de.up.ling.irtg.sampling.rule_weighting

Class RegularizedKLRuleWeighting

java.lang.Object

de.up.ling.irtg.sampling.rule_weighting.RegularizedKLRuleWeighting

All Implemented Interfaces:

RuleWeighting

Direct Known Subclasses:

AutomatonWeighted

public abstract class RegularizedKLRuleWeighting
extends Object
implements RuleWeighting

Implements adaption by minimizing the KL between the target distribution and the proposal distribution as described in our adaptive importance sampling paper. This class implements all functionality necessary for a rule weighting except the computation of the target weight.

Author:

teichmann

Field Summary

Fields

Modifier and Type	Field and Description
static double	ALMOST_ZERO The threshold for when we consider a number close enough to 0.0 to pick a choice point.

Constructor Summary

Constructors

Constructor and Description
RegularizedKLRuleWeighting(TreeAutomaton basis, int regularizationExponent, double regularizationDivisor, LearningRate rate) Creates a new instance with the regularization given by the specified values.

Method Summary

Modifier and Type	Method and Description
void	adapt(TreeSample<Rule> treSamp, boolean deterministic) Adapts the proposal distribution with the assumption that treSamp is an importance sample generated from this proposal distribution.
TreeAutomaton	getAutomaton() Returns the underlying tree automaton from which the rules and start states are drawn.
double	getLogProbability(int state, int ruleNumber) Returns the log of the probability of choosing the rule identified by the number given the state.
double	getLogProbability(Rule r) Returns the proposal probability of the given rule given its parent.
int	getNumberOfStartStates() Returns the overall number of start states available.
Rule	getRuleByNumber(int state, int number) Returns the rule identified by the given state and number.
int	getRuleNumber(int state, double choicePoint) Returns the number of the first rule such that the cumulative probability of the earlier rules plus this one is larger than choicePoint.
int	getStartStateByNumber(int number) Returns the start state with the given number.
int	getStartStateNumber(double choicePoint) Returns the start first start state such that the cumulative probability of earlier start states plus this one is larger than choicePoint.
double	getStateStartLogProbability(int position) Returns the log of the proposal probability of the given start state.
void	prepareProbability(int state) Tells the class to recompute the proposal probabilities for the rules of the given state.
void	prepareStartProbability() Recomputes the proposal probabilities for the start states.
void	reset() Resets any adaption of the proposal distribution.

Methods inherited from class java.lang.Object

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

Methods inherited from interface de.up.ling.irtg.sampling.RuleWeighting

getLogTargetProbability

Field Detail

ALMOST_ZERO

public static double ALMOST_ZERO

The threshold for when we consider a number close enough to 0.0 to pick a choice point. This is to ensure that we never throw an error because of rounding error in computing probabilities.

Constructor Detail

RegularizedKLRuleWeighting

public RegularizedKLRuleWeighting(TreeAutomaton basis,
                                  int regularizationExponent,
                                  double regularizationDivisor,
                                  LearningRate rate)

Creates a new instance with the regularization given by the specified values.

Parameters:

basis -

regularizationExponent -

regularizationDivisor -

rate -

Method Detail

getLogProbability

public double getLogProbability(int state,
                                int ruleNumber)

Description copied from interface: RuleWeighting

Returns the log of the probability of choosing the rule identified by the number given the state. The value should be those computed by the last prepareProbability call.

Specified by:

getLogProbability in interface RuleWeighting

Returns:

prepareProbability

public void prepareProbability(int state)

Description copied from interface: RuleWeighting

Tells the class to recompute the proposal probabilities for the rules of the given state. If it is certain that they have not changed, then it is acceptable to do nothing.

Specified by:

prepareProbability in interface RuleWeighting

getStateStartLogProbability

public double getStateStartLogProbability(int position)

Description copied from interface: RuleWeighting

Returns the log of the proposal probability of the given start state. May throw an error if there is no such state. The value is always the last one computed by prepareStartProbability.

Specified by:

getStateStartLogProbability in interface RuleWeighting

Returns:

prepareStartProbability

public void prepareStartProbability()

Description copied from interface: RuleWeighting

Recomputes the proposal probabilities for the start states. If they are guaranteed to be unchanged, then this step can be skipped by the implementing class.

Specified by:

prepareStartProbability in interface RuleWeighting

reset

public void reset()

Description copied from interface: RuleWeighting

Resets any adaption of the proposal distribution.

Specified by:

reset in interface RuleWeighting

adapt

public void adapt(TreeSample<Rule> treSamp,
                  boolean deterministic)

Description copied from interface: RuleWeighting

Adapts the proposal distribution with the assumption that treSamp is an importance sample generated from this proposal distribution.

Specified by:

adapt in interface RuleWeighting

deterministic - indicates whether we can assume the underlying automaton to be unambiguous.

getAutomaton

public TreeAutomaton getAutomaton()

Description copied from interface: RuleWeighting

Returns the underlying tree automaton from which the rules and start states are drawn.

Specified by:

getAutomaton in interface RuleWeighting

Returns:

getNumberOfStartStates

public int getNumberOfStartStates()

Description copied from interface: RuleWeighting

Returns the overall number of start states available.

Specified by:

getNumberOfStartStates in interface RuleWeighting

Returns:

getStartStateByNumber

public int getStartStateByNumber(int number)

Description copied from interface: RuleWeighting

Returns the start state with the given number. May throw an error if there is no such state.

Specified by:

getStartStateByNumber in interface RuleWeighting

Returns:

getRuleNumber

public int getRuleNumber(int state,
                         double choicePoint)

Description copied from interface: RuleWeighting

Returns the number of the first rule such that the cumulative probability of the earlier rules plus this one is larger than choicePoint. May throw an error if there is no such rule.

Specified by:

getRuleNumber in interface RuleWeighting

Returns:

getRuleByNumber

public Rule getRuleByNumber(int state,
                            int number)

Description copied from interface: RuleWeighting

Returns the rule identified by the given state and number. May throw an error if there is no such rule.

Specified by:

getRuleByNumber in interface RuleWeighting

Returns:

getStartStateNumber

public int getStartStateNumber(double choicePoint)

Description copied from interface: RuleWeighting

Returns the start first start state such that the cumulative probability of earlier start states plus this one is larger than choicePoint. May throw an error if there is no such state.

Specified by:

getStartStateNumber in interface RuleWeighting

Returns:

getLogProbability

public double getLogProbability(Rule r)

Description copied from interface: RuleWeighting

Returns the proposal probability of the given rule given its parent.

Specified by:

getLogProbability in interface RuleWeighting

Returns: