A general purpose Library for Evolutionary Algorithms in Python.
New Features
leap_ec.distrib.asynchronous;
note that the API may change in the futureAPI changes
CGPDecoder.initialize() method for convenience, offering a default genome initializern_ary_crossover and uniform_crossover functions with classes NAryCrossover and UniformCrossover
persist_children flag to crossover operators, which allows offspring pairs
to be used with steady-state algorithmsuuid field to the Individual base class, and Individual now also tracks parent & offspring UUIDs; this
moved UUID support from DistributedIndividual
parents attribute to Individual base class that tracks the
UUIDs of the parents via clone or crossoverPopulationMetricsPlotProbe and FitnessPlotProbe
bounds for mutation operators; previously was inconsistent nomenclature between
hard_bounds and bounds
New Features
FitnessOffsetProblem convenience wrapper to the problem moduleParabaloidProblem and QuadraticFamilyProblem to the real_rep.problems moduleCGPWithParametersDecoder
ImageXYProblem to executable_rep.problems, and a cgp_images.py example demonstrating itmutate_gaussian() to allow transforming genes by a linear functioncheck_constraints() operator to the CGPDecoder class, to help verify custom algorithmsLeadingOnes, DeceptiveTrap, and TwoMax problems to binary_rep.problems moduleSumPhenotypePlotProbe, and a new example using it to visualizing MaxOnes-style problemsmultiobjective sub-package that provides support for NSGA-II
multiobjective.nsga2.nsga2() top-level monolithic functionmultiobjective.problems.MultiObjectiveProblem is new abstract base class for multiobjective problemsmultiobjective.ops contains supporting pipeline operators, though most users will not see those if they use nsga()
API changes
Individual now has a phenome propertymutate_gaussian() and mutate_binomial()) can now be passed a list of std values to adjust the mutation width by gene.real_rep.problems.CosineFamilyProblem
reset method on PopulationMetricsPlotProbe
util.inc_generation() now takes a start_generation argumentgenome_mutate_gaussian() is now a curried function instead of a closureplot_2d_problem() and plot_2d_function() now accept extra kwargs to forward to MatplotlibMaxOnes now takes an optional target_string to generalize it to other target patternsNew features
ops.sus_selection() and ops.proportional_selection()
API changes
numpy arrays (instead of lists) the default representation for most LEAP operators and examples, for a significant speedup.indices parameter to ops.random_selection()
plot_2d_problem() now defaults to checking the problem.bounds field for xlim and ylim valuesea_solve() now accepts optional Dask Client object to enable
parallel evaluationsgenerational_ea() now supports elitism by defaultDrop support for Python 3.6
New features
landscape_features package with some initial exploratory landscape analysis toolsmutate_binomial() operator for integer representationsSimpleNeuralNetworkExecutable phenotypesHistPhenotypePlotProbe
ops.grouped_evaluate() for evaluating batches of individualsExternalProcessproblem for using external programs as fitness functionsDocumentation
leap_ec.context and updated software development
guidelines to encourage its use if tracking persistent state outside of
function calls was necessary.CI/CD
make test-slow harnessexamples/ scriptsBugfixes
viz parameter when calling simple.ea_solve()
real_rep.problems.NoisyQuarticProblem
dask that functions are impure by default, to make sure it doesn't cache resultsMakefile to use pip install -e . instead of the deprecated python setup.py develop
API changes
executable_rep.rules package to simplify learning classifier systemsleap_ec.__version__ attributehard_bounds flag to ea_solve() to tell it to respect the bounds at all times (rather than just initialization); defaults to True
Individual, Representation) into the top-level packagegenerations parameter of generational_ea() to max_generations and added an optional stop parameter for other stopping conditionsuniform_crossover operatormutate_gaussian now accepts a list of gene-wise hard boundselect_worst Boolean parameter to tournament_selection
notes columns parameter to FitnessStatsCSVProbe
pad_inputs parameter to TruthTableProblem to handle varying-dimension inputspad parameter to CartesianPhenotypePlotProbe to plot 2D projections of higher-D functionsFitnessPlotProbe as a convenience wrapper for PopulationMetricsPlotProbe
x_axis_value parameter to FitnessPlotProbe and PopulationMetricsPlotProbe
PlotTrajectoryProbe to the more descriptive CartesianPhenotypePlotProbe
PopulationPlotProbe to the more descriptive PopulationMetricsPlotProbe
leap_ec.distributed to leap_ec.distrib to reduce name space
confusion with dask.distributed
leap_ec.context to leap_ec.global_vars
Individual.decoder and Representation.decoder now uses a phenotypic representation (IdentityDecoder) by defaultp_swap = 0.2 for uniform_crossover, instead of 0.5num_points = 2 for n_ary_crossover, instead of 1context parameter on probes, so users needn't set itcontext last function argument that defaults to
leap_ec.context.context
n_ary_crossover operatorexecutable_rep.neural_network, and made it the default for examples/openai_gym.py
Executable interface to act as a Callable object (rather than using a custom output() method)statistical_helpers to assist with writing unit tests for stochastic algorithmsint_rep packageexecutable_rep.cgp, with example in examples/cgp.py
examples/multitask_island_model.py
Individual has been moved to RobustIndividual
DistributedIndividual now inherits from RobustIndividual
core.py has been broken out to separate modules
Individual and RobustIndividual now in individual.py
binary_rep
and real_rep
Representation now in representation.py
Decoder now in decoder.py
A minor release, consisting mostly of code cleaning and documentation.
An initial release, corresponding to the version of LEAP described in our GECCO 2020 workshop paper.