pycma        

DOI BibTeX cite as:

Nikolaus Hansen, Youhei Akimoto, and Petr Baudis. CMA-ES/pycma on Github. Zenodo, DOI:10.5281/zenodo.2559635, February 2019.

---

pycma is a Python implementation of CMA-ES and some related numerical optimization tools.

The CMA-ES (Covariance Matrix Adaptation Evolution Strategy) is a randomized derivative-free numerical optimization algorithm for difficult (non-convex, ill-conditioned, multi-modal, rugged, noisy) optimization problems in continuous and mixed-integer search spaces. This package provides an implementation of the CMA-ES algorithm that includes the handling of

• bound constraints via the 'bounds' = [lower, upper] option or the cma.BoundDomainTransform wrapper
• linear and nonlinear constraints via the constraints argument to fmin2 or fmin_con2
• noise via noise_handler=True as argument to fmin2
• integer variables for mixed-integer problems via the 'integer_variables'=index_list option

Documentation and Getting Started (Links)

• A quick start guide with a few usage examples

• The notebooks folder provides example code in Jupyter notebooks, namely

  • Basic use cases notebook
  • Constraints handling notebook
  • The ask-and-tell interface
  • Noise handling notebook

• API Documentation

• Hints for how to use this (kind of) optimization module in practice

• FAQs and HowTos (under development).

Installation of the latest release

In a system shell, type

    python -m pip install cma

to install the latest release from the Python Package Index (PyPI). Type install -U instead of install to upgrade a current installation to the latest release. The release link also provides more installation hints and a quick start guide.

Installation from Github

The quick way to install the code from, for example, the development branch (this requires git to be installed):

    pip install git+https://github.com/CMA-ES/pycma.git@development

The long way:

• get the package

  • either download and unzip the code by clicking the green button above
  • or, with git installed, type git clone https://github.com/CMA-ES/pycma.git

• "install" the package

  • either copy (or move) the cma source code folder into a folder which is in the Python path (e.g. the current folder)

  • or modify the Python path to point to the folder where the cma package folder can be found. In both cases, import cma works without any further installation.

  • or install the cma package by typing

        pip install -e .
    

    in the (pycma) folder where the cma package folder can be found. Moving the cma folder away from its location invalidates this installation.

It may be necessary to replace pip with python -m pip and/or prefixing either of these with sudo.

Version History

• Release 4.5.0

  • console output now starts to overwrite the last line after 4 regular lines, see option verb_disp_overwrite which now expresses "number of lines"
  • introduce number of runs attribute in CMAEvolutionStrategyResult2
  • introduce all_stoppings and all_best module attributes with information collected from restarts
  • major revision of step-size implementations in sigma_adaptation.py
  • polishing of AugmentedLagrangian constraints handler and initialization bug fix
  • cma.evolution_strategy.TPA_dimension = 300 sets TPA as default step-size adaptation for dimension >= 300.
  • improve implementation of ridge fitness functions
  • implement low effective dimension functions and transformations
  • implement optimization_tools.EvaluationsIterator to compute non-discretized ERT values from non-f-targeted data
  • implement s.figpolish and s.grid for convenience
  • implement CountWarnings instance deliver_warnings to control the number of warnings shown

• Release 4.4.4

  • fix the fix of issue 343

• Release 4.4.3

  • Addressing issue 231, failures in corner cases with large population size, by increasing the step-size damping of CSA and TPA. This seems also to improve the performance on bbob-f24 in 10 and 20-D while worsening the performance on bbob-f23 in 10 and 40-D.
  • Provide option 'TPA_dampfac' analogous to 'CSA_dampfac'.
  • Plots now show the current best solution and the distribution mean in two subplots.
  • New:
    • Provide the two (by far) most useful statistical tests with a tidy interface in cma.utilities.math.test...
    • Provide a more_algorithms sub-package containing purecma and CompactGA.
    • Provide a provisional experimentation module (requires import cma.experimentation).
  • A few smaller fixes and improvements.

• Release 4.4.2

  • Fix compatibility issues (with comocma):
    • add back the (deprecated) cma.constraints_handling.BoundTransform class which was missing since 4.1.0. Note that cma.BoundTransform is the recommended way to access this class
    • remove dependency of OOOptimizer.optimize on self.result
  • fix issue 337 where plotting bails with some recent version of matplotlib>3.8.0.
  • Various improvements of the logger and plotting.
  • Remove default f-offset from binary test functions (cma.fitness_functions.binary_foffset = 0 by default now)
  • A few new module settings to (better) control corner case behavior.

• Release 4.4.1

  • fmin2 accepts a constraints function as constraints keyword argument
  • an improved CMAEvolutionStrategyResult2 class which also contains the best feasible solution
  • a reset_options method which also clears the current termination status
  • polish the output of .optimize() and of .result_pretty()
  • catch final .stop() value displayed with cma.plot

• Release 4.4.0

  • constraints handling is available also in the ask-and-tell interface (addressing issues #258, #287, and #167)
  • ask has an ignore_integer_variables argument to not mutate integer variables
  • an on/off switch for integer centering, cma.integer_centering.centering_on (by default True)
  • polishing and minor bug fixes
  • code internals:
    • move integer rounding code (applied to delivered solutions) to the cma.transformations.RoundIntegerVariables class
    • utils.SolutionDict can behave like a queue too

• Release 4.3.0

  • integer variables of candidate solutions are rounded (addressing also issue #286)
  • moved main docstring from fmin to fmin2
  • experimental plots for error estimates and sensitivities
  • fix numpy scalar type representations at various places
  • replace ineffective use_archives flag with archive_sent_solutions and archive_after_sent

• Release 4.2.0

  • a stand-alone boundary handling function wrapper BoundDomainTransform
  • streamline plot docs, fix symlog plot with newest matplotlib, plots display the value of .stop() and the version number
  • a few more minor fixes and improvements
  • replace setup.py with pyproject.toml
  • Version 4.1.0 (already since 5a30571f)
    • move boundary handling into a separate module
    • various small-ish fixes and improvements, in particular an edge case in the initialization of the Lagrange multipliers in the constraints handling

• Release 4.0.0

  • majorly improved mixed-integer handling based on a more concise lower bound of variances and on so-called integer centering
  • moved options and parameters code into a new file
  • many small-ish fixes and improvements

• Release 3.4.0

  • fix compatibility to numpy 2.0 (thanks to Sait Cakmak)
  • improved interface to noise_handler argument which accepts True as value
  • improved interface to ScaleCoordinates now also with lower and upper value mapping to [0, 1], see issue #210
  • changed: 'ftarget' triggers with <= instead of <
  • assign surrogate attribute (for the record) when calling fmin_lq_surr
  • various (minor) bug fixes
  • various (small) improvements of the plots and their usability
    • display iterations, evaluations and population size and termination criteria in the plots
    • subtract any recorded x from the plotted x-values by x_opt=index
  • plots are now versus iteration number instead of evaluations by default
  • provide legacy bbobbenchmarks without downloading
  • new: CMADataLogger.zip allows sharing plotting data more easily by a zip file
  • new: tolxstagnation termination condition for when the incumbent seems stuck
  • new: collect restart terminations in cma.evalution_strategy.all_stoppings
  • new: stall_sigma_change_on_divergence_iterations option to stall sigma change when the median fitness is worsening
  • new: limit active C update for integer variables
  • new: provide a COCO single function

• Release 3.3.0 implements

  • diagonal acceleration via diagonal decoding (option CMA_diagonal_decoding, by default still off).
  • fmin_lq_surr2 for running the surrogate assisted lq-CMA-ES.
  • optimization_tools.ShowInFolder to facilitate rapid experimentation.
  • verb_disp_overwrite option starts to overwrite the last line of the display output instead of continuing adding lines to avoid screen flooding with longish runs (off by default).
  • various smallish improvements, bug fixes and additional features and functions.

• Release 3.2.2 fixes some smallish interface and logging bugs in ConstrainedFitnessAL and a bug when printing a warning. Polishing mainly in the plotting functions. Added a notebook for how to use constraints.

• Release 3.2.1 fixes plot of principal axes which were shown squared by mistake in version 3.2.0.

• Release 3.2.0 provides a new interface for constrained optimization ConstrainedFitnessAL and fmin_con2 and many other minor fixes and improvements.

• Release 3.1.0 fixes the return value of fmin_con, improves its usability and provides a best_feasible attribute in CMAEvolutionStrategy, in addition to various other more minor code fixes and improvements.

• Release 3.0.3 provides parallelization with OOOptimizer.optimize(..., n_jobs=...) (fix for 3.0.1/2) and improved pickle support.

• Release 3.0.0 provides non-linear constraints handling, improved plotting and termination options and better resilience to injecting bad solutions, and further various fixes.

• Version 2.7.1 allows for a list of termination callbacks and a light copy of CMAEvolutionStrategy instances.

• Release 2.7.0 logger now writes into a folder, new fitness model module, various fixes.

• Release 2.6.1 allow possibly much larger condition numbers, fix corner case with growing more-to-write list.

• Release 2.6.0 allows initial solution x0 to be a callable.

• Version 2.4.2 added the function cma.fmin2 which, similar to cma.purecma.fmin, returns (x_best:numpy.ndarray, es:cma.CMAEvolutionStrategy) instead of a 10-tuple like cma.fmin. The result 10-tuple is accessible in es.result:namedtuple.

• Version 2.4.1 included bbob testbed.

• Version 2.2.0 added VkD CMA-ES to the master branch.

• Version 2.* is a multi-file split-up of the original module.

• Version 1.x.* is a one file implementation and not available in the history of this repository. The latest 1.* version 1.1.7 can be found here.