io.improbable.keanu.algorithms.particleFiltering

Class ParticleFilter

java.lang.Object

io.improbable.keanu.algorithms.particleFiltering.ParticleFilter

public class ParticleFilter
extends java.lang.Object

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	ParticleFilter.Particle

Constructor Summary

Constructors

Constructor and Description
ParticleFilter()

Method Summary

Modifier and Type	Method and Description
static java.util.List<ParticleFilter.Particle>	getProbableValues(java.util.Collection<? extends Vertex<?>> vertices, int numParticles, int resamplingCycles, double resamplingProportion, java.util.Random random) A particle filtering approach is used to find probable values for the latent vertices of a Bayesian network, given a set of observed vertices.

Methods inherited from class java.lang.Object

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

Constructor Detail

ParticleFilter

public ParticleFilter()

Method Detail

getProbableValues

public static java.util.List<ParticleFilter.Particle> getProbableValues(java.util.Collection<? extends Vertex<?>> vertices,
                                                                        int numParticles,
                                                                        int resamplingCycles,
                                                                        double resamplingProportion,
                                                                        java.util.Random random)

A particle filtering approach is used to find probable values for the latent vertices of a Bayesian network, given a set of observed vertices. This is done by incrementally increasing the proportion of the graph under consideration, randomly sampling values for newly added latent variables at each stage. Each increment involves the addition of one new observed vertex and the latent vertices that it depend on. This is done for a specified number of 'particles', each of which represents one set of randomly sampled values of the latent vertices in the Bayesian network and has an associated probability. As the proportion of the graph under consideration incrementally grows, less probable particles are culled and more probable particles are duplicated, resulting in a final set of relatively probable particles. This methodology is similar to the Sequential Importance Resampling Algorithm described here (https://www.lancaster.ac.uk/pg/turnerl/PartileFiltering.pdf).

Parameters:

vertices - the vertices of a Bayesian network to find probable values for

numParticles - the number of particles to generate (a larger number will yield better results but is more computationally expensive)

resamplingCycles - the number of times low probability particles are culled and high probability particles are replicated each time the subgraph under consideration is expanded

resamplingProportion - the proportion of particles to cull (e.g. the 25% of least probably particles could be culled)

random - a random number generator

Returns: