edu.udo.cs.myRVM
Class ConstructiveRegression

java.lang.Object
  edu.udo.cs.myRVM.RVMBase
      edu.udo.cs.myRVM.ConstructiveRegression

public class ConstructiveRegression
extends RVMBase

Constructive RVM for regression problems (see bla).

Version:

$Id: ConstructiveRegression.java,v 1.7 2006/09/30 00:05:31 ingomierswa Exp $

Author:

Piotr Kasprzak

Field Summary

protected Jama.Matrix	A
protected double[]	alpha
protected java.util.LinkedList<java.lang.Integer>	basisSet
protected double	beta
protected Jama.Matrix	mu
protected double[][]	phi
protected Jama.Matrix	PHI_t
protected double	q
protected double	s
protected Jama.Matrix	SIGMA
protected Jama.Matrix	SIGMA_chol
protected double[][]	t
protected double[]	tVector
protected double[][]	x Data shared accross various methods

Fields inherited from class edu.udo.cs.myRVM.RVMBase

model, parameter, problem

Constructor Summary

ConstructiveRegression(RegressionProblem problem, Parameter parameter)
Constructor

Method Summary

protected double[]	convertListToDoubleArray(java.util.List list) Take a list holding "Double"-objects and return an "double[]"
protected void	deleteBasis(int selectedBasis) Delete a basis function from the model.
protected void	includeBasis(int selectedBasis) Include a basis function into the model.
protected double	innerProduct(double[] x, double[] y) Return the inner product of x and y (x.length == y.length assumed)
Model	learn() The hard work is done here
protected void	prune(java.util.LinkedList<java.lang.Integer> basisSet) Create pruned versions of all important matrices / vectors so that only rows / columns matching the indices in basisSet are kept.
protected void	reestimateAlpha(int selectedBasis) Reestimate alpha by setting it to the value which maximizes the marginal likelihood: alpha_i = s^2_i / (q^2_i - s_i)
java.lang.String	toString() Identify the RVM
protected void	updateBeta() Update beta (same as for the "normal" regression rvm)
protected void	updateCriteriumScalars(int selectedBasis) Compute the scalars s_m, q_m which are part of the criterium for inclusion / deletion of the given basis m: S_m = beta * phi^t_m * phi_m - beta^2 * phi^t_m * PHI * SIGMA * PHI^t * phi_m Q_m = beta * phi^t_m * t - beta^2 * phi^t_m * PHI * SIGMA * PHI^t * t s_m = alpha_m * S_m / (alpha_m - S_m) q_m = alpha_m * Q_m / (alpha_m - S_m)
protected void	updateMu() Update the mean of the weight posterior distribution (mu): mu = beta * SIGMA * PHI^t * t
protected void	updateSIGMA() Update the covariance Matrix of the weight posterior distribution (SIGMA) along with its cholesky factor: SIGMA = (A + beta * PHI^t * PHI)^{-1} SIGMA_chol with SIGMA_chol * SIGMA_chol^t = SIGMA

Methods inherited from class edu.udo.cs.myRVM.RVMBase

getModel

Methods inherited from class java.lang.Object

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

Field Detail

x

protected double[][] x

Data shared accross various methods

t

protected double[][] t

tVector

protected double[] tVector

phi

protected double[][] phi

PHI_t

protected Jama.Matrix PHI_t

alpha

protected double[] alpha

beta

protected double beta

A

protected Jama.Matrix A

SIGMA

protected Jama.Matrix SIGMA

SIGMA_chol

protected Jama.Matrix SIGMA_chol

mu

protected Jama.Matrix mu

s

protected double s

q

protected double q

basisSet

protected java.util.LinkedList<java.lang.Integer> basisSet

Constructor Detail

ConstructiveRegression

public ConstructiveRegression(RegressionProblem problem,
                              Parameter parameter)

Constructor

Method Detail

convertListToDoubleArray

protected double[] convertListToDoubleArray(java.util.List list)

Take a list holding "Double"-objects and return an "double[]"

innerProduct

protected double innerProduct(double[] x,
                              double[] y)

Return the inner product of x and y (x.length == y.length assumed)

prune

protected void prune(java.util.LinkedList<java.lang.Integer> basisSet)

Create pruned versions of all important matrices / vectors so that only rows / columns matching the indices in basisSet are kept.

updateSIGMA

protected void updateSIGMA()

Update the covariance Matrix of the weight posterior distribution (SIGMA) along with its cholesky factor: SIGMA = (A + beta * PHI^t * PHI)^{-1} SIGMA_chol with SIGMA_chol * SIGMA_chol^t = SIGMA

updateMu

protected void updateMu()

Update the mean of the weight posterior distribution (mu): mu = beta * SIGMA * PHI^t * t

updateCriteriumScalars

protected void updateCriteriumScalars(int selectedBasis)

Compute the scalars s_m, q_m which are part of the criterium for inclusion / deletion of the given basis m: S_m = beta * phi^t_m * phi_m - beta^2 * phi^t_m * PHI * SIGMA * PHI^t * phi_m Q_m = beta * phi^t_m * t - beta^2 * phi^t_m * PHI * SIGMA * PHI^t * t s_m = alpha_m * S_m / (alpha_m - S_m) q_m = alpha_m * Q_m / (alpha_m - S_m)

reestimateAlpha

protected void reestimateAlpha(int selectedBasis)

Reestimate alpha by setting it to the value which maximizes the marginal likelihood: alpha_i = s^2_i / (q^2_i - s_i)

includeBasis

protected void includeBasis(int selectedBasis)

Include a basis function into the model.

deleteBasis

protected void deleteBasis(int selectedBasis)

Delete a basis function from the model.

updateBeta

protected void updateBeta()

Update beta (same as for the "normal" regression rvm)

learn

public Model learn()

The hard work is done here

Specified by:

learn in class RVMBase

toString

public java.lang.String toString()

Identify the RVM

Overrides:

toString in class java.lang.Object