Step size adaptation based on the success of the new population compared to the old. More...

#include <shark/Algorithms/DirectSearch/Operators/PopulationBasedStepSizeAdaptation.h>

Public Member Functions
	PopulationBasedStepSizeAdaptation ()

double	targetSuccessRate () const

double &	targetSuccessRate ()

double	learningRate () const

double &	learningRate ()

double	dampingFactor () const

double &	dampingFactor ()

double	stepSize () const

void	init (double initialStepSize)
	Initializes a new trial by setting the initial learning rate and resetting the internal values. More...

template<class Population >
void	update (Population const &offspring)
	updates the step size using the newly sampled population More...

Protected Attributes
double	m_stepSize
	current value for the step size More...

RealVector	m_prevFitness
	fitness values of the previous iteration for ranking More...

double	m_s
	The time average of the population success. More...

double	m_targetSuccessRate

double	m_c

double	m_d

Detailed Description

Step size adaptation based on the success of the new population compared to the old.

This is the step size adaptation algorithm as proposed in Ilya Loshchilov, "A Computationally Efficient Limited Memory CMA-ES for Large Scale Optimization"

It ranks the old and new population together and checks whether the mean rank of the new population is lower than the old one in this combined population. If this is true, the step size is increased in an exponential fashion. More formally, let \( r_t(i) \) be the rank of the i-th individual in the current population in the combined ranking and \( r_{t-1}(i) \) the rank of the i-th previous individual. Then we have

\[ z_t \leftarrow \frac 1 {\lamba^2} \sum_i^{\lambda} r_{t-1}(i) - r_t(i) - z*\]

where \( z* \) is a target success value, which defaults to 0.25 this statistic is stabilised using an exponential average:

\[ s_t \leftarrow (1-c)*s_{t-1} + c*z_t \]

where the learning rate c defaults to 0.3 finally we adapt the step size sigma by

\[ \sigma_t = \sigma_{t-1} exp(s_t/d) \]

where the damping factor d defaults to 1

Definition at line 53 of file PopulationBasedStepSizeAdaptation.h.

Constructor & Destructor Documentation

◆ PopulationBasedStepSizeAdaptation()

shark::PopulationBasedStepSizeAdaptation::PopulationBasedStepSizeAdaptation ( )

inline

Definition at line 55 of file PopulationBasedStepSizeAdaptation.h.

Member Function Documentation

◆ dampingFactor() [1/2]

double shark::PopulationBasedStepSizeAdaptation::dampingFactor ( ) const

inline

Definition at line 75 of file PopulationBasedStepSizeAdaptation.h.

References m_d.

◆ dampingFactor() [2/2]

double& shark::PopulationBasedStepSizeAdaptation::dampingFactor ( )

inline

Definition at line 79 of file PopulationBasedStepSizeAdaptation.h.

References m_d.

◆ init()

void shark::PopulationBasedStepSizeAdaptation::init ( double initialStepSize )

inline

Initializes a new trial by setting the initial learning rate and resetting the internal values.

Definition at line 88 of file PopulationBasedStepSizeAdaptation.h.

References m_prevFitness, m_s, and m_stepSize.

◆ learningRate() [1/2]

double shark::PopulationBasedStepSizeAdaptation::learningRate ( ) const

inline

Definition at line 67 of file PopulationBasedStepSizeAdaptation.h.

References m_c.

◆ learningRate() [2/2]

double& shark::PopulationBasedStepSizeAdaptation::learningRate ( )

inline

Definition at line 71 of file PopulationBasedStepSizeAdaptation.h.

References m_c.

◆ stepSize()

double shark::PopulationBasedStepSizeAdaptation::stepSize ( ) const

inline

Definition at line 83 of file PopulationBasedStepSizeAdaptation.h.

References m_stepSize.

◆ targetSuccessRate() [1/2]

double shark::PopulationBasedStepSizeAdaptation::targetSuccessRate ( ) const

inline

Definition at line 59 of file PopulationBasedStepSizeAdaptation.h.

References m_targetSuccessRate.

◆ targetSuccessRate() [2/2]

double& shark::PopulationBasedStepSizeAdaptation::targetSuccessRate ( )

inline

Definition at line 63 of file PopulationBasedStepSizeAdaptation.h.

References m_targetSuccessRate.

◆ update()

template<class Population >

void shark::PopulationBasedStepSizeAdaptation::update ( Population const & offspring )

inline

updates the step size using the newly sampled population

The offspring is assumed to be ordered in ascending order by their penalizedFitness (this is the same as ordering by the unpenalized fitness in an unconstrained setting)

Definition at line 99 of file PopulationBasedStepSizeAdaptation.h.

References lambda, m_c, m_d, m_prevFitness, m_s, m_stepSize, m_targetSuccessRate, and shark::penalizedFitness().