PGAPy 1.1.6

Naming conventions in PGAPy

When you extend PGAPy – remember not all functions of PGAPack are wrapped yet and you may need additional functions – you should stick to my naming conventions when making changes. The following naming conventions were used for the wrapper:

• Constants of PGAPack like PGA_REPORT_STRING are used as-is in uppercase. These constants can be directly imported from the wrapper module. Not all constants are wrapped so far, if you need more, add them to the constdef array in pgamodule.c and send me a patch.

• For methods of the pga.PGA class I’ve removed the PGA prefix used throughout PGAPack and converted the method to lowercase with underscores between uppercase words in the original function name, so PGARun becomes run, PGACheckStoppingConditions becomes check_stopping_conditions. An exception of the lowercase-rule is whenever a name contains “GA” (for “genetic algorithm”), So PGASetMaxGAIterValue becomes max_GA_iter.

• Where possible I’ve made a single class method where PGAPack needs a separate function for each datatype, so PGAGetBinaryAllele, PGAGetCharacterAllele, PGAGetIntegerAllele, PGAGetRealAllele all become get_allele. Same holds true for set_allele.

• Whenever a name in PGApack has a “Value” or “Flag” suffix, I’ve left this out, so PGAGetFitnessCmaxValue becomes fitness_cmax and PGAGetMutationAndCrossoverFlag becomes mutation_and_crossover, the only exception to this rule is for the two functions PGAGetMutationRealValue and PGAGetMutationIntegerValue which become mutation_value not just mutation.

• Some fields can take multiple values (they are implemented by ORing integer constants, in python they are specified as a list or tuple of constants). These are converted to plural (if not already plural in PGApack), e.g., PGASetStoppingRuleType becomes stopping_rule_types.

• Internal method names in the wrapper program have a leading PGA_ – so the class method set_allele is implemented by the C-function PGA_set_allele in pgamodule.c.

Constructor Parameters

PGApack has a lot of PGASet and PGAGet functions for setting parameters. These are reflected in constructor parameters on the one hand and in read-only properties of a PGA object on the other hand. The following table gives an overview of all the original PGApack names and the names of the python wrapper. For the PGApack name I’ve only listed the PGASet function, in many cases there is a corresponding PGAGet function. If a corresponding read-only property exists for a constructor parameter this is indicated in the “Prop” column. In some cases properties are missing because no corresponding PGAGet function is implemented in PGApack, in other cases returning a numeric value that has a symbolic constant in PGApy doesn’t make much sense. The properties have the same name as the constructor parameter. There is one Property that is no constructor parameter, namely the GA_iter property that returns the current GA generation. In the type column I’m listing the Python type. If the type is followed by a number, more than one item of that type is specified (a sequence in Python). Some entries contain “sym”, these are integer values with a symbolic constant, the value “msym” indicates that several values denoted by a list of symbolic constants can be given. A special case are the PGASetRealInitRange, PGASetRealInitPercent, PGASetIntegerInitRange functions. These take two values for each allele of the gene. In python this is a sequence of 2-tuples. Note that this means that you can have different ranges of allowed values for each allele.

The num_eval property is special: Due to limitations of the C programming language, for multiple evaluations in C the first evaluation is returned as the function return-value of the evaluate function and all other parameters are returned in an auxiliary array. PGApack specifies the number of auxiliary evaluations to be returned. In Python the evaluation function can always return a sequence of evaluation values and the num_eval is one more than PGAGetNumAuxEval would return. The default for num_eval is 1.

The first two (mandatory) constructor parameters are the type of the gene (this takes a Python type, e.g., bool for a binary genome or int for an integer genome) and the length. Note that the string_length is implicitly set with the length parameter. The string_length is also available as the length of the PGA object using the Python built-in len function.

PGA Object Methods

The following are the methods that can be used during the run of the genetic search. The run method is used to start the search. This can be used, to, e.g., set an allele during hill-climbing in a custom endofgen method. Note that some methods only apply to certain gene types, e.g. the encode_int_ methods can only be used on binary alleles (they encode an integer value as a binary or gray code representation into the gene). Other methods take or return different types depending on the type of gene, e.g. get_allele or set_allele, they call different backend functions depending on the gene type. With the set_random_seed method, the random number generator can be re-seeded. It is usually best to seed the generator once at (before) the beginning by specifying random_seed in the constructor. For further details consult the user guide. The method get_evaluation will return a double for a single evaluation and a tuple of double for multiple evaluations (when num_eval is >1)

User-Methods

PGApack has the concept of user functions. These allow customization of different areas of a genetic algorihm. In Python they are implemented as methods that can be changed in a derived class. One of the methods that must be implemented in a derived class is the evaluate function (although technically it is not a user function in PGApack). It interprets the gene and returns an evaluation value or a sequence of evaluation values if you set the num_eval constructor parameter. PGApack computes a fitness from the raw evaluation value. For some methods an up-call into the PGA class is possible, for some methods this is not possible (and in most cases not reasonable). Note that for the stop_cond method, the standard check for stopping conditions can be called with:

self.check_stopping_conditions()

The following table lists the overridable methods with their parameters (for the function signature the first parameter self is omitted). Note that in PGApack there are additional user functions that are needed for user-defined data types which are currently not exposed in Python. In the function signatures p denotes the index of the individual and pop denotes the population. If more than one individual is specified (e.g., for crossover) these can be followed by a number. For crossover c1 and c2 denote the destination individuals (children). The propability for the mutation method is a floating-point value between 0 and 1. Remember to count the number of mutations that happen, and return that value for the mutation method!

Constants

The following PGApack constants are available:

Missing Features

As already mentioned, not all functions and constants of PGAPack are wrapped yet – still for many applications the given set should be enough. If you need additional functions, you may want to wrap these and send me a patch.

Another feature of PGAPack is currently not implemented in the wrapper, the usage of custom datatypes. With PGAPack you can define your own datatypes complete with their custom implementations of the genetic algorithm functionality like crossover, mutation, etc. I don’t expect problems implementing these, though.

Reporting Bugs

Please use the Sourceforge Bug Tracker or the Github Bug Tracker and

• give a short description of what you think is the correct behaviour

• give a description of the observed behaviour

• tell me exactly what you did.

• if you can publish your source code this makes it a lot easier to debug for me

Resources

Project information and download from Sourceforge main page

or checkout from Github

or directly install via pypi.

Installation

PGApy, as the name suggests, supports parallelizing the evaluation function of the genetic algorithm. This uses the Message Passing Interface (MPI) standard.

To install a serial version (without parallel programming using MPI) you can simply install from pypi using pip. Alternatively when you have unpacked or checked out from sources you can install with:

python3 setup.py install --prefix=/usr/local

If you want a parallel version using an MPI (Message-Passing Interface) library you will have to install a parallel version of PGApack first. The easiest way to do this is to use my pgapack debian package builder from github. Clone this repository, check out the branch debian/sid, install the build dependencies, they’re listed in the file debian/control and build the debian packages using:

dpkg-buildpackage -rfakeroot

This builds pgapack debian packages for all supported MPI libraries in debian, currently these are mpich, openmpi, and lam. In addition to the MPI libraries a serial version of the pgapack library is also built. Proceed by installing the package pgapack and the MPI backend library of choice. If you don’t have a preference for an MPI library, libpgapack-openmpi is the package that uses the Debians default preferences of an MPI library.

Once a parallel version of PGApack is installed, you can install PGApy as follows: You set environment variables for the PGA_PARALLEL_VARIANT (one of mpich, openmpi, or lam) and set the PGA_MODULE to module_from_parallel_install. Finally you envoke the setup, e.g.:

export PGA_PARALLEL_VARIANT=openmpi
export PGA_MODULE=module_from_parallel_install
python3 setup.py install --prefix=/usr/local

If your MPI library is installed in a different place you should study the Extension configurations in setup.py to come up with an Extension definition that fits your installation. If your installation is interesting to more people, feel free to submit a patch that adds your Extension-configuration to the standard setup.py.

Changes

Version 1.1.6: Polynomial mutation and simulated binary crossover (SBX)

• Simulated binary crossover (SBX)

• Polynomial mutation

Version 1.1.1-1.1.5: Small PGAPack updates, fixes for non-debian

• Fix setup.py for non-debian systems

• Update to latest PGAPack with small changes

Version 1.1: Add multi-objective optimization with NSGA-II

• Wrap latest pgapack version 1.4

• This add multi-objective optimization using the Nondominated Sorting Genetic Algorithm version 2 (NSGA-II) by Deb et. al. This makes use of the previously-introduced option to return more than one value in the objective function. To use the feature you need to set the num_constraint parameter to a value that leave some of the function values returned by your evaluation function as objective function values (and not as constraints). See example in examples/multi.py.

Version 1.0: Add constraint handling

• Wrap latest pgapack version 1.3

• This adds auxiliary evaluations. Now your evaluation function can return multiple floating-point values as a sequence if you set the num_eval paramter >1 in the constructor. Currently additional evaluation values are used for constraint handling. Constraint values are minimized. Once they reach zero or a negative value they no longer count: The sum of all positive constraints is the overall constraint violation. For details see paper by Deb, 2000, see user guide for citation. If you’re not using constraints, nothing in your code needs changes.

• This release may change the path an optimization takes. So for the same seed of the random number generator you will get a different result, at least if during the search there are individuals with the same evaluation (and different genetic material). This is due to a change of the sort function in pgapack (it switched to a stable sort from the C standard library).

Version 0.9: Allow installation of parallel version

• Pass argv (or sys.argv) to PGACreate

• Add a stanza to setup.py to allow a parallel installation with a given pgapack variant compiled for an MPI library. This currently needs a pre-installed pgapack debian package.

Version 0.8: Bugfix in real mutation

• Fix a core-dump in the latest pgapack

Version 0.7: Major changes in wrapping

• Now Differential Evolution is implemented, see the minfloat example and the user guide of pgapack.

Version 0.6: Major changes in wrapping

• Now the wrapping uses the standard Python recommendations on how to create a custom class.

• Update documentation

• Rename fitness_cmax (from fitness_cmax_value)

• Better error checking of parameters

Version 0.5: Bug-fix release

• Now the setup.py works, previous version had an encoding problem

• Wrap some minor new methods

• Bug-fix in PGAPack truncation selection

Version 0.4: Bundle PGAPack

• The PGAPack package is now included as a git submodule. By default we build against this library

• License fixes: The module long shipped a COPYING file that includes the 2-clause BSD license. But the headers of setup.py and pgamodule.c still included another license. This has been corrected.

Version 0.3: Feature enhancements, Bug fixes

Port to Python3, Python2 is still supported, license change.

• C-Code of wrapper updated to support both, Python2 and Python3

• Update documentation

• Fix some memory leaks that could result when errors occurred during some callback methods

• License change: We now have the 2-clause BSD license (similar to the MPICH license of PGAPack), this used to be LGPL.

Version 0.2: Feature enhancements, Bug fixes

64 bit support, more PGAPack functions and attributes wrapped, Readme-update: Sourceforge logo, Changes chapter.

• Bug-fixes for 64 bit architectures

• More functions and attributes of PGAPack wrapped

• Add a build-rule to setup.py to allow building for standard-install of PGAPack – this currently needs editing of setup.py – should use autodetect here but this would require that I set up a machine with standard install of PGAPack for testing.

• Add Sourceforge logo as required

• Add Changes chapter for automagic releases

• Add the __module__ string to class PGA in module pga. Now calling:: help (pga) in python works as expected, previously no help-text was given for the included module

Version 0.1: Initial freshmeat announcement

PGAPy is a wrapper for PGAPack, the parallel genetic algorithm library, a powerful genetic algorithm library. PGAPy wraps this library for use with Python. Pgapack is one of the most complete and accurate genetic algorithm implementations out there with a lot of features for experimentation.

• Initial Release