Models a CMAChromosomeof the elitist (MO-)CMA-ES that encodes strategy parameters. More...

#include <shark/Algorithms/DirectSearch/CMA/Chromosome.h>

Public Types
enum	IndividualSuccess { Successful = 1 , Unsuccessful = 2 , Failure = 3 }

Public Member Functions
	CMAChromosome ()

	CMAChromosome (std::size_t searchSpaceDimension, double successThreshold, double initialStepSize)

void	updateAsOffspring ()
	Updates a \((\mu+1)\)-MO-CMA-ES chromosome of an successful offspring individual. It is assumed that unsuccessful individuals are not selected for future mutation.

void	updateAsParent (IndividualSuccess offspringSuccess)
	Updates a \((\mu+1)\)-MO-CMA-ES chromosome of a parent individual.

template<typename Archive >
void	serialize (Archive &archive, const unsigned int version)
	Serializes the CMAChromosometo the supplied archive.

Public Attributes
MultiVariateNormalDistributionCholesky	m_mutationDistribution
	Models the search distribution using a cholsky matrix.

RealVector	m_evolutionPath
	Low-pass filtered accumulation of successful mutative steps.

RealVector	m_lastStep
	The most recent mutative step.

RealVector	m_lastZ
	The sample from N(0,I) that produced the last step.

double	m_stepSize
	The step-size used to scale the normally-distributed mutative steps. Dynamically adapted during the run.

double	m_stepSizeDampingFactor
	Damping factor \(d\) used in the step-size update procedure.

double	m_stepSizeLearningRate
	The learning rate for the step-size.

double	m_successProbability
	Current success probability of this parameter set.

double	m_targetSuccessProbability
	Target success probability, close \( \frac{1}{5}\).

double	m_evolutionPathLearningRate
	Learning rate (constant) for updating the evolution path.

double	m_covarianceMatrixLearningRate
	Learning rate (constant) for updating the covariance matrix.

double	m_covarianceMatrixUnlearningRate
	Learning rate (constant) for unlearning unsuccessful directions from the covariance matrix.

double	m_successThreshold
	Success threshold \(p_{\text{thresh}}\) for cutting off evolution path updates.

Detailed Description

Models a CMAChromosomeof the elitist (MO-)CMA-ES that encodes strategy parameters.

Definition at line 42 of file Chromosome.h.

Member Enumeration Documentation

◆ IndividualSuccess

enum shark::CMAChromosome::IndividualSuccess

Enumerator
Successful
Unsuccessful
Failure

Definition at line 43 of file Chromosome.h.

Constructor & Destructor Documentation

◆ CMAChromosome() [1/2]

shark::CMAChromosome::CMAChromosome ( )

inline

Definition at line 65 of file Chromosome.h.

◆ CMAChromosome() [2/2]

shark::CMAChromosome::CMAChromosome	(	std::size_t	searchSpaceDimension,
		double	successThreshold,
		double	initialStepSize
	)

inline

Definition at line 66 of file Chromosome.h.

References m_covarianceMatrixLearningRate, m_covarianceMatrixUnlearningRate, m_evolutionPath, m_evolutionPathLearningRate, m_lastStep, m_lastZ, m_mutationDistribution, m_stepSizeDampingFactor, m_stepSizeLearningRate, m_successProbability, m_targetSuccessProbability, shark::MultiVariateNormalDistributionCholesky::resize(), and shark::sqr().

Member Function Documentation

◆ serialize()

template<typename Archive >

void shark::CMAChromosome::serialize	(	Archive &	archive,
		const unsigned int	version
	)

inline

Serializes the CMAChromosometo the supplied archive.

Template Parameters

Archive The type of the archive the CMAChromosomeshall be serialized to.

Parameters

[in,out]	archive	The archive to serialize to.
[in]	version	Version information (optional and not used here).

Definition at line 151 of file Chromosome.h.

References m_covarianceMatrixLearningRate, m_covarianceMatrixUnlearningRate, m_evolutionPath, m_evolutionPathLearningRate, m_lastStep, m_mutationDistribution, m_stepSize, m_stepSizeDampingFactor, m_stepSizeLearningRate, m_successProbability, m_successThreshold, and m_targetSuccessProbability.

◆ updateAsOffspring()

void shark::CMAChromosome::updateAsOffspring ( )

inline

Updates a \((\mu+1)\)-MO-CMA-ES chromosome of an successful offspring individual. It is assumed that unsuccessful individuals are not selected for future mutation.

Updates strategy parameters according to:

\begin{align*} \bar{p}_{\text{succ}} & \leftarrow & (1-c_p)\bar{p}_{\text{succ}} + c_p \\ \sigma & \leftarrow & \sigma \cdot e^{\frac{1}{d}\frac{\bar{p}_{\text{succ}} - p^{\text{target}}_{\text{succ}}}{1-p^{\text{target}}_{\text{succ}}}}\\ \vec{p}_c & \leftarrow & (1-c_c) \vec{p}_c + \mathbb{1}_{\bar{p}_{\text{succ}} < p_{\text{thresh}}} \sqrt{c_c (2 - c_c)} \vec{x}_{\text{step}} \\ \vec{C} & \leftarrow & (1-c_{\text{cov}}) \vec{C} + c_{\text{cov}} \left( \vec{p}_c \vec{p}_c^T + \mathbb{1}_{\bar{p}_{\text{succ}} \geq p_{\text{thresh}}} c_c (2 - c_c) \vec{C} \right) \end{align*}

Definition at line 100 of file Chromosome.h.

References m_covarianceMatrixLearningRate, m_evolutionPath, m_evolutionPathLearningRate, m_lastStep, m_stepSize, m_stepSizeDampingFactor, m_stepSizeLearningRate, m_successProbability, m_successThreshold, and m_targetSuccessProbability.

Referenced by shark::CMAIndividual< FitnessType >::updateAsOffspring().

◆ updateAsParent()

void shark::CMAChromosome::updateAsParent ( IndividualSuccess offspringSuccess )

inline

Updates a \((\mu+1)\)-MO-CMA-ES chromosome of a parent individual.

This is called when the parent individual survived the last selection process. The update process depends now on how the offspring fares: It can be successful, unsuccesful or a complete failure.

Based on whether it is succesful or not, the global stepsize is adapted as for the child. In the case of a failure the direction of that individual is actively purged from the Covariance matrix to make this offspring less likely.

Definition at line 123 of file Chromosome.h.

References Failure, m_covarianceMatrixUnlearningRate, m_lastStep, m_lastZ, m_stepSize, m_stepSizeDampingFactor, m_stepSizeLearningRate, m_successProbability, m_successThreshold, m_targetSuccessProbability, and Successful.