ACM Surrogate strategy for CMA-ES, follows: 'Surrogate-Assisted Evolutionary Algorithms', Ilya Loshchilov, PhD Thesis, Universite Paris-Sud 11, 2013. http://www.loshchilov.com/phd.html see Chapter 4. More...

#include <libcmaes/surrogatestrategy.h>

Inheritance diagram for libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >:

Public Member Functions
	ACMSurrogateStrategy (FitFunc &func, CMAParameters< TGenoPheno > &parameters)
	constructor

dMat	ask ()
	Generates a set of candidate points. Uses the pre-sampling of a larger than usual number of offprings, controled by 'lambdaprime', as needed.

void	eval (const dMat &candidates, const dMat &phenocandidates=dMat(0, 0))
	Evaluates a set of candiates against the objective function or the surrogate model, as needed.

void	tell ()
	Updates the state of the stochastic search, and prepares for the next iteration by training the surrogate model, as needed.

int	optimize ()

bool	do_train () const
	whether to train the model

void	set_lambdaprime (const int &lp)
	sets the number of true objective function calls per iteration

int	get_lambdaprime () const
	returns the number of calls to the true objective function per iteration

void	set_prelambda (const int &pl)
	sets the number of pre-screened offsprings (sampled)

int	get_prelambda () const
	returns the current number of pre-screened offpsrings (sampled)

void	set_theta_sel0 (const double &s)
	sets the standard deviation of selection sampling step 0

double	get_theta_sel0 () const
	returns the standard deviation of selection sampling step 0

void	set_theta_sel1 (const double &s)
	sets the standard deviation of selection sampling step 1

double	get_theta_sel1 () const
	returns the standard deviation of selection sampling step 0

Public Member Functions inherited from libcmaes::SurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >
	SurrogateStrategy (FitFunc &func, CMAParameters< TGenoPheno > &parameters)
	constructor

int	train (const std::vector< Candidate > &candidates, const dMat &cov)
	train a surrogate model

int	predict (std::vector< Candidate > &candidates, const dMat &cov)
	predict from a surrogate model

double	compute_error (const std::vector< Candidate > &test_set, const dMat &cov=dMat(0, 0))
	compute surrogate model error (copies and sorts the test_set)

bool	do_train () const
	conditionals on training, to be specialized in inherited surrogate strategies

void	set_ftrain (const CSurrFunc &train)
	sets the training function

void	set_fpredict (const SurrFunc &predict)
	sets the prediction function

void	set_l (const int &l)
	sets the size of the training set (number of points)

int	get_l () const
	gets the size of the training set (number of points)

void	set_exploit (const bool &exploit)
	sets whether to exploit the surrogate model

bool	get_exploit () const
	gets the state of surrogate model exploitation

double	get_train_error () const
	returns the surrogate model training error

double	get_test_error () const
	returns the surrogate model test error

void	set_train_error (const double &err)
	sets training error

void	set_test_error (const double &err)
	sets the test error and updates the smoothed test err.

void	add_to_training_set (const Candidate &c)
	adds a point to the training set (candidate = points + objective function value)

void	set_nsteps (const int &nsteps)
	sets the lifelength of the surrogate, i.e. the number of steps in between to training steps

void	reset_training_set ()
	resets training set and related information, useful when using algorithms with restarts

int	get_nsteps () const
	returns the current surrogate lifelength

Protected Member Functions
void	pre_selection_eval (const dMat &candidates)
	pre-selection + candidate evaluation scheme. Called by eval, evaluates lambdaprime candidates with surrogate model, then subsample the population in order to evaluate them with the original objective function, test the surrogate model and grow the training set with new points

Protected Attributes
double	_prelambda = 500

double	_theta_sel0 = 0.4

double	_theta_sel1 = 0.8

int	_lambdaprime

Protected Attributes inherited from libcmaes::SurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >
bool	_exploit = true

int	_l = 200

std::vector< Candidate >	_tset

CSurrFunc	_train

SurrFunc	_predict

double	_train_err = 0.0

double	_test_err = 0.0

double	_smooth_test_err = 0.5

double	_beta_err = 0.2

int	_nsteps = 1

int	_auto_nsteps = false

Private Member Functions
void	init_rd ()

Private Attributes
std::random_device	_rd

std::normal_distribution< double >	_norm_sel0

std::normal_distribution< double >	_norm_sel1

std::mt19937	_gen0

std::mt19937	_gen1

Detailed Description

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>
class libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >

ACM Surrogate strategy for CMA-ES, follows: 'Surrogate-Assisted Evolutionary Algorithms', Ilya Loshchilov, PhD Thesis, Universite Paris-Sud 11, 2013. http://www.loshchilov.com/phd.html see Chapter 4.

Implements a single-objective strategy for CMA-ES and related algorithms, that samples ans pre-screens a larger than usual number of offsprings at each generation, rank them with a rank-based surrogate model, and consumes a small portion of offsprings with the original (supposedly expensive) objective function.

This strategy overrides the ask/eval/tell functions of the base optimization strategy

Constructor & Destructor Documentation

◆ ACMSurrogateStrategy()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate , class TGenoPheno >

libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::ACMSurrogateStrategy	(	FitFunc &	func,
		CMAParameters< TGenoPheno > &	parameters
	)

constructor

Parameters

func	objective function to minimize
parameters	optimization parameters

Member Function Documentation

◆ ask()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate , class TGenoPheno >

dMat libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::ask ( )

Generates a set of candidate points. Uses the pre-sampling of a larger than usual number of offprings, controled by 'lambdaprime', as needed.

Note: this function overrides the default ESOStrategy::ask

Returns: A matrix whose rows contain the candidate points.

◆ do_train()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

bool libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::do_train ( ) const

inline

whether to train the model

Returns: whether to train the model

◆ eval()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate , class TGenoPheno >

void libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::eval	(	const dMat &	candidates,
		const dMat &	phenocandidates = `dMat(0,0)`
	)

Evaluates a set of candiates against the objective function or the surrogate model, as needed.

Note: this function overrides the default CMAStrategy::eval

Parameters

candidates	A matrix whose rows contain the candidates.
phenocandidates	The candidates transformed into phenotype, leave empty if no pheno transform.

◆ get_lambdaprime()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

int libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::get_lambdaprime ( ) const

inline

returns the number of calls to the true objective function per iteration

Returns: calls per iteration to the true objective function

◆ get_prelambda()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

int libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::get_prelambda ( ) const

inline

returns the current number of pre-screened offpsrings (sampled)

Returns: number of offsprings

◆ get_theta_sel0()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

double libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::get_theta_sel0 ( ) const

inline

returns the standard deviation of selection sampling step 0

Returns: standard deviation

◆ get_theta_sel1()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

double libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::get_theta_sel1 ( ) const

inline

returns the standard deviation of selection sampling step 0

Returns: standard deviation

◆ pre_selection_eval()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate , class TGenoPheno >

void libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::pre_selection_eval ( const dMat & candidates )

protected

pre-selection + candidate evaluation scheme. Called by eval, evaluates lambdaprime candidates with surrogate model, then subsample the population in order to evaluate them with the original objective function, test the surrogate model and grow the training set with new points

Parameters

candidates A matrix whose rows contain the candidates.

◆ set_lambdaprime()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

void libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::set_lambdaprime ( const int & lp )

inline

sets the number of true objective function calls per iteration

Parameters

lp	true objective function calls per iteration

◆ set_prelambda()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

void libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::set_prelambda ( const int & pl )

inline

sets the number of pre-screened offsprings (sampled)

Parameters

number of offsprings

◆ set_theta_sel0()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

void libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::set_theta_sel0 ( const double & s )

inline

sets the standard deviation of selection sampling step 0

Parameters

s	standard deviation

◆ set_theta_sel1()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

void libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::set_theta_sel1 ( const double & s )

inline

sets the standard deviation of selection sampling step 1

Parameters

s	standard deviation

◆ tell()

template<template< class U, class V > class TStrategy, class TCovarianceUpdate , class TGenoPheno >

void libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::tell ( )

Updates the state of the stochastic search, and prepares for the next iteration by training the surrogate model, as needed.

Note: this function overrides the default CMAStrategy::tell

Member Data Documentation

◆ _lambdaprime

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

int libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::_lambdaprime

protected

true objective function calls per iteration.

◆ _prelambda

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

double libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::_prelambda = 500

protected

number of pre-screened offsprings.

◆ _theta_sel0

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

double libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::_theta_sel0 = 0.4

protected

standard deviation of selection sampling step 0.

◆ _theta_sel1

template<template< class U, class V > class TStrategy, class TCovarianceUpdate = CovarianceUpdate, class TGenoPheno = GenoPheno<NoBoundStrategy>>

double libcmaes::ACMSurrogateStrategy< TStrategy, TCovarianceUpdate, TGenoPheno >::_theta_sel1 = 0.8

protected

standard deviation of selection sampling step 1.

The documentation for this class was generated from the following files:

include/libcmaes/surrogatestrategy.h
src/surrogatestrategy.cc

Public Member Functions

Protected Member Functions

Protected Attributes

Private Member Functions

Private Attributes

Detailed Description

Constructor & Destructor Documentation

◆ ACMSurrogateStrategy()

Member Function Documentation

◆ ask()

◆ do_train()

◆ eval()

◆ get_lambdaprime()

◆ get_prelambda()

◆ get_theta_sel0()

◆ get_theta_sel1()

◆ pre_selection_eval()

◆ set_lambdaprime()

◆ set_prelambda()

◆ set_theta_sel0()

◆ set_theta_sel1()

◆ tell()

Member Data Documentation

◆ _lambdaprime

◆ _prelambda

◆ _theta_sel0

◆ _theta_sel1