Class AlgorithmRVM

java.lang.Object
  Algorithm
      AlgorithmRVM

All Implemented Interfaces:

java.lang.Runnable

public class AlgorithmRVM
extends Algorithm

Algorithm Relevance Vector Machines

Constructor Summary

AlgorithmRVM()

Method Summary

void	computeDecisionRegions() Computes the line of discrimination for the classification algorithms when the corresponding flags have been initialized
void	computeErrors() Computes the classification error for the data points
double	computeLikelihood() Computes the log likelihood of the weights given the data.
boolean	computeSigma() Computes the sigma valued defined on pp.
boolean	computeVarianceCholesky() Computes the diagonal elements of the inverse from the cholesky decomposition
boolean	initFullTrain() initializes the data structures in preparation for a full training pass
boolean	initialize() Overrides the initialize() method in the base class.
boolean	irlsTrain() Completes one pass of IRLS training to update the weights given the currently assigne hyperparamters.
boolean	pruneAndUpdate() prunes off vectors whose hyperparameters have gone to infinity and updates working data sets
boolean	pruneWeights() Prunes off vectors which attain a zero weight during training.
void	run() Implementation of the run function from the Runnable interface.
boolean	trainFull() this method trains an RVM probabilistic classifier on the input data and targets provided.
boolean	updateHyperparametersFull() Updates the hyperparameter values

Methods inherited from class Algorithm

computeMeans, disableControl, enableControl, nextStep, prevStep, scaleToFitData, setDataPoints, setOutputPanel, setProcessBox

Methods inherited from class java.lang.Object

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

Constructor Detail

AlgorithmRVM

public AlgorithmRVM()

Method Detail

initialize

public boolean initialize()

Overrides the initialize() method in the base class. Initializes member data and prepares for execution of first step. This method "resets" the algorithm.

Specified by:

initialize in class Algorithm

Returns:

true

run

public void run()

Implementation of the run function from the Runnable interface. Determines what the current step is and calls the appropriate method.

Specified by:

run in interface java.lang.Runnable

Specified by:

run in class Algorithm

trainFull

public boolean trainFull()

this method trains an RVM probabilistic classifier on the input data and targets provided. the inputs and targets should have a one-to-one correspondence and all targets should be either 0 (out-of-class) or 1 (in-class). The training scheme follows that of [1] section 3. It is assumed that the class data and targets are already set when this method is called. the training algorithm in [1] for RVMs proceeds in three iterative steps 1. prune away any weights whose hyperparameters have gone to infinity 2. estimate most probable weights: in this step we find those weights that maximize equation (24) of [1]. The iteratively reweighted least squares algorithm is used to find w_mp 3. estimate the covariance of a Gaussian approximation to the posterior distribution (the posterior is what we want to model in the end) over the weights centered at the weights, w_mp. 4. estimate the hyperparameters that govern the weights. this is done by evaluating the derivative over the hyperparameters and finding the maximizing hyperparameters. 1, 2, 3, and 4 are carried out iteratively until a suitable convergence criteria is satisfied.

Returns:

boolean value indicating status

initFullTrain

public boolean initFullTrain()

initializes the data structures in preparation for a full training pass

Returns:

boolean value indicating status

updateHyperparametersFull

public boolean updateHyperparametersFull()

Updates the hyperparameter values

Returns:

boolean value indicating status

irlsTrain

public boolean irlsTrain()

Completes one pass of IRLS training to update the weights given the currently assigne hyperparamters. In this step, we find those weights that maximize equation (24) of [1]. The iteratively reweighted least squares algorithm is used to find w_mp. It proceeds as follows: a. initialize b. loop until convergence 1. compute the B matrix 2. compute the Hessian and the gradient 3. update weights with formula new weights = weights - inv(Hessian) * gradient 4. check convergence

Returns:

boolean value indicating status

pruneAndUpdate

public boolean pruneAndUpdate()

prunes off vectors whose hyperparameters have gone to infinity and updates working data sets

Returns:

true

pruneWeights

public boolean pruneWeights()

Prunes off vectors which attain a zero weight during training. Auxiliary data structures that are indexed according to the working set are also updated.

Returns:

true

computeVarianceCholesky

public boolean computeVarianceCholesky()

Computes the diagonal elements of the inverse from the cholesky decomposition

Returns:

true

computeSigma

public boolean computeSigma()

Computes the sigma valued defined on pp. 218 of [1].

Returns:

true

computeLikelihood

public double computeLikelihood()

Computes the log likelihood of the weights given the data. We would expect this value to increase as training proceeds

Returns:

Likelihood of weights given the data

computeDecisionRegions

public void computeDecisionRegions()

Computes the line of discrimination for the classification algorithms when the corresponding flags have been initialized

computeErrors

public void computeErrors()

Computes the classification error for the data points