javaslam.filter
Class InformationFilter

java.lang.Object
  |
  +--javaslam.filter.InformationFilter

All Implemented Interfaces:

Filter

public class InformationFilter

extends Object

implements Filter

An Information filter.

This class records counts of all floating point operations using Flops.count(long) (except those used in the service of debugging and avoiding numerical errors).

Field Summary

protected Gaussian

p The canonical potential representing the belief state.

Constructor Summary

InformationFilter(Gaussian p)
Constructor.

Method Summary

static Gaussian	getConditionalGaussian(ListSet vars, Matrix y0, Matrix C, Matrix R, Matrix y) Creates a linear-Gaussian measurement potential.
Gaussian	getDistribution() Returns the current filtered belief state.
Gaussian	getMarginal(Set vars) Extracts the filtered marginal distribution.
Map	getMarginals(Collection vars) Extracts a set of unary marginals.
static Gaussian	getUnconditionalGaussian(ListSet newVars, ListSet vars, Matrix x0, Matrix A, Matrix Q) Constructs a linear-Gaussian potential that introduces a set of new unobserved variables.
Set	getVariables() Gets an unmodifiable set of the Variables in the filtered belief state.
void	marginalizeOut(Set mvars) Marginalizes a set of variables out of the belief state.
void	measurement(ListSet vars, double[] y0, double[][] C, double[][] R, double[] y) Performs a linear-Gaussian measurement update.
void	time(ListSet vars, double[] x0, double[][] A, double[][] Q) Performs a linear-Gaussian time update.

Methods inherited from class java.lang.Object

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

Field Detail

p

protected Gaussian p

The canonical potential representing the belief state.

Constructor Detail

InformationFilter

public InformationFilter(Gaussian p)

Constructor.

Parameters:

p - the initial belief state

Method Detail

getDistribution

public Gaussian getDistribution()

Returns the current filtered belief state.

Returns:

the current filtered belief state

marginalizeOut

public void marginalizeOut(Set mvars)

Marginalizes a set of variables out of the belief state.

Specified by:

marginalizeOut in interface Filter

Parameters:

mvars - a set of Variables to marginalize out

getConditionalGaussian

public static Gaussian getConditionalGaussian(ListSet vars,
                                              Matrix y0,
                                              Matrix C,
                                              Matrix R,
                                              Matrix y)

Creates a linear-Gaussian measurement potential. The parameters vars, C and R define the measurement equation as follows:

y=y0 + Cx(vars)+w

where w is a white-noise variable with covariance R. Given the actual measurement y, this method computes the potential over vars that, when multiplied into a distribution over vars, has the effect of conditioning on y.

Parameters:

vars - an ordered set of the variables with sum dimension n in the belief state that causally influenced this measurement; any variables in this list that are not currently in the belief state are added with uninformative priors.

y0 - a k-vector giving the constant term

C - a k by n observation matrix that defines the linear measurement model (and whose columns are ordered consistently with the order of vars)

R - a k by k symmetric positive definite matrix giving the covariance of the measurement white noise

y - the measurement k-vector

Returns:

a Gaussian in the canonical parameterization which, when multiplied into a graphical model, has the effect of conditioning on the observation y

Throws:

IllegalArgumentException - if there are any dimension mismatches

measurement

public void measurement(ListSet vars,
                        double[] y0,
                        double[][] C,
                        double[][] R,
                        double[] y)

Performs a linear-Gaussian measurement update. The parameters vars, C and R define the measurement equation as follows:

y=y0 + Cx(vars)+w

where w is a white-noise variable with covariance R. Given the actual measurement y, this method updates the belief state in O(1) time.

Specified by:

measurement in interface Filter

Parameters:

vars - an ordered set of the variables with sum dimension n in the belief state that causally influenced this measurement; any variables in this list that are not currently in the belief state are added with uninformative priors.

y0 - a k-vector giving the constant term

C - a k by n observation matrix that defines the linear measurement model (and whose columns are ordered consistently with the order of vars)

R - a k by k symmetric positive definite matrix giving the covariance of the measurement white noise

y - the measurement k-vector

Throws:

IllegalArgumentException - if there are any dimension mismatches

getUnconditionalGaussian

public static Gaussian getUnconditionalGaussian(ListSet newVars,
                                                ListSet vars,
                                                Matrix x0,
                                                Matrix A,
                                                Matrix Q)

Constructs a linear-Gaussian potential that introduces a set of new unobserved variables. The parameters vars, newVars, x0, A and Q define the distribution of the new variables as follows:

newVars=x0 Avars+v

where v is a white-noise variable with covariance Q.

Multiplying the returned potential into a junction tree will not result in message passing; the junction tree is already consistent. This is because the new variables are barren nodes, i.e., unobserved children of the other variables, and thus do not change their distribution.

Parameters:

newVars - an ordered set of variables with sum dimension n that are not present in the belief state

vars - an ordered set of variables with sum dimension m that are currently in the belief state

x0 - an n-vector giving the constant term

A - an n by m linear coefficient matrix

Q - an n by n symmetric positive definite matrix giving the covariance of the white noise

Returns:

a Gaussian in the canonical parameterization which, when multiplied into a graphical model, has the effect of adding newVars into the belief state as a directed child of vars

Throws:

IllegalArgumentException - if there are any dimension mismatches

time

public void time(ListSet vars,
                 double[] x0,
                 double[][] A,
                 double[][] Q)

Performs a linear-Gaussian time update. The parameters vars, A and Q define the state evolution equation as follows:

x_{t + 1}(vars)=x0 Ax_t(vars)+v

where v is a white-noise variable with covariance Q. All variables not in vars are assumed stationary.

Specified by:

time in interface Filter

Parameters:

vars - an ordered set of the variables with sum dimension n in the belief state that evolve over time

x0 - an n-vector giving the constant term

A - an n by n evolution matrix that defines the linear evolution model (and whose blocks are ordered consistently with the order of vars)

Q - an n by n symmetric positive definite matrix giving the covariance of the evolution white noise (and whose blocks are ordered consistently with the order of vars)

Throws:

IllegalArgumentException - if there are any dimension mismatches or vars contains variables that are not in the current belief state

getMarginal

public Gaussian getMarginal(Set vars)

Extracts the filtered marginal distribution.

Specified by:

getMarginal in interface Filter

Parameters:

vars - the set of Variables whose filtered marginal is to be computed

Returns:

a filtered marginal potential over vars

getMarginals

public Map getMarginals(Collection vars)

Extracts a set of unary marginals.

Specified by:

getMarginals in interface Filter

Parameters:

vars - a collection of Variables, or null to indicate all variables in the belief state

Returns:

a map whose keys are the (distinct) elements of vars and whose values are the corresponding marginals (in the moment parameterization)

getVariables

public Set getVariables()

Gets an unmodifiable set of the Variables in the filtered belief state.

Specified by:

getVariables in interface Filter