io.improbable.keanu.algorithms.variational.optimizer.gradient

Class GradientOptimizer

java.lang.Object

io.improbable.keanu.algorithms.variational.optimizer.gradient.GradientOptimizer

All Implemented Interfaces:

Optimizer

public class GradientOptimizer
extends java.lang.Object
implements Optimizer

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	GradientOptimizer.GradientOptimizerBuilder

Constructor Summary

Constructors

Constructor and Description
GradientOptimizer()

Method Summary

Modifier and Type	Method and Description
void	addFitnessCalculationHandler(java.util.function.BiConsumer<java.util.Map<VariableReference,DoubleTensor>,java.lang.Double> fitnessCalculationHandler) Adds a callback to be called whenever the optimizer evaluates the fitness of a point.
void	addGradientCalculationHandler(java.util.function.BiConsumer<java.util.Map<VariableReference,DoubleTensor>,java.util.Map<? extends VariableReference,DoubleTensor>> gradientCalculationHandler) Adds a callback to be called whenever the optimizer evaluates the gradient at a point.
static GradientOptimizer.GradientOptimizerBuilder	builder()
OptimizedResult	maxAPosteriori() This will use MAP estimation to optimize the values of latent vertices such that the probability of the whole Bayesian network is maximised.
OptimizedResult	maxLikelihood() This method will use Maximum Likelihood estimation to optimize the values of latent vertices such that the probability of the observed vertices is maximised.
void	removeFitnessCalculationHandler(java.util.function.BiConsumer<java.util.Map<VariableReference,DoubleTensor>,java.lang.Double> fitnessCalculationHandler) Removes a callback function that previously would have been called whenever the optimizer evaluated the fitness of a point.
void	removeGradientCalculationHandler(java.util.function.BiConsumer<java.util.Map<VariableReference,DoubleTensor>,java.util.Map<? extends VariableReference,DoubleTensor>> gradientCalculationHandler) Removes a callback function that previously would have been called whenever the optimizer evaluated the gradient at a point.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Methods inherited from interface io.improbable.keanu.algorithms.variational.optimizer.Optimizer

convertFromPoint, convertToArrayPoint, convertToMapPoint, createFitnessStatusBar, getAsDoubleTensors, numDimensions, toDoubleTensorVariable, totalNumberOfLatentDimensions

Constructor Detail

GradientOptimizer

public GradientOptimizer()

Method Detail

builder

public static GradientOptimizer.GradientOptimizerBuilder builder()

addGradientCalculationHandler

public void addGradientCalculationHandler(java.util.function.BiConsumer<java.util.Map<VariableReference,DoubleTensor>,java.util.Map<? extends VariableReference,DoubleTensor>> gradientCalculationHandler)

Adds a callback to be called whenever the optimizer evaluates the gradient at a point.

Parameters:

gradientCalculationHandler - a function to be called whenever the optimizer evaluates the gradient at a point. The first argument to the handler represents the point being evaluated. The second argument to the handler represents the gradient of that point.

removeGradientCalculationHandler

public void removeGradientCalculationHandler(java.util.function.BiConsumer<java.util.Map<VariableReference,DoubleTensor>,java.util.Map<? extends VariableReference,DoubleTensor>> gradientCalculationHandler)

Removes a callback function that previously would have been called whenever the optimizer evaluated the gradient at a point. If the callback is not registered then this function will do nothing.

Parameters:

gradientCalculationHandler - the function to be removed from the list of gradient evaluation callbacks

addFitnessCalculationHandler

public void addFitnessCalculationHandler(java.util.function.BiConsumer<java.util.Map<VariableReference,DoubleTensor>,java.lang.Double> fitnessCalculationHandler)

Description copied from interface: Optimizer

Adds a callback to be called whenever the optimizer evaluates the fitness of a point. E.g. for logging.

Specified by:

addFitnessCalculationHandler in interface Optimizer

Parameters:

fitnessCalculationHandler - a function to be called whenever the optimizer evaluates the fitness of a point. The first argument to the handler represents the point being evaluated. The second argument to the handler represents the fitness of that point.

removeFitnessCalculationHandler

public void removeFitnessCalculationHandler(java.util.function.BiConsumer<java.util.Map<VariableReference,DoubleTensor>,java.lang.Double> fitnessCalculationHandler)

Description copied from interface: Optimizer

Removes a callback function that previously would have been called whenever the optimizer evaluated the fitness of a point. If the callback is not registered then this function will do nothing.

Specified by:

removeFitnessCalculationHandler in interface Optimizer

Parameters:

fitnessCalculationHandler - the function to be removed from the list of fitness evaluation callbacks

maxAPosteriori

public OptimizedResult maxAPosteriori()

Description copied from interface: Optimizer

This will use MAP estimation to optimize the values of latent vertices such that the probability of the whole Bayesian network is maximised. This method will modify in place the Bayesian network that was used to create this object.

Specified by:

maxAPosteriori in interface Optimizer

Returns:

the natural logarithm of the Maximum a posteriori (MAP)

maxLikelihood

public OptimizedResult maxLikelihood()

Description copied from interface: Optimizer

This method will use Maximum Likelihood estimation to optimize the values of latent vertices such that the probability of the observed vertices is maximised. This method will modify in place the Bayesian network that was used to create this object.

Specified by:

maxLikelihood in interface Optimizer

Returns:

the natural logarithm of the maximum likelihood (MLE)