Package com.bayesserver.inference

Class VariableEliminationInference

java.lang.Object

com.bayesserver.inference.VariableEliminationInference

All Implemented Interfaces:

Inference

public final class VariableEliminationInference
extends Object
implements Inference

An exact inference algorithm for Bayesian networks and Dynamic Bayesian networks, loosely based on the Variable Elimination algorithm.

See Also:

RelevanceTreeInference, Inference

Constructor Summary

Constructors

Constructor	Description
VariableEliminationInference(Network network)	Initializes a new instance of the VariableEliminationInference class, with the target Bayesian network.

Method Summary

Modifier and Type	Method	Description
Evidence	getBaseEvidence()	Optional evidence which can be used to calculate the lift of queries.
Evidence	getEvidence()	Gets the evidence (case data, e.g.
Network	getNetwork()	The target Bayesian network.
QueryDistributionCollection	getQueryDistributions()	The collection of distributions required from a Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput).
QueryFunctionCollection	getQueryFunctions()	Gets the collection of functions to evaluate, after QueryDistributions have been calculated.
QueryLifecycle	getQueryLifecycle()	Optional, allowing callers to hook into query lifecycle events, such as begin query and end query.
void	query(QueryOptions queryOptions, QueryOutput queryOutput)	Calculates a number of distributions, e.g.
void	setBaseEvidence(Evidence value)	Optional evidence which can be used to calculate the lift of queries.
void	setEvidence(Evidence value)	Sets the evidence (case data, e.g.
void	setQueryDistributions(QueryDistributionCollection value)	The collection of distributions required from a Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput).
void	setQueryFunctions(QueryFunctionCollection value)	Sets the collection of functions to evaluate, after QueryDistributions have been calculated.
void	setQueryLifecycle(QueryLifecycle value)	Optional, allowing callers to hook into query lifecycle events, such as begin query and end query.

Methods inherited from class java.lang.Object

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

Constructor Detail

VariableEliminationInference

public VariableEliminationInference(Network network)

Initializes a new instance of the VariableEliminationInference class, with the target Bayesian network.

Parameters:

network - The target Network.

Throws:

NullPointerException - Raised if [network] is null.

Method Detail

getEvidence

public Evidence getEvidence()

Gets the evidence (case data, e.g. row in a database) used in a query. The distributions are only recalculated when Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput) is called, not each time evidence is changed.

Specified by:

getEvidence in interface Inference

setEvidence

public void setEvidence(Evidence value)

Sets the evidence (case data, e.g. row in a database) used in a query. The distributions are only recalculated when Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput) is called, not each time evidence is changed.

Specified by:

setEvidence in interface Inference

getBaseEvidence

public Evidence getBaseEvidence()

Optional evidence which can be used to calculate the lift of queries. Defaults to null. When QueryDistribution.getComparison() is true queries are adjusted based on marginals calculated with this evidence or no evidence if this value is null.

Specified by:

getBaseEvidence in interface Inference

See Also:

QueryDistribution.getComparison()

setBaseEvidence

public void setBaseEvidence(Evidence value)

Optional evidence which can be used to calculate the lift of queries. Defaults to null. When QueryDistribution.getComparison() is true queries are adjusted based on marginals calculated with this evidence or no evidence if this value is null.

Specified by:

setBaseEvidence in interface Inference

See Also:

QueryDistribution.getComparison()

getQueryDistributions

public QueryDistributionCollection getQueryDistributions()

The collection of distributions required from a Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput). Only request those that you need.

Each distribution added, such as a Table, acts like a buffer. I.e. the values are populated by the Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput) method. This means that the same distributions can be resused across multiple calls to Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput), without the need to recreate the distribution each time.

It is important not to request distributions you do not need, because the computations will take longer. For example, it is common to request all marginal probabilities, P(A), P(B), P(C), ... , P(Z) given the evidence, however if you only need to know P(A) and P(B) given the evidence, then limit the query to just these distributions.

Typically the distributions requested are a number of marginal propabilities, such as P(A), P(B), P(C) etc... given the evidence, however it is also possible to request distributions over more than one variable, so we might query P(A), P(B,C) given the evidence.

If you are performing a batch of queries, there is no need to remove those distributions that you require, however happen to have evidence set for a particular query.

Specified by:

getQueryDistributions in interface Inference

setQueryDistributions

public void setQueryDistributions(QueryDistributionCollection value)

The collection of distributions required from a Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput). Only request those that you need.

Each distribution added, such as a Table, acts like a buffer. I.e. the values are populated by the Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput) method. This means that the same distributions can be resused across multiple calls to Inference.query(com.bayesserver.inference.QueryOptions, com.bayesserver.inference.QueryOutput), without the need to recreate the distribution each time.

It is important not to request distributions you do not need, because the computations will take longer. For example, it is common to request all marginal probabilities, P(A), P(B), P(C), ... , P(Z) given the evidence, however if you only need to know P(A) and P(B) given the evidence, then limit the query to just these distributions.

Typically the distributions requested are a number of marginal propabilities, such as P(A), P(B), P(C) etc... given the evidence, however it is also possible to request distributions over more than one variable, so we might query P(A), P(B,C) given the evidence.

If you are performing a batch of queries, there is no need to remove those distributions that you require, however happen to have evidence set for a particular query.

Specified by:

setQueryDistributions in interface Inference

getQueryFunctions

public QueryFunctionCollection getQueryFunctions()

Gets the collection of functions to evaluate, after QueryDistributions have been calculated. If required, a query function collection can be dynamically attached or detached. Also note that individual elements can be enabled or disabled on a per query basis.

Specified by:

getQueryFunctions in interface Inference

setQueryFunctions

public void setQueryFunctions(QueryFunctionCollection value)

Sets the collection of functions to evaluate, after QueryDistributions have been calculated. If required, a query function collection can be dynamically attached or detached. Also note that individual elements can be enabled or disabled on a per query basis.

Specified by:

setQueryFunctions in interface Inference

getQueryLifecycle

public QueryLifecycle getQueryLifecycle()

Optional, allowing callers to hook into query lifecycle events, such as begin query and end query.

Specified by:

getQueryLifecycle in interface Inference

setQueryLifecycle

public void setQueryLifecycle(QueryLifecycle value)

Optional, allowing callers to hook into query lifecycle events, such as begin query and end query.

Specified by:

setQueryLifecycle in interface Inference

getNetwork

public Network getNetwork()

The target Bayesian network.

Specified by:

getNetwork in interface Inference

query

public void query(QueryOptions queryOptions,
                  QueryOutput queryOutput)
           throws InconsistentEvidenceException

Calculates a number of distributions, e.g. P(A) and P(B) given the evidence (case data, e.g. row in a database), and if requested the log-likelihood of the data. Each time this method is called the distributions have their values replaced, acting like buffers.

As well as requesting distributions of the form P(A), P(B) it is also possible to request distributions over a number of variables such as P(A,B).

If CLGaussian distributions are requested that include instantiated discrete variables, an algorithm may legitimately return mean and covariance values equal to NaN for entries corresponding to inconsistent discrete combinations. For example, if a CLGaussian is requested that includes a discrete variable 'Gender' with states 'Male' and 'Female' and evidence is set to 'Female' then any mean and covariance entries that correspond to 'Male' may be set to NaN. NaN values will always be accompanied by a corresponding zero value in the CLGaussian.getTable().

Specified by:

query in interface Inference

Parameters:

queryOptions - Options governing which calculations are performed and how.

queryOutput - Returns any information, in addition to the distributions, that is requested. For example the log-likelihood.

Throws:

InconsistentEvidenceException - Raised if inconsistent evidence is detected, or underflow/overflow occurs when calculating queries or the conflict measure. Note that log-likelihood calculations do not raise this exception, but instead report -Infinity or +Infinity.