org.apache.mahout.classifier.sequencelearning.hmm

Class HmmTrainer

java.lang.Object

org.apache.mahout.classifier.sequencelearning.hmm.HmmTrainer

public final class HmmTrainer
extends Object

Class containing several algorithms used to train a Hidden Markov Model. The three main algorithms are: supervised learning, unsupervised Viterbi and unsupervised Baum-Welch.

Method Summary

Modifier and Type	Method and Description
static HmmModel	trainBaumWelch(HmmModel initialModel, int[] observedSequence, double epsilon, int maxIterations, boolean scaled) Iteratively train the parameters of the given initial model wrt the observed sequence using Baum-Welch training.
static HmmModel	trainSupervised(int nrOfHiddenStates, int nrOfOutputStates, int[] observedSequence, int[] hiddenSequence, double pseudoCount) Create an supervised initial estimate of an HMM Model based on a sequence of observed and hidden states.
static HmmModel	trainSupervisedSequence(int nrOfHiddenStates, int nrOfOutputStates, Collection<int[]> hiddenSequences, Collection<int[]> observedSequences, double pseudoCount) Create an supervised initial estimate of an HMM Model based on a number of sequences of observed and hidden states.
static HmmModel	trainViterbi(HmmModel initialModel, int[] observedSequence, double pseudoCount, double epsilon, int maxIterations, boolean scaled) Iteratively train the parameters of the given initial model wrt to the observed sequence using Viterbi training.

Methods inherited from class java.lang.Object

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

Method Detail

trainSupervised

public static HmmModel trainSupervised(int nrOfHiddenStates,
                                       int nrOfOutputStates,
                                       int[] observedSequence,
                                       int[] hiddenSequence,
                                       double pseudoCount)

Create an supervised initial estimate of an HMM Model based on a sequence of observed and hidden states.

Parameters:

nrOfHiddenStates - The total number of hidden states

nrOfOutputStates - The total number of output states

observedSequence - Integer array containing the observed sequence

hiddenSequence - Integer array containing the hidden sequence

pseudoCount - Value that is assigned to non-occurring transitions to avoid zero probabilities.

Returns:

An initial model using the estimated parameters

trainSupervisedSequence

public static HmmModel trainSupervisedSequence(int nrOfHiddenStates,
                                               int nrOfOutputStates,
                                               Collection<int[]> hiddenSequences,
                                               Collection<int[]> observedSequences,
                                               double pseudoCount)

Create an supervised initial estimate of an HMM Model based on a number of sequences of observed and hidden states.

Parameters:

nrOfHiddenStates - The total number of hidden states

nrOfOutputStates - The total number of output states

hiddenSequences - Collection of hidden sequences to use for training

observedSequences - Collection of observed sequences to use for training associated with hidden sequences.

pseudoCount - Value that is assigned to non-occurring transitions to avoid zero probabilities.

Returns:

An initial model using the estimated parameters

trainViterbi

public static HmmModel trainViterbi(HmmModel initialModel,
                                    int[] observedSequence,
                                    double pseudoCount,
                                    double epsilon,
                                    int maxIterations,
                                    boolean scaled)

Iteratively train the parameters of the given initial model wrt to the observed sequence using Viterbi training.

Parameters:

initialModel - The initial model that gets iterated

observedSequence - The sequence of observed states

pseudoCount - Value that is assigned to non-occurring transitions to avoid zero probabilities.

epsilon - Convergence criteria

maxIterations - The maximum number of training iterations

scaled - Use Log-scaled implementation, this is computationally more expensive but offers better numerical stability for large observed sequences

Returns:

The iterated model

trainBaumWelch

public static HmmModel trainBaumWelch(HmmModel initialModel,
                                      int[] observedSequence,
                                      double epsilon,
                                      int maxIterations,
                                      boolean scaled)

Iteratively train the parameters of the given initial model wrt the observed sequence using Baum-Welch training.

Parameters:

initialModel - The initial model that gets iterated

observedSequence - The sequence of observed states

epsilon - Convergence criteria

maxIterations - The maximum number of training iterations

scaled - Use log-scaled implementations of forward/backward algorithm. This is computationally more expensive, but offers better numerical stability for long output sequences.

Returns:

The iterated model