_64-2.7/cython/interface.o
cython/interface.c:4:20: fatal error: Python.h: No such file or directory
#include "Python.h"
^
compilation terminated.
error: command 'x86_64-linux-gnu-gcc' failed with exit status 1

Under Linux

  sudo apt-get install python-dev

should do the trick.

Module Update/Install Does Not Propagate

We have observed a case where the update of the cocoex Python module seemed to have no effect. In this case it has been successful to remove all previously installed versions, see here for a few more details.

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.

Postprocessing not installable

If for some reason, your python installation is corrupted and running python do.py install-postprocessing crashes with an error message like

[...]
    safe = scan_module(egg_dir, base, name, stubs) and safe
  File "C:\Users\dimo\Anaconda2\lib\site-packages\setuptools\command\bdist_egg.py", line 392, in sca
n_module
    code = marshal.load(f)
EOFError: EOF read where object expected
[...]

try adding zip_safe=False to the setup.py.in file in the code-postprocessing folder. More details can be found in the issue #1373.

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).

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.

  • We continuously test the code through the open source automation server Jenkins on one ubuntu 12.04 machine, one OSX 10.9 machine, and two 32-bit Windows 7 machines (one with and one without cygwin).

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, 2020. [arXiv version]

@ARTICLE{hansen2020cocoplat, 
  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},
  year = 2020
}

Links and Documentation

Comprehensive List of Documentations