org.ptolemy.machineLearning.hsmm

Class HSMMParameterEstimator

java.lang.Object

ptolemy.kernel.util.NamedObj

ptolemy.kernel.InstantiableNamedObj

ptolemy.kernel.Entity<T>

ptolemy.kernel.ComponentEntity<T>

ptolemy.actor.AtomicActor<TypedIOPort>

ptolemy.actor.TypedAtomicActor

org.ptolemy.machineLearning.hmm.ParameterEstimator

org.ptolemy.machineLearning.hsmm.HSMMParameterEstimator

All Implemented Interfaces:

java.lang.Cloneable, Actor, Executable, FiringsRecordable, Initializable, TypedActor, Changeable, Debuggable, DebugListener, Derivable, Instantiable, ModelErrorHandler, MoMLExportable, Moveable, Nameable

Direct Known Subclasses:

HSMMGaussianEstimator, HSMMMultinomialEstimator

public abstract class HSMMParameterEstimator
extends ParameterEstimator

This actor implements the Expectation-Maximization(EM) algorithm for parameter estimation in a family of graphical stochastic models, known as the Hidden Semi-Markov Model family.

In addition to estimating the parameters of a hidden markov model,

References

[1] Jordan, Michael I., et al. An introduction to variational methods for graphical models, Springer Netherlands, 1998.

[2] Bilmes, Jeff A. A gentle tutorial of the EM algorithm and its application to parameter estimation for Gaussian mixture and hidden Markov models. International Computer Science Institute 4.510 (1998): 126.

Since:

Ptolemy II 10.0

Version:

$Id$

Author:

Ilge Akkaya

See Also:

ParameterEstimator

Pt.AcceptedRating:

Pt.ProposedRating:

Red (ilgea)

Nested Class Summary

Nested classes/interfaces inherited from class ptolemy.kernel.Entity

Entity.ContainedObjectsIterator

Field Summary

Fields

Modifier and Type	Field and Description
protected double[][]	_D current duration distribution.
protected double[][]	_D0 initial duration distribution.
protected double[]	_dPriors0 The user defined duration priors.
protected double[]	_durationPriors Duration priors - an nStates-by-nDurations matrix.
protected int	_etaDimension Total number of categories among all observation dimensions.
protected java.util.List<java.lang.Double>	_likelihoodHistory Likelihood history over iterations.
protected int	_maxDuration maximum duration ( in time steps).
TypedIOPort	clusterAssignments Hidden-State Assignments.
protected double[][]	D_new new duration distribution.
TypedIOPort	durationEstimates DurationEstimates.
TypedIOPort	durationPriorEstimates Duration Prior estimates.
Parameter	durationProbabilities The user-provided initial guess on the prior probability distribution.
Parameter	maxStateDuration Maximum Duration.
Parameter	priorDurationDistribution The user-provided initial guess on the prior probability distribution.

Fields inherited from class org.ptolemy.machineLearning.hmm.ParameterEstimator

_A0, _likelihood, _likelihoodThreshold, _nIterations, _nStates, _obsDimension, _observations, _priorIn, _priors, _randomize, _transitionMatrix, A0, input, likelihood, likelihoodOut, likelihoodThreshold, maxIterations, newEstimates, nStates, priorDistribution, priorEstimates, randomizeGuessVectors, transitionMatrix

Fields inherited from class ptolemy.actor.TypedAtomicActor

_typesValid

Fields inherited from class ptolemy.actor.AtomicActor

_actorFiringListeners, _initializables, _notifyingActorFiring, _stopRequested

Fields inherited from class ptolemy.kernel.util.NamedObj

_changeListeners, _changeLock, _changeRequests, _debugging, _debugListeners, _deferChangeRequests, _elementName, _isPersistent, _verbose, _workspace, ATTRIBUTES, CLASSNAME, COMPLETE, CONTENTS, DEEP, FULLNAME, LINKS

Fields inherited from interface ptolemy.actor.Executable

COMPLETED, NOT_READY, STOP_ITERATING

Constructor Summary

Constructors

Constructor and Description
HSMMParameterEstimator(CompositeEntity container, java.lang.String name) Construct an actor with the given container and name.

Method Summary

Modifier and Type	Method and Description
protected abstract boolean	_checkForConvergence(int i) Check whether the gradient-descent algorithm has converged.
protected boolean	_EMParameterEstimation() Expectation-Maximization, which internally executes a gradient-descent algorithm for parameter estimation.
protected void	_initializeArrays() Initialize arrays to be used in parameter estimation.
protected abstract void	_initializeEMParameters() Initialize parameters used in ExpectationMaximization here.
protected abstract void	_iterateEM() One step EM iteration.
protected abstract void	_updateEstimates() Update parameter estimates.
void	attributeChanged(Attribute attribute) React to a change in an attribute.
java.lang.Object	clone(Workspace workspace) Clone the actor into the specified workspace.
protected abstract double	durationProbability(int y, int hiddenState) Return the duration probability of duration y at hidden state.
protected abstract double	emissionProbability(double[] y, int hiddenState) Computes the emission probability.
void	fire() Do nothing.
protected java.util.HashMap	HSMMAlphaBetaRecursion(double[][] y, double[][] A, double[] prior, int[] nCategories) Java implementation of the Baum-Welch algorithm (Alpha-Beta Recursion) for parameter estimation and cluster assignment.
void	preinitialize() Create receivers and declare delay dependencies.

Methods inherited from class org.ptolemy.machineLearning.hmm.ParameterEstimator

HMMAlphaBetaRecursion

Methods inherited from class ptolemy.actor.TypedAtomicActor

_containedTypeConstraints, _customTypeConstraints, _defaultTypeConstraints, _fireAt, _fireAt, attributeTypeChanged, clone, isBackwardTypeInferenceEnabled, newPort, typeConstraintList, typeConstraints

Methods inherited from class ptolemy.actor.AtomicActor

_actorFiring, _actorFiring, _declareDelayDependency, addActorFiringListener, addInitializable, connectionsChanged, createReceivers, declareDelayDependency, getCausalityInterface, getDirector, getExecutiveDirector, getManager, initialize, inputPortList, isFireFunctional, isStrict, iterate, newReceiver, outputPortList, postfire, prefire, pruneDependencies, recordFiring, removeActorFiringListener, removeDependency, removeInitializable, setContainer, stop, stopFire, terminate, wrapup

Methods inherited from class ptolemy.kernel.ComponentEntity

_adjustDeferrals, _checkContainer, _getContainedObject, _propagateExistence, getContainer, instantiate, isAtomic, isOpaque, moveDown, moveToFirst, moveToIndex, moveToLast, moveUp, propagateExistence, setName

Methods inherited from class ptolemy.kernel.Entity

_addPort, _description, _exportMoMLContents, _removePort, _validateSettables, connectedPortList, connectedPorts, containedObjectsIterator, getAttribute, getPort, getPorts, linkedRelationList, linkedRelations, portList, removeAllPorts, setClassDefinition, uniqueName

Methods inherited from class ptolemy.kernel.InstantiableNamedObj

_setParent, exportMoML, getChildren, getElementName, getParent, getPrototypeList, isClassDefinition, isWithinClassDefinition

Methods inherited from class ptolemy.kernel.util.NamedObj

_addAttribute, _adjustOverride, _attachText, _cloneFixAttributeFields, _containedDecorators, _copyChangeRequestList, _debug, _debug, _debug, _debug, _debug, _executeChangeRequests, _getIndentPrefix, _isMoMLSuppressed, _markContentsDerived, _notifyHierarchyListenersAfterChange, _notifyHierarchyListenersBeforeChange, _propagateValue, _removeAttribute, _splitName, _stripNumericSuffix, addChangeListener, addDebugListener, addHierarchyListener, attributeDeleted, attributeList, attributeList, decorators, deepContains, depthInHierarchy, description, description, event, executeChangeRequests, exportMoML, exportMoML, exportMoML, exportMoML, exportMoMLPlain, getAttribute, getAttributes, getChangeListeners, getClassName, getDecoratorAttribute, getDecoratorAttributes, getDerivedLevel, getDerivedList, getDisplayName, getFullName, getModelErrorHandler, getName, getName, getSource, handleModelError, isDeferringChangeRequests, isOverridden, isPersistent, lazyContainedObjectsIterator, message, notifyOfNameChange, propagateValue, propagateValues, removeAttribute, removeChangeListener, removeDebugListener, removeHierarchyListener, requestChange, setClassName, setDeferringChangeRequests, setDerivedLevel, setDisplayName, setModelErrorHandler, setPersistent, setSource, sortContainedObjects, toplevel, toString, validateSettables, workspace

Methods inherited from class java.lang.Object

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

Methods inherited from interface ptolemy.actor.Actor

createReceivers, getCausalityInterface, getDirector, getExecutiveDirector, getManager, inputPortList, newReceiver, outputPortList

Methods inherited from interface ptolemy.actor.Executable

isFireFunctional, isStrict, iterate, postfire, prefire, stop, stopFire, terminate

Methods inherited from interface ptolemy.actor.Initializable

addInitializable, initialize, removeInitializable, wrapup

Methods inherited from interface ptolemy.kernel.util.Nameable

description, getContainer, getDisplayName, getFullName, getName, getName, setName

Methods inherited from interface ptolemy.kernel.util.Derivable

getDerivedLevel, getDerivedList, propagateValue

Field Detail

priorDurationDistribution

public Parameter priorDurationDistribution

The user-provided initial guess on the prior probability distribution.

durationProbabilities

public Parameter durationProbabilities

The user-provided initial guess on the prior probability distribution.

durationEstimates

public TypedIOPort durationEstimates

DurationEstimates.

durationPriorEstimates

public TypedIOPort durationPriorEstimates

Duration Prior estimates.

clusterAssignments

public TypedIOPort clusterAssignments

Hidden-State Assignments.

maxStateDuration

public Parameter maxStateDuration

Maximum Duration.

_durationPriors

protected double[] _durationPriors

Duration priors - an nStates-by-nDurations matrix.

D_new

protected double[][] D_new

new duration distribution.

_D0

protected double[][] _D0

initial duration distribution.

_D

protected double[][] _D

current duration distribution.

_dPriors0

protected double[] _dPriors0

The user defined duration priors. Note that while _durationPriors gets updated on Baum-Welch Iterations, this array stays set at the initial user setting such that if the algorithm fails to converge, the value of _durationPriors is rolled back to _dPriors0.

_maxDuration

protected int _maxDuration

maximum duration ( in time steps).

_etaDimension

protected int _etaDimension

Total number of categories among all observation dimensions.

_likelihoodHistory

protected java.util.List<java.lang.Double> _likelihoodHistory

Likelihood history over iterations. Used for debugging at the moment.

Constructor Detail

HSMMParameterEstimator

public HSMMParameterEstimator(CompositeEntity container,
                              java.lang.String name)
                       throws NameDuplicationException,
                              IllegalActionException

Construct an actor with the given container and name.

Parameters:

container - The container.

name - The name of this actor

Throws:

IllegalActionException - If the actor cannot be contained by the proposed container.

NameDuplicationException - If the container already has an actor with this name.

Method Detail

attributeChanged

public void attributeChanged(Attribute attribute)
                      throws IllegalActionException

Description copied from class: NamedObj

React to a change in an attribute. This method is called by a contained attribute when its value changes. In this base class, the method does nothing. In derived classes, this method may throw an exception, indicating that the new attribute value is invalid. It is up to the caller to restore the attribute to a valid value if an exception is thrown.

Overrides:

attributeChanged in class ParameterEstimator

Parameters:

attribute - The attribute that changed.

Throws:

IllegalActionException - If the change is not acceptable to this container (not thrown in this base class).

clone

public java.lang.Object clone(Workspace workspace)
                       throws java.lang.CloneNotSupportedException

Description copied from class: TypedAtomicActor

Clone the actor into the specified workspace.

Overrides:

clone in class ParameterEstimator

Parameters:

workspace - The workspace for the new object.

Returns:

A new actor.

Throws:

java.lang.CloneNotSupportedException - If a derived class contains an attribute that cannot be cloned.

See Also:

NamedObj.exportMoML(Writer, int, String), NamedObj.setDeferringChangeRequests(boolean)

preinitialize

public void preinitialize()
                   throws IllegalActionException

Description copied from class: AtomicActor

Create receivers and declare delay dependencies. Derived classes can override this method to perform additional initialization functions, but they should call this base class methods or create the receivers themselves. This method gets executed exactly once prior to any other action methods. It cannot produce output data since type resolution is typically not yet done. It also gets invoked prior to any static scheduling that might occur in the domain, so it can change scheduling information.

Specified by:

preinitialize in interface Initializable

Overrides:

preinitialize in class AtomicActor<TypedIOPort>

Throws:

IllegalActionException - Not thrown in this base class.

fire

public void fire()
          throws IllegalActionException

Description copied from class: AtomicActor

Do nothing. Derived classes override this method to define their primary run-time action.

Specified by:

fire in interface Executable

Overrides:

fire in class ParameterEstimator

Throws:

IllegalActionException - Not thrown in this base class.

_EMParameterEstimation

protected boolean _EMParameterEstimation()

Description copied from class: ParameterEstimator

Expectation-Maximization, which internally executes a gradient-descent algorithm for parameter estimation.

Overrides:

_EMParameterEstimation in class ParameterEstimator

Returns:

whether parameter estimation has succeded

emissionProbability

protected abstract double emissionProbability(double[] y,
                                              int hiddenState)

Description copied from class: ParameterEstimator

Computes the emission probability. Implemented by the child class.

Specified by:

emissionProbability in class ParameterEstimator

Parameters:

y - input observation

hiddenState - index of hidden state

Returns:

P(Y=y | X=hiddenState)

durationProbability

protected abstract double durationProbability(int y,
                                              int hiddenState)

Return the duration probability of duration y at hidden state.

Parameters:

y - The duration index

hiddenState - The hidden state index

Returns:

p(y|hiddenState)

_initializeArrays

protected void _initializeArrays()
                          throws IllegalActionException

Description copied from class: ParameterEstimator

Initialize arrays to be used in parameter estimation.

Overrides:

_initializeArrays in class ParameterEstimator

Throws:

IllegalActionException - Not thrown in this base class.

_initializeEMParameters

protected abstract void _initializeEMParameters()

Description copied from class: ParameterEstimator

Initialize parameters used in ExpectationMaximization here.

Specified by:

_initializeEMParameters in class ParameterEstimator

_iterateEM

protected abstract void _iterateEM()

Description copied from class: ParameterEstimator

One step EM iteration.

Specified by:

_iterateEM in class ParameterEstimator

_checkForConvergence

protected abstract boolean _checkForConvergence(int i)

Description copied from class: ParameterEstimator

Check whether the gradient-descent algorithm has converged.

Specified by:

_checkForConvergence in class ParameterEstimator

Parameters:

i - Current iteration index

Returns:

boolean indicating whether algorithm has converged

_updateEstimates

protected abstract void _updateEstimates()

Description copied from class: ParameterEstimator

Update parameter estimates.

Specified by:

_updateEstimates in class ParameterEstimator

HSMMAlphaBetaRecursion

protected java.util.HashMap HSMMAlphaBetaRecursion(double[][] y,
                                                   double[][] A,
                                                   double[] prior,
                                                   int[] nCategories)

Java implementation of the Baum-Welch algorithm (Alpha-Beta Recursion) for parameter estimation and cluster assignment. This method uses normalized alpha values for computing the conditional probabilities of input sequences, to ensure numerical stability. Set nCategories to zero for continuous distribution types

Parameters:

y - Observation array

A - transition probability matrix

prior - State priors

nCategories - Number of categories in multinomial distribution

Returns:

a Map of parameter estimates