metaheuristic_designer.strategies.classic.CMA_ES module

CMA-ES (Covariance Matrix Adaptation Evolution Strategy) implementation.

Warning

The current implementation is architecturally a temporary solution. It will be refactored once the EDA (Distribution-based) interface is finalized.

class CMA_ES(initializer, survivor_sel=None, name='CMA-ES', offspring_size=None, rng=None, mean=None, sigma=None, cond_tol=100000000.0, sigma_tol=1e-10, **kwargs)

Bases: EDAStrategy

Covariance Matrix Adaptation Evolution Strategy (CMA-ES).

This is a population-based algorithm that samples new solutions from a multivariate normal distribution whose mean and covariance are adapted each generation based on the best individuals.

Note

The architecture of this class is provisional. It currently overrides initialize() and perturb() directly. Once the distribution-based (EDA) abstraction is in place, CMA-ES will be rewritten to use that common interface.

Parameters:

initializerInitializer

Provides population size and genotype shape, but does not generate the initial solutions.

survivor_selSurvivorSelection, optional

How survivors are selected. Defaults to the strategy’s default (generational).

namestr, optional

Display name (default "CMA-ES").

offspring_sizeint or SchedulableParameter, optional

Number of offspring per generation. If None, the initializer’s population size is used.

rngRNGLike, optional

Random number generator.

meanVectorLike, optional

Initial mean vector. If not given, it is computed from the objective’s bounds (or randomly if no bounds exist).

sigmaVectorLike, optional

Initial step size. If not given, a default is computed.

**kwargs

Forwarded to VariablePopulation.

Attributes:

params

Access parameter values by attribute-style lookup.

population_size

Gets the amount of individuals in the population.

Parameters:

• initializer (Initializer)

• survivor_sel (SurvivorSelection)

• name (str)

• offspring_size (Optional[int | SchedulableParameter])

• mean (Optional[VectorLike])

• sigma (Optional[VectorLike])

• cond_tol (float)

• sigma_tol (float)

Methods

estimate_parameters(population)	Update the distribution parameters
extra_report()	Hook called at the end of the optimization (intended for subclasses).
extra_step_info()	Hook called after each generation (intended for subclasses).
gather_parameters()	Collect the current parameters from all sub-components.
get_params()	Return a copy of the current parameter dictionary.
get_state()	Gets the current state of the search strategy as a dictionary.
initialize(objfunc)	Create the initial population by sampling from the current distribution.
step(prev_population, objfunc)	Performs a single iteration of the algorithm on a given population.
store_kwargs([progress])	Store keyword arguments and evaluate them at the given progress.
update(progress)	Advances the state of the search by one iteration.
update_kwargs([progress])	Add or replace parameters and immediately evaluate them.

reset

initialize(objfunc)

Create the initial population by sampling from the current distribution.

Return type:

Population

Parameters:

objfuncObjectiveFunc

The objective function, used to infer bounds if mean or sigma are not provided.

Returns:

Population

A freshly sampled population with unevaluated fitness.

Parameters:

objfunc (ObjectiveFunc)

estimate_parameters(population)

Update the distribution parameters

The parents (the best μ individuals from the previous generation) are used to update mean, sigma, covariance, and the evolution paths.

Parameters:

populationPopulation

The selected parents (must be already evaluated).

Returns:

Population

Offspring population of size offspring_size.