|
GenericProposer.java
|
/* * Copyright (c) 2005, Regents of the University of California * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * * * Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * * * Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in * the documentation and/or other materials provided with the * distribution. * * * Neither the name of the University of California, Berkeley nor * the names of its contributors may be used to endorse or promote * products derived from this software without specific prior * written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED * OF THE POSSIBILITY OF SUCH DAMAGE. */ package blog; import java.util.*; import common.Util; import common.MultiMap; import common.HashMultiMap;
|
Implements a generic MCMC proposal algorithm. Each step changes
the value of exactly one instantiated RV (selected uniformly at
random from the instantiated RVs), then samples values for
previously uninstantiated RVs to make the resulting instantiation
self-supporting. It also uninstantiates any RVs that become barren.
Each variable is sampled from its prior distribution given its parents. The forward proposal probability is just the probability of choosing the selected variable to change, times the product of the probabilities of values sampled (for that variable and any other variables that must be instantiated to make the instantiation self-supporting). The backward probability is the probability of choosing the same variable again, choosing its old value, and re-instantiating all the newly uninstantiated variables to their old values. No other moves can yield the same proposed world because each move changes the value of exactly one instantiated variable (it just instantiates and uninstantiates other variables). GenericProposer does not use identifiers in the worlds it constructs. |
public class GenericProposer implements Proposer {
| Creates a new GenericProposer that proposes possible worlds for the given model. The properties table, specifying configuration parameters, is passed to the LWSampler that is used to generate the initial world. |
public GenericProposer(Model model, Properties properties) {
this.model = model;
initialStateSampler = new LWSampler(model, properties);
}
| Set up the initial state. We repeatedly sample a world until we get one where the observed values of the evidence variables have positive probability. |
public PartialWorldDiff initialize(Evidence evidence, List queries) {
this.evidence = new Evidence();
evidenceVars = new HashSet();
numBasicEvidenceVars = 0;
this.queries = new LinkedList();
queryVars = new LinkedHashSet();
add(evidence);
addQueries(queries);
return constructInitialState();
}
| Adds extra evidence to the evidence the proposer must consider. |
public void add(Evidence evidence) {
this.evidence.addAll(evidence);
evidenceVars.add(evidence.getEvidenceVars());
for (Iterator iter = evidenceVars.iterator(); iter.hasNext(); ) {
if (iter.next() instanceof BasicVar) {
numBasicEvidenceVars++;
}
}
}
| Adds extra queries to the ones the proposer must consider. |
public void addQueries(List queries) {
this.queries.addAll(queries);
for(Iterator iter=queries.iterator(); iter.hasNext(); ) {
queryVars.addAll(((Query)iter.next()).getVariables());
}
}
Actually builds initial state, after whatever initializations
are necessary. Separated from { so extensions can change it alone.
|
protected PartialWorldDiff constructInitialState() {
++numTrials;
numInitialStateTriesThisTrial = 0;
initialStateSampler.initialize(evidence, queries);
while (true) {
initialStateSampler.nextSample();
PartialWorld initWorld = initialStateSampler.getLatestWorld();
++totalNumInitialStateTries;
++numInitialStateTriesThisTrial;
if (initialStateSampler.getLatestWeight() > 0) {
if (Util.verbose()) {
System.out.println("Probability of " +
numInitialStateTriesThisTrial +
"th initial state = " +
initialStateSampler.getLatestWeight());
initWorld.print(System.out);
}
PartialWorld underlying
= new DefaultPartialWorld(initWorld.getIdTypes());
return new PartialWorldDiff(underlying, initWorld);
} else { // world is inconsistent with evidence, try again
if (Util.verbose()) {
System.out.println(numInitialStateTriesThisTrial +
"th initial world rejected.");
}
}
}
}
protected class PickVarToSampleResult {
public VarWithDistrib varToSample;
public int numberOfChoices;
public PickVarToSampleResult(VarWithDistrib varToSample,
int numberOfChoices) {
this.varToSample = varToSample;
this.numberOfChoices = numberOfChoices;
}
}
protected PickVarToSampleResult pickVarToSample(PartialWorld world) {
Set eligibleVars = new HashSet(world.getInstantiatedVars());
eligibleVars.removeAll(evidenceVars);
// Uniformly sample a random variable from the support network.
VarWithDistrib varToSample
= (VarWithDistrib) Util.uniformSample(eligibleVars);
return new PickVarToSampleResult(varToSample, eligibleVars.size());
}
| Proposes a next state for the Markov chain given the current state. The proposedWorld argument is a PartialWorldDiff that the proposer can modify to create the proposal; the saved version of this PartialWorldDiff is the state before the proposal. Returns the log proposal ratio: log (q(x | x') / q(x' | x)) |
public double proposeNextState(PartialWorldDiff world) {
if (evidence == null) {
throw new IllegalStateException
("initialize() has not been called on proposer.");
}
logProbForward = 0;
logProbBackward = 0;
PickVarToSampleResult result = pickVarToSample(world);
if(result.varToSample == null) return 1.0;
if(Util.verbose())
System.out.println(" sampling " + result.varToSample);
// Multiply in the probability of this uniform sample.
logProbForward += (-Math.log(result.numberOfChoices));
// Sample value for variable and update forward and backward probs
sampleValue(result.varToSample, world);
// Remove barren variables
LinkedList newlyBarren = new LinkedList(world.getNewlyBarrenVars());
while (!newlyBarren.isEmpty()) {
BayesNetVar var = (BayesNetVar) newlyBarren.removeFirst();
if(!evidenceVars.contains(var) && !queryVars.contains(var)) {
// Remember its parents.
Set parentSet = world.getBayesNet().getParents(var);
if (var instanceof VarWithDistrib) {
// Multiply in the probability of sampling this
// variable again. Since the parent value may have
// changed, must use the old world.
logProbBackward
+= world.getSaved().getLogProbOfValue(var);
// Uninstantiate
world.setValue((VarWithDistrib) var, null);
}
// Check to see if its parents are now barren.
for(Iterator parentIter = parentSet.iterator();
parentIter.hasNext(); ) {
// If parent is barren, add to the end of this
// linked list. Note that if a parent has two
// barren children, it will only be added to the
// end of the list once, when the last child is
// considered.
BayesNetVar parent = (BayesNetVar) parentIter.next();
if(world.getBayesNet().getChildren(parent).isEmpty())
newlyBarren.addLast(parent);
}
}
}
// Uniform sampling from new world.
logProbBackward += (-Math.log(world.getInstantiatedVars().size()
- numBasicEvidenceVars));
return (logProbBackward - logProbForward);
}
public void printStats() {
System.out.println("===== " + getClass().getName() + " Stats ====");
System.out.println("Initial world attempts: "
+ numInitialStateTriesThisTrial);
if (numTrials > 0) {
System.out.println
("\tRunning average (for trials so far): " +
(totalNumInitialStateTries / (double) numTrials));
}
}
public void updateStats(boolean accepted) {
}
// Samples a new value for the given variable (which must be
// supported in <code>world</code>) and sets this new value as the
// value of the variable in <code>world</code>. Then ensures that
// <code>world</code> is self-supporting by calling
// ensureDetAndSupported on each of the variable's children. Also
// updates the logProbForward and logProbBackward variables.
private void sampleValue(VarWithDistrib varToSample, PartialWorld world) {
// Save child set before graph becomes out of date
Set children = world.getBayesNet().getChildren(varToSample);
DependencyModel.Distrib distrib = varToSample.getDistrib
(new DefaultEvalContext(world, true));
Object oldValue = world.getValue(varToSample);
logProbBackward += Math.log(distrib.getCPD().getProb
(distrib.getArgValues(), oldValue));
Object newValue = distrib.getCPD().sampleVal
(distrib.getArgValues(), varToSample.getType());
world.setValue(varToSample, newValue);
logProbForward += Math.log(distrib.getCPD().getProb
(distrib.getArgValues(), newValue));
// Make the world self-supporting. The only variables whose active
// parent sets could have changed are the children of varToSample.
InstantiatingEvalContext instantiator
= new InstantiatingEvalContext(world);
for (Iterator childrenIter = children.iterator();
childrenIter.hasNext(); ) {
BayesNetVar child = (BayesNetVar) childrenIter.next();
child.ensureDetAndSupported(instantiator);
}
logProbForward += instantiator.getLogProbability();
}
protected Model model;
protected Sampler initialStateSampler;
protected Evidence evidence = null;
protected List queries; // of Query
protected Set evidenceVars;
protected int numBasicEvidenceVars;
protected Set queryVars;
protected double logProbForward;
protected double logProbBackward;
protected int numTrials = 0;
protected int totalNumInitialStateTries = 0;
protected int numInitialStateTriesThisTrial = 0;
}
This file was generated on Tue Jun 08 17:53:36 PDT 2010 from file GenericProposer.java
by the ilog.language.tools.Hilite Java tool written by Hassan Aït-Kaci