Package de.jstacs.sequenceScores.statisticalModels.trainable

Provides all TrainableStatisticalModels, which can be learned from a single DataSet.

See:
Description

Interface Summary

TrainableStatisticalModel

This interface defines all methods for a probabilistic model.

Class Summary

AbstractTrainableStatisticalModel	Abstract class for a model for pattern recognition.
CompositeTrainSM	This class is for modelling sequences by modelling the different positions of the each sequence by different models.
DifferentiableStatisticalModelWrapperTrainSM	This model can be used to use a DifferentiableStatisticalModel as model.
TrainableStatisticalModelFactory	This class allows to easily create some frequently used models.
UniformTrainSM	This class represents a uniform model.
VariableLengthWrapperTrainSM	This class allows to train any TrainableStatisticalModel on DataSets of Sequences with variable length if each individual length is at least SequenceScore.getLength().

Package de.jstacs.sequenceScores.statisticalModels.trainable Description

Provides all TrainableStatisticalModels, which can be learned from a single DataSet. Often, parameter learning follows a learning principle like maximum likelihood or maximum a-posteriori. Parameter learning typically is performed analytically like for the homogeneous and inhomogeneous models in the de.jstacs.sequenceScores.statisticalModels.trainable.discrete sub-package.
Notable exceptions are hidden Markov models (de.jstacs.sequenceScores.statisticalModels.trainable.hmm), which are learned by Baum-Welch or Viterbi training, and mixture models (de.jstacs.sequenceScores.statisticalModels.trainable.mixture), which are learned by expectation-maximization (EM) or Gibbs sampling.
After a TrainableStatisticalModel has been trained, it can be used to compute the likelihood of new sequences.
Any combination of TrainableStatisticalModels can be used to build a TrainSMBasedClassifier, which can be used to classify new sequences and which can be evaluated using a ClassifierAssessment.