autopas::GaussianProcess Class Reference

Gaussian process is a stochastical model. More...

#include <GaussianProcess.h>

Public Member Functions
	GaussianProcess (size_t dims, double sigma, Random &rngRef)
	Constructor.
void	setDimension (size_t dims)
	Change input dimension.
void	clear ()
	Discard all evidence.
size_t	numEvidence () const
	Get the number of evidence provided.
std::pair< const std::vector< Vector > &, const Vector & >	getEvidence () const
	Get all currently stored evidence.
void	addEvidence (const Vector &input, double output, bool tuneHypers)
	Provide a input-output pair as evidence.
const Vector &	getEvidenceMax () const
	Get the evidence with the highest output value.
double	predictMean (const Vector &input) const
	Try to predict f(x) using the evidence provided so far.
double	getDefaultVar () const
	Get the variance if evidence are ignored.
double	predictVar (const Vector &input) const
	The variance of the predicted f(x) from predictMean().
double	predictOutputPDF (const Vector &input, double output) const
	Calculate the probability density of provided output given provided input.
double	predictOutputScaledPDF (const Vector &input, double output) const
	Calculate the scaled probability density of provided output given provided input.
double	calcAcquisition (AcquisitionFunctionOption af, const Vector &input) const
	Calculates the acquisition function for given input.
Vector	sampleAquisitionMax (AcquisitionFunctionOption af, const std::vector< Vector > &samples) const
	Find the input in samples which maximizes given aquisition function.
std::vector< GaussianHyperparameters > &	getHyperparameters ()
	Get current hyperparameters.
void	setHyperparameters (const std::vector< double > &sample_means, const std::vector< double > &sample_thetas, const std::vector< autopas::GaussianProcess::Vector > &sample_dimScales)
	Set the hyperparameters: means, theta, dimScale.
void	normalizeHyperparameters ()
	Normalize weights of hyperparameters and truncate lowest weights.

Static Public Member Functions
static std::tuple< std::vector< double >, std::vector< double >, std::vector< autopas::GaussianProcess::Vector > >	generateHyperparameterSamples (size_t sampleSize, Random &rng, size_t dims, double sigma, double evidenceMinValue, double evidenceMaxValue)
	Generate hyperparameter samples.

Detailed Description

Gaussian process is a stochastical model.

It predicts the output of a blackbox function f(x) for given input x. To do so, some sample input-output pairs (x,f(x)) should be provided as evidence.

Currently the squared exponential kernel is used. TODO: maybe offer some options.

Constructor & Destructor Documentation

GaussianProcess()

autopas::GaussianProcess::GaussianProcess	(	size_t	dims,
		double	sigma,
		Random &	rngRef
	)

Constructor.

Parameters

dims	number of input dimensions
sigma	fixed noise
rngRef	reference to rng

Member Function Documentation

addEvidence()

void autopas::GaussianProcess::addEvidence	(	const Vector &	input,
		double	output,
		bool	tuneHypers
	)

Provide a input-output pair as evidence.

Each evidence improve the quality of future predictions.

Parameters

input	x
output	f(x)
tuneHypers	if false hyperparemeters need to be set manually

calcAcquisition()

double autopas::GaussianProcess::calcAcquisition	(	AcquisitionFunctionOption	af,
		const Vector &	input
	)		const

Calculates the acquisition function for given input.

Parameters

af	acquisition function a:input->double
input	i

Returns

a(i). This value can be compared with values a(x) of other inputs x to weigh which input would give the most gain if its evidence were provided.

generateHyperparameterSamples()

std::tuple< std::vector< double >, std::vector< double >, std::vector< autopas::GaussianProcess::Vector > > autopas::GaussianProcess::generateHyperparameterSamples ( size_t  sampleSize, autopas::Random &  rng, size_t  dims, double  sigma, double  evidenceMinValue, double  evidenceMaxValue  )

static

Generate hyperparameter samples.

Parameters

sampleSize	size
rng	random number generator
dims	number of input dimension
sigma	fixed noise
evidenceMinValue	current lowest evidence output
evidenceMaxValue	current highest evidence output

Returns

tuple [means, thetas, dimScales]

getDefaultVar()

double autopas::GaussianProcess::getDefaultVar

(

)

const

Get the variance if evidence are ignored.

Returns

getEvidence()

std::pair< const std::vector< autopas::GaussianProcess::Vector > &, const autopas::GaussianProcess::Vector & > autopas::GaussianProcess::getEvidence

(

)

const

Get all currently stored evidence.

Returns

pair of inputs and outputs

getEvidenceMax()

const autopas::GaussianProcess::Vector & autopas::GaussianProcess::getEvidenceMax

(

)

const

Get the evidence with the highest output value.

Returns

input of max

getHyperparameters()

std::vector< autopas::GaussianHyperparameters > & autopas::GaussianProcess::getHyperparameters

(

)

Get current hyperparameters.

Returns

numEvidence()

size_t autopas::GaussianProcess::numEvidence

(

)

const

Get the number of evidence provided.

Returns

predictMean()

double autopas::GaussianProcess::predictMean

(

const Vector &

input

)

const

Try to predict f(x) using the evidence provided so far.

Parameters

input

x

Returns

expected output of f(x)

predictOutputPDF()

double autopas::GaussianProcess::predictOutputPDF	(	const Vector &	input,
		double	output
	)		const

Calculate the probability density of provided output given provided input.

Parameters

input
output

Returns

predictOutputScaledPDF()

double autopas::GaussianProcess::predictOutputScaledPDF	(	const Vector &	input,
		double	output
	)		const

Calculate the scaled probability density of provided output given provided input.

The probability density is scaled such that the maximum is 1.

Parameters

input
output

Returns

predictVar()

double autopas::GaussianProcess::predictVar

(

const Vector &

input

)

const

The variance of the predicted f(x) from predictMean().

Parameters

input

x

Returns

variance

sampleAquisitionMax()

autopas::GaussianProcess::Vector autopas::GaussianProcess::sampleAquisitionMax	(	AcquisitionFunctionOption	af,
		const std::vector< Vector > &	samples
	)		const

Find the input in samples which maximizes given aquisition function.

TODO: maybe add parameters for hyperparameters of aquisition functions

Parameters

af	function to maximize
samples

Returns

setDimension()

void autopas::GaussianProcess::setDimension

(

size_t

dims

)

Change input dimension.

Current evidence will be discarded.

Parameters

dims

setHyperparameters()

void autopas::GaussianProcess::setHyperparameters	(	const std::vector< double > &	sample_means,
		const std::vector< double > &	sample_thetas,
		const std::vector< autopas::GaussianProcess::Vector > &	sample_dimScales
	)

Set the hyperparameters: means, theta, dimScale.

The samples are scored equal to the probability that given evidence and hyperparameter-sample generates given output. Hyperparameters weights should be normalized.

Parameters

sample_means
sample_thetas
sample_dimScales

---

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

• autopas/tuning/tuningStrategy/GaussianModel/GaussianProcess.h
• autopas/tuning/tuningStrategy/GaussianModel/GaussianProcess.cpp