In every tuning phase, this strategy makes a prediction about the run time for every configuration. More...

#include <PredictiveTuning.h>

Inheritance diagram for autopas::PredictiveTuning:

Collaboration diagram for autopas::PredictiveTuning:

Public Types
using	PredictionsType = std::unordered_map< Configuration, long, ConfigHash >
	Shorthand for the type used to store predictions in.

Public Member Functions
	PredictiveTuning (double relativeOptimum, unsigned int maxTuningIterationsWithoutTest, unsigned int testsUntilFirstPrediction, ExtrapolationMethodOption extrapolationMethodOption, const std::string &outputSuffix="")
	Constructor for the PredictiveTuning that generates the search space from the allowed options.

TuningStrategyOption	getOptionType () const override
	Get this object's associated TuningStrategyOption type.

void	addEvidence (const Configuration &configuration, const Evidence &evidence) override
	Notifies the strategy about empirically collected information for the given configuration.

bool	reset (size_t iteration, size_t tuningPhase, std::vector< Configuration > &configQueue, const autopas::EvidenceCollection &evidenceCollection) override
	Reset all internal parameters to the beginning of a new tuning phase.

bool	optimizeSuggestions (std::vector< Configuration > &configQueue, const EvidenceCollection &evidenceCollection) override
	Optimizes the queue of configurations to process.

void	rejectConfiguration (const Configuration &configuration, bool indefinitely) override
	Notify the strategy about a configuration that is (currently) invalid and thus can potentially be dropped from some internal storage.

PredictiveTuning::PredictionsType	calculatePredictions (size_t iteration, size_t tuningPhase, const std::vector< Configuration > &configurations, const autopas::EvidenceCollection &evidenceCollection)
	For all given configuration use the chosen extrapolation method to calculate a prediction based on the given evidence.

Public Member Functions inherited from autopas::TuningStrategyInterface
virtual TuningStrategyOption	getOptionType () const =0
	Get this object's associated TuningStrategyOption type.

virtual void	addEvidence (const Configuration &configuration, const Evidence &evidence)
	Notifies the strategy about empirically collected information for the given configuration.

virtual bool	optimizeSuggestions (std::vector< Configuration > &configQueue, const EvidenceCollection &evidenceCollection)=0
	Optimizes the queue of configurations to process.

virtual bool	reset (size_t iteration, size_t tuningPhase, std::vector< Configuration > &configQueue, const autopas::EvidenceCollection &evidenceCollection)=0
	Reset all internal parameters to the beginning of a new tuning phase.

virtual bool	needsLiveInfo () const
	Returns whether this tuning strategy wants to get a LiveInfo object passed before a new tuning phase.

virtual void	receiveLiveInfo (const LiveInfo &info)
	Virtual method that subclasses can override to receive the LiveInfo object before a tuning phase if they return true in needsLiveInfo().

virtual void	rejectConfiguration (const Configuration &configuration, bool indefinitely)
	Notify the strategy about a configuration that is (currently) invalid and thus can potentially be dropped from some internal storage.

virtual bool	needsSmoothedHomogeneityAndMaxDensity () const
	Indicate whether the strategy needs smoothed values of homogeneity and max density.

virtual void	receiveSmoothedHomogeneityAndMaxDensity (double homogeneity, double maxDensity)
	Method to pass smoothed homogeneity and the maximal density to the tuning strategy.

Detailed Description

In every tuning phase, this strategy makes a prediction about the run time for every configuration.

Then only those are tested which have the best predictions. In the end, the configuration that performed best during testing is selected.

Predictions about run time are extrapolations from previous measurements. There are multiple extrapolation methods available. Depending on this choice a certain number of tuning phases is necessary where everything is tested.

Additional features:

Retesting of old configurations: Configurations are guaranteed to be retested after a given amount of tuning phases so that their predictions do not rely on data points that are too old.

The strategy works by having multiple sets of configurations (e.g. the whole search space, optimal search space, search space of configurations that were not tested for a long time). _currentConfig is a iterator to any of them and might be switched between the sets depending on what is currently tested.

Member Typedef Documentation

◆ PredictionsType

using autopas::PredictiveTuning::PredictionsType = std::unordered_map<Configuration, long, ConfigHash>

Shorthand for the type used to store predictions in.

Map < Configuration, PredictionValue >

Constructor & Destructor Documentation

◆ PredictiveTuning()

autopas::PredictiveTuning::PredictiveTuning	(	double	relativeOptimum,
		unsigned int	maxTuningIterationsWithoutTest,
		unsigned int	testsUntilFirstPrediction,
		ExtrapolationMethodOption	extrapolationMethodOption,
		const std::string &	outputSuffix = `""`
	)

Constructor for the PredictiveTuning that generates the search space from the allowed options.

Parameters

relativeOptimum
maxTuningIterationsWithoutTest
testsUntilFirstPrediction
extrapolationMethodOption
outputSuffix

Member Function Documentation

◆ addEvidence()

void autopas::PredictiveTuning::addEvidence	(	const Configuration &	configuration,
		const Evidence &	evidence
	)

overridevirtual

Notifies the strategy about empirically collected information for the given configuration.

All evidence is stored centrally in the AutoTuner and its EvidenceCollection is passed to the tuning strategies during optimization.

Implementing this function is only necessary if the tuning strategy processes evidence differently than EvidenceCollection.

Parameters

configuration	Configuration used to obtain the evidence.
evidence	Measurement and when it was taken.

Reimplemented from autopas::TuningStrategyInterface.

◆ calculatePredictions()

PredictiveTuning::PredictionsType autopas::PredictiveTuning::calculatePredictions	(	size_t	iteration,
		size_t	tuningPhase,
		const std::vector< Configuration > &	configurations,
		const autopas::EvidenceCollection &	evidenceCollection
	)

For all given configuration use the chosen extrapolation method to calculate a prediction based on the given evidence.

Parameters

iteration	The iteration for which to calculate the predictions.
tuningPhase	The current tuning phase number.
configurations	Configurations for which to calculate predictions.
evidenceCollection	Data on which all extrapolation is based on.

Returns: For each configuration a predictions.

◆ getOptionType()

TuningStrategyOption autopas::PredictiveTuning::getOptionType ( ) const

overridevirtual

Get this object's associated TuningStrategyOption type.

Returns: TuningStrategyOption

Implements autopas::TuningStrategyInterface.

◆ optimizeSuggestions()

bool autopas::PredictiveTuning::optimizeSuggestions	(	std::vector< Configuration > &	configQueue,
		const EvidenceCollection &	evidenceCollection
	)

overridevirtual

Optimizes the queue of configurations to process.

This function is called once before each iteration in a tuning phase so all tuning strategies can give their input on which configuration to try next. This is done by reordering configQueue so that the next configuration to try is at the end (FIFO).

Parameters

configQueue	Queue of configurations to be tested. The tuning strategy should edit this queue.
evidenceCollection	All collected evidence until now.

Returns: boolean value to signal if the tuning strategy has intentionally wiped the config queue

Implements autopas::TuningStrategyInterface.

◆ rejectConfiguration()

void autopas::PredictiveTuning::rejectConfiguration	(	const Configuration &	configuration,
		bool	indefinitely
	)

overridevirtual

Notify the strategy about a configuration that is (currently) invalid and thus can potentially be dropped from some internal storage.

Parameters

configuration
indefinitely	Whether the given configuration will never be valid

Reimplemented from autopas::TuningStrategyInterface.

◆ reset()

bool autopas::PredictiveTuning::reset	(	size_t	iteration,
		size_t	tuningPhase,
		std::vector< Configuration > &	configQueue,
		const autopas::EvidenceCollection &	evidenceCollection
	)

overridevirtual

Reset all internal parameters to the beginning of a new tuning phase.

This can also mean to reorder the configQueue to some initially expected state.

Parameters

iteration	Gives the current iteration to the tuning strategy.
tuningPhase	Gives the current tuning phase to the tuning strategy.
configQueue	Queue of configurations to be tested. The tuning strategy should edit this queue.
evidenceCollection	All collected evidence until now.

Returns: boolean value to signal if the tuning strategy has intentionally wiped the config queue

Implements autopas::TuningStrategyInterface.

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

autopas/tuning/tuningStrategy/PredictiveTuning.h
autopas/tuning/tuningStrategy/PredictiveTuning.cpp

Public Types

Public Member Functions

Detailed Description

Member Typedef Documentation

◆ PredictionsType

Constructor & Destructor Documentation

◆ PredictiveTuning()

Member Function Documentation

◆ addEvidence()

◆ calculatePredictions()

◆ getOptionType()

◆ optimizeSuggestions()

◆ rejectConfiguration()

◆ reset()