de.jstacs.sequenceScores.statisticalModels.trainable.hmm.models
Class HigherOrderHMM

java.lang.Object
  de.jstacs.sequenceScores.statisticalModels.trainable.AbstractTrainableStatisticalModel
      de.jstacs.sequenceScores.statisticalModels.trainable.hmm.AbstractHMM
          de.jstacs.sequenceScores.statisticalModels.trainable.hmm.models.HigherOrderHMM

All Implemented Interfaces:

SequenceScore, StatisticalModel, TrainableStatisticalModel, Storable, Cloneable

Direct Known Subclasses:

DifferentiableHigherOrderHMM, SamplingHigherOrderHMM

public class HigherOrderHMM
extends AbstractHMM

This class implements a higher order hidden Markov model. Currently, the modeling of the transitions is higher order, but is easily possible to extend this to emissions.
This implementation allows to have a set of final states . A state is denoted final states if it is allowed at the end of a path. Hence, any valid path always ends with a final state. Using the method AbstractTrainableStatisticalModel.getLogProbFor(Sequence) for sequence returns the value

Setting to all states leads to the computation of the likelihood.

Author:

Jens Keilwagen

Nested Class Summary

protected static class

HigherOrderHMM.Type This enum defined different types of computations that will be done using the backward algorithm.

Field Summary

protected double[]	backwardIntermediate Helper variable = only for internal use.
protected int[]	container Helper variable = only for internal use.
protected double[]	logEmission Helper variable = only for internal use.
protected int[][]	numberOfSummands Helper variable = only for internal use.
protected boolean	skipInit Indicates if the model should be initialized (randomly) before optimization
protected IntList	stateList Helper variable = only for internal use.

Fields inherited from class de.jstacs.sequenceScores.statisticalModels.trainable.hmm.AbstractHMM

bwdMatrix, emission, emissionIdx, finalState, forward, fwdMatrix, name, sostream, START_NODE, states, threads, trainingParameter, transition

Fields inherited from class de.jstacs.sequenceScores.statisticalModels.trainable.AbstractTrainableStatisticalModel

alphabets, length

Constructor Summary

HigherOrderHMM(HMMTrainingParameterSet trainingParameterSet, String[] name, Emission[] emission, BasicHigherOrderTransition.AbstractTransitionElement... te)
This is a convenience constructor.

HigherOrderHMM(HMMTrainingParameterSet trainingParameterSet, String[] name, int[] emissionIdx, boolean[] forward, Emission[] emission, BasicHigherOrderTransition.AbstractTransitionElement... te)
This is the main constructor.

HigherOrderHMM(StringBuffer xml)
The standard constructor for the interface Storable.

Method Summary

protected void	appendFurtherInformation(StringBuffer xml) This method appends further information to the XML representation.
protected double	baumWelch(int startPos, int endPos, double weight, Sequence seq) This method computes the likelihood and modifies the sufficient statistics according to the Baum-Welch algorithm.
HigherOrderHMM	clone() Follows the conventions of Object's clone()-method.
protected void	createHelperVariables() This method instantiates all helper variables that are need inside the model for instance for filling forward and backward matrix, ...
protected void	createStates() This method creates states for the internal usage.
protected void	estimateFromStatistics() This method estimates the parameters of all emissions and the transition using their sufficient statistics.
protected void	extractFurtherInformation(StringBuffer xml) This method extracts further information from the XML representation.
protected void	fillBwdMatrix(int startPos, int endPos, Sequence seq) This method fills the backward-matrix for a given sequence.
protected void	fillBwdOrViterbiMatrix(HigherOrderHMM.Type t, int startPos, int endPos, double weight, Sequence seq) This method computes the entries of the backward or the viterbi matrix.
protected void	fillFwdMatrix(int startPos, int endPos, Sequence seq) This method fills the forward-matrix for a given sequence.
protected void	fillLogStatePosteriorMatrix(double[][] statePosterior, int startPos, int endPos, Sequence seq, boolean silentZero) This method fills the log state posterior of Sequence seq in a given matrix.
protected void	finalize()
ResultSet	getCharacteristics() Returns some information characterizing or describing the current instance.
String	getInstanceName() Should return a short instance name such as iMM(0), BN(2), ...
double	getLogPriorTerm() Returns a value that is proportional to the log of the prior.
double	getLogProbForPath(IntList path, int startPos, Sequence seq)
double[]	getLogScoreFor(DataSet data) This method computes the logarithm of the scores of all sequences in the given sample.
void	getLogScoreFor(DataSet data, double[] res) This method computes and stores the logarithm of the scores for any sequence in the sample in the given double-array.
byte	getMaximalMarkovOrder() This method returns the maximal used Markov order, if possible.
NumericalResultSet	getNumericalCharacteristics() Returns the subset of numerical values that are also returned by SequenceScore.getCharacteristics().
Pair<IntList,Double>	getViterbiPathFor(int startPos, int endPos, Sequence seq)
protected String	getXMLTag() Returns the tag for the XML representation.
protected void	initialize(DataSet data, double[] weight) This method initializes all emissions and the transition.
protected void	initializeRandomly() This method initializes all emissions and the transition randomly.
boolean	isInitialized() This method can be used to determine whether the instance is initialized.
protected void	resetStatistics() This method resets all sufficient statistics of all emissions and the transition.
void	samplePath(IntList path, int startPos, int endPos, Sequence seq) This method samples a valid path for the given sequence seq using the internal parameters.
void	train(DataSet data, double[] weights) Trains the TrainableStatisticalModel object given the data as DataSet using the specified weights.
protected double	viterbi(IntList path, int startPos, int endPos, double weight, Sequence seq) This method computes the viterbi score of a given sequence seq.

Methods inherited from class de.jstacs.sequenceScores.statisticalModels.trainable.hmm.AbstractHMM

createMatrixForStatePosterior, decodePath, decodeStatePosterior, determineFinalStates, fromXML, getFinalStatePosterioriMatrix, getGraphvizRepresentation, getGraphvizRepresentation, getGraphvizRepresentation, getGraphvizRepresentation, getLogProbFor, getLogStatePosteriorMatrixFor, getLogStatePosteriorMatrixFor, getNumberOfStates, getNumberOfThreads, getRunTimeException, getStatePosteriorMatrixFor, getStatePosteriorMatrixFor, getViterbiPathFor, getViterbiPathsFor, initTransition, logProb, provideMatrix, setOutputStream, toString, toXML, train

Methods inherited from class de.jstacs.sequenceScores.statisticalModels.trainable.AbstractTrainableStatisticalModel

check, emitDataSet, getAlphabetContainer, getLength, getLogProbFor, getLogProbFor, getLogScoreFor, getLogScoreFor, getLogScoreFor

Methods inherited from class java.lang.Object

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

Field Detail

container

protected int[] container

Helper variable = only for internal use. This array is used for Transition.fillTransitionInformation(int, int, int, int[]).

logEmission

protected double[] logEmission

Helper variable = only for internal use. This array is used for compute the emissions at each position of a sequence only once, which might be beneficial for higher order models.

See Also:

AbstractHMM.emission

backwardIntermediate

protected double[] backwardIntermediate

Helper variable = only for internal use. This array is used to compute the backward matrix. It stores intermediate results. #see numberOfSummands

numberOfSummands

protected int[][] numberOfSummands

Helper variable = only for internal use. This array is used to compute the forward and backward matrix. It stores the number of intermediate results.

stateList

protected IntList stateList

Helper variable = only for internal use. This field is used in the method samplePath(IntList, int, int, Sequence).

skipInit

protected boolean skipInit

Indicates if the model should be initialized (randomly) before optimization

Constructor Detail

HigherOrderHMM

public HigherOrderHMM(HMMTrainingParameterSet trainingParameterSet,
                      String[] name,
                      Emission[] emission,
                      BasicHigherOrderTransition.AbstractTransitionElement... te)
               throws Exception

This is a convenience constructor. It assumes that state i used emission i on the forward strand.

Parameters:

trainingParameterSet - the ParameterSet that determines the training algorithm and contains the necessary Parameters

name - the names of the states

emission - the emissions

te - the BasicHigherOrderTransition.AbstractTransitionElements building a transition

Throws:

Exception - if

• some component could not be cloned
• some the length of name, emissionIdx, or forward is not equal to the number of states
• not all emissions use the same AlphabetContainer
• the states can not be handled by the transition

See Also:

HigherOrderHMM(HMMTrainingParameterSet, String[], int[], boolean[], Emission[], de.jstacs.sequenceScores.statisticalModels.trainable.hmm.transitions.BasicHigherOrderTransition.AbstractTransitionElement...)

HigherOrderHMM

public HigherOrderHMM(HMMTrainingParameterSet trainingParameterSet,
                      String[] name,
                      int[] emissionIdx,
                      boolean[] forward,
                      Emission[] emission,
                      BasicHigherOrderTransition.AbstractTransitionElement... te)
               throws Exception

This is the main constructor.

Parameters:

trainingParameterSet - the ParameterSet that determines the training algorithm and contains the necessary Parameters

name - the names of the states

emissionIdx - the indices of the emissions that should be used for each state, if null state i will use emission i

forward - a boolean array that indicates whether the symbol on the forward or the reverse complementary strand should be used, if null all states use the forward strand

emission - the emissions

te - the BasicHigherOrderTransition.AbstractTransitionElements building a transition

Throws:

Exception - if

• some component could not be cloned
• some the length of name, emissionIdx, or forward is not equal to the number of states
• not all emissions use the same AlphabetContainer
• the states can not be handled by the transition

HigherOrderHMM

public HigherOrderHMM(StringBuffer xml)
               throws NonParsableException

The standard constructor for the interface Storable. Constructs an HigherOrderHMM out of an XML representation.

Parameters:

xml - the XML representation as StringBuffer

Throws:

NonParsableException - if the HigherOrderHMM could not be reconstructed out of the StringBuffer xml

Method Detail

createHelperVariables

protected void createHelperVariables()

Description copied from class: AbstractHMM

This method instantiates all helper variables that are need inside the model for instance for filling forward and backward matrix, ...

Specified by:

createHelperVariables in class AbstractHMM

getXMLTag

protected String getXMLTag()

Description copied from class: AbstractHMM

Returns the tag for the XML representation.

Specified by:

getXMLTag in class AbstractHMM

Returns:

the tag for the XML representation

See Also:

AbstractHMM.fromXML(StringBuffer), AbstractHMM.toXML()

appendFurtherInformation

protected void appendFurtherInformation(StringBuffer xml)

Description copied from class: AbstractHMM

This method appends further information to the XML representation. It allows subclasses to save further parameters that are not defined in the superclass.

Specified by:

appendFurtherInformation in class AbstractHMM

Parameters:

xml - the XML representation

extractFurtherInformation

protected void extractFurtherInformation(StringBuffer xml)
                                  throws NonParsableException

This method extracts further information from the XML representation. It allows subclasses to cast further parameters that are not defined in the superclass.

Specified by:

extractFurtherInformation in class AbstractHMM

Parameters:

xml - the XML representation

Throws:

NonParsableException - if the information could not be reconstructed out of the StringBuffer xml

clone

public HigherOrderHMM clone()
                     throws CloneNotSupportedException

Description copied from class: AbstractTrainableStatisticalModel

Follows the conventions of Object's clone()-method.

Specified by:

clone in interface SequenceScore

Specified by:

clone in interface TrainableStatisticalModel

Overrides:

clone in class AbstractHMM

Returns:

an object, that is a copy of the current AbstractTrainableStatisticalModel (the member-AlphabetContainer isn't deeply cloned since it is assumed to be immutable). The type of the returned object is defined by the class X directly inherited from AbstractTrainableStatisticalModel. Hence X's clone()-method should work as:
1. Object o = (X)super.clone();
2. all additional member variables of o defined by X that are not of simple data-types like int, double, ... have to be deeply copied
3. return o

Throws:

CloneNotSupportedException - if something went wrong while cloning

createStates

protected void createStates()

Description copied from class: AbstractHMM

This method creates states for the internal usage.

Specified by:

createStates in class AbstractHMM

getLogPriorTerm

public double getLogPriorTerm()

Description copied from interface: StatisticalModel

Returns a value that is proportional to the log of the prior. For maximum likelihood (ML) 0 should be returned.

Returns:

a value that is proportional to the log of the prior

getLogProbForPath

public double getLogProbForPath(IntList path,
                                int startPos,
                                Sequence seq)
                         throws Exception

Specified by:

getLogProbForPath in class AbstractHMM

Parameters:

path - the given state path

startPos - the start position within the sequence(s) (inclusive)

seq - the sequence(s)

Returns:

the logarithm of the probability for the given path and the given sequence(s)

Throws:

Exception - if the probability for the sequence given path could not be computed, for instance if the model is not trained, ...

fillLogStatePosteriorMatrix

protected void fillLogStatePosteriorMatrix(double[][] statePosterior,
                                           int startPos,
                                           int endPos,
                                           Sequence seq,
                                           boolean silentZero)
                                    throws Exception

Description copied from class: AbstractHMM

This method fills the log state posterior of Sequence seq in a given matrix.

Specified by:

fillLogStatePosteriorMatrix in class AbstractHMM

Parameters:

statePosterior - the matrix for the log state posterior

startPos - the start position

endPos - the end position

seq - the sequence

silentZero - true if the state posterior for silent states is defined to be zero, otherwise false

Throws:

Exception - if an error occurs during the computation

See Also:

AbstractHMM.getLogStatePosteriorMatrixFor(int, int, Sequence), AbstractHMM.createMatrixForStatePosterior(int, int)

fillFwdMatrix

protected void fillFwdMatrix(int startPos,
                             int endPos,
                             Sequence seq)
                      throws OperationNotSupportedException,
                             WrongLengthException

Description copied from class: AbstractHMM

This method fills the forward-matrix for a given sequence.

Specified by:

fillFwdMatrix in class AbstractHMM

Parameters:

startPos - the start position (inclusive) in the sequence

endPos - the end position (inclusive) in the sequence

seq - the sequence

Throws:

OperationNotSupportedException

WrongLengthException

fillBwdMatrix

protected void fillBwdMatrix(int startPos,
                             int endPos,
                             Sequence seq)
                      throws Exception

Description copied from class: AbstractHMM

This method fills the backward-matrix for a given sequence.

Specified by:

fillBwdMatrix in class AbstractHMM

Parameters:

startPos - the start position (inclusive) in the sequence

endPos - the end position (inclusive) in the sequence

seq - the sequence

Throws:

Exception - if some error occurs during the computation

fillBwdOrViterbiMatrix

protected void fillBwdOrViterbiMatrix(HigherOrderHMM.Type t,
                                      int startPos,
                                      int endPos,
                                      double weight,
                                      Sequence seq)
                               throws Exception

This method computes the entries of the backward or the viterbi matrix. Additionally it allows to modify the sufficient statistics as needed for Baum-Welch training.

Parameters:

t - a switch to decide which computation mode

startPos - start position of the sequence

endPos - end position of the sequence

weight - the given external weight of the sequence (only used for Baum-Welch)

seq - the sequence

Throws:

Exception - forwarded from TrainableState.addToStatistic(int, int, double, de.jstacs.data.sequences.Sequence) and State.getLogScoreFor(int, int, Sequence)

getViterbiPathFor

public Pair<IntList,Double> getViterbiPathFor(int startPos,
                                              int endPos,
                                              Sequence seq)
                                       throws Exception

Specified by:

getViterbiPathFor in class AbstractHMM

Parameters:

startPos - the start position within the sequence

endPos - the end position within the sequence

seq - the sequence

Returns:

a Pair containing the viterbi state path and the corresponding score

Throws:

Exception - if the viterbi path could not be computed, for instance if the model is not trained, ...

viterbi

protected double viterbi(IntList path,
                         int startPos,
                         int endPos,
                         double weight,
                         Sequence seq)
                  throws Exception

This method computes the viterbi score of a given sequence seq. Furthermore, it allows either to modify the sufficient statistics according to the viterbi training algorithm or to compute the viterbi path, which will in this case be returned in path.

Parameters:

path - if null viterbi training, otherwise computation of the viterbi path

startPos - the start position

endPos - the end position

weight - the sequence weight, in most cases this is 1

seq - the sequence

Returns:

the viterbi score of the sequence

Throws:

Exception - an error occurs during the computation

baumWelch

protected double baumWelch(int startPos,
                           int endPos,
                           double weight,
                           Sequence seq)
                    throws Exception

This method computes the likelihood and modifies the sufficient statistics according to the Baum-Welch algorithm.

Parameters:

startPos - the start position

endPos - the end position

weight - the sequence weight, in most cases this is 1

seq - the sequence

Returns:

the likelihood of the sequence

Throws:

Exception - an error occurs during the computation

train

public void train(DataSet data,
                  double[] weights)
           throws Exception

Description copied from interface: TrainableStatisticalModel

Trains the TrainableStatisticalModel object given the data as DataSet using the specified weights. The weight at position i belongs to the element at position i. So the array weight should have the number of sequences in the sample as dimension. (Optionally it is possible to use weight == null if all weights have the value one.)
This method should work non-incrementally. That means the result of the following series: train(data1); train(data2) should be a fully trained model over data2 and not over data1+data2. All parameters of the model were given by the call of the constructor.

Parameters:

data - the given sequences as DataSet

weights - the weights of the elements, each weight should be non-negative

Throws:

Exception - if the training did not succeed (e.g. the dimension of weights and the number of sequences in the sample do not match)

See Also:

DataSet.getElementAt(int), DataSet.ElementEnumerator

initialize

protected void initialize(DataSet data,
                          double[] weight)
                   throws Exception

This method initializes all emissions and the transition. The initialization might use the data, but the default implementation refers to initializeRandomly().

Parameters:

data - the data set

weight - the weights for each sequence of the data set

Throws:

Exception - if an error occurs during the initialization

initializeRandomly

protected void initializeRandomly()

This method initializes all emissions and the transition randomly.

resetStatistics

protected void resetStatistics()

This method resets all sufficient statistics of all emissions and the transition.

estimateFromStatistics

protected void estimateFromStatistics()

This method estimates the parameters of all emissions and the transition using their sufficient statistics.

getMaximalMarkovOrder

public final byte getMaximalMarkovOrder()
                                 throws UnsupportedOperationException

Description copied from interface: StatisticalModel

This method returns the maximal used Markov order, if possible.

Specified by:

getMaximalMarkovOrder in interface StatisticalModel

Overrides:

getMaximalMarkovOrder in class AbstractTrainableStatisticalModel

Returns:

maximal used Markov order

Throws:

UnsupportedOperationException - if the model can't give a proper answer

getCharacteristics

public ResultSet getCharacteristics()
                             throws Exception

Description copied from interface: SequenceScore

Returns some information characterizing or describing the current instance. This could be e.g. the number of edges for a Bayesian network or an image showing some representation of the instance. The set of characteristics should always include the XML-representation of the instance. The corresponding result type is StorableResult.

Specified by:

getCharacteristics in interface SequenceScore

Overrides:

getCharacteristics in class AbstractTrainableStatisticalModel

Returns:

the characteristics of the current instance

Throws:

Exception - if some of the characteristics could not be defined

See Also:

StorableResult

getInstanceName

public String getInstanceName()

Description copied from interface: SequenceScore

Should return a short instance name such as iMM(0), BN(2), ...

Returns:

a short instance name

getLogScoreFor

public double[] getLogScoreFor(DataSet data)
                        throws Exception

Description copied from interface: SequenceScore

This method computes the logarithm of the scores of all sequences in the given sample. The values are stored in an array according to the index of the respective sequence in the sample.

The score for any sequence shall be computed independent of all other sequences in the sample. So the result should be exactly the same as for the method SequenceScore.getLogScoreFor(Sequence).

Specified by:

getLogScoreFor in interface SequenceScore

Overrides:

getLogScoreFor in class AbstractTrainableStatisticalModel

Parameters:

data - the sample of sequences

Returns:

an array containing the logarithm of the score of all sequences of the sample

Throws:

Exception - if something went wrong

See Also:

SequenceScore.getLogScoreFor(Sequence)

getLogScoreFor

public void getLogScoreFor(DataSet data,
                           double[] res)
                    throws Exception

Description copied from interface: SequenceScore

This method computes and stores the logarithm of the scores for any sequence in the sample in the given double-array.

The score for any sequence shall be computed independent of all other sequences in the sample. So the result should be exactly the same as for the method SequenceScore.getLogScoreFor(Sequence).

Specified by:

getLogScoreFor in interface SequenceScore

Overrides:

getLogScoreFor in class AbstractTrainableStatisticalModel

Parameters:

data - the sample of sequences

res - the array for the results, has to have length data.getNumberOfElements() (which returns the number of sequences in the sample)

Throws:

Exception - if something went wrong

See Also:

SequenceScore.getLogScoreFor(Sequence), SequenceScore.getLogScoreFor(DataSet)

getNumericalCharacteristics

public NumericalResultSet getNumericalCharacteristics()
                                               throws Exception

Description copied from interface: SequenceScore

Returns the subset of numerical values that are also returned by SequenceScore.getCharacteristics().

Returns:

the numerical characteristics of the current instance

Throws:

Exception - if some of the characteristics could not be defined

isInitialized

public boolean isInitialized()

Description copied from interface: SequenceScore

This method can be used to determine whether the instance is initialized. If the instance is initialized you should be able to invoke SequenceScore.getLogScoreFor(Sequence).

Returns:

true if the instance is initialized, false otherwise

finalize

protected void finalize()
                 throws Throwable

Overrides:

finalize in class AbstractHMM

Throws:

Throwable

samplePath

public void samplePath(IntList path,
                       int startPos,
                       int endPos,
                       Sequence seq)
                throws Exception

This method samples a valid path for the given sequence seq using the internal parameters.

Parameters:

path - an IntList containing the path after using this method

startPos - the start position

endPos - the end position

seq - the sequence

Throws:

Exception - if an error occurs during computation