/**
* Simulate next step for the given particle.
*
* @param nextPts
* @param particleNum
*/
public void simulateForwardStep(ArrayList<Point> nextPts, Particle thisParticle) {
HashMap<Vertex, Point> thisObservedPosMap = new HashMap<Vertex, Point>();
HashMap<Vertex, AntPath> thisAntPathMap = new HashMap<Vertex, AntPath>();
thisParticle.currentFalsePositives = new ArrayList<Point>();
for (Point pt : nextPts) {
Vertex vx = new ObsVertex();
thisObservedPosMap.put(vx, pt);
}
for (AntPath ap : thisParticle.getPaths()) {
Vertex vx = new PathVertex();
thisAntPathMap.put(vx, ap);
}
/*
* Compute probability graph for this particle.
*/
SimpleWeightedGraph<Vertex, DefaultWeightedComparableEdge> probabilityGraph =
new SimpleWeightedGraph<Vertex, DefaultWeightedComparableEdge>(
DefaultWeightedComparableEdge.class);
ValueSortedMap<DefaultWeightedComparableEdge, Double> edgeMap =
new ValueSortedMap<DefaultWeightedComparableEdge, Double>(true);
computeLogProbabilityGraph(probabilityGraph, edgeMap, thisObservedPosMap, thisAntPathMap);
// displayLogProbabilityGraph(probabilityGraph,thisObservedPosMap,thisAntPathMap,vertexSums);
/*
* Generate new ant locations based on probability graph. We do this by sampling the most likely event,
* removing this event from the probability graph, updating the prob. graph, then repeating until all
* observations/causes are accounted for.
*
* To do this (relatively) efficiently, we maintain a sorted hash-map of possible observation/cause pairs.
* To avoid having to rescale the entire hash-map every step, we maintain the current sum of weights.
*
* We're also keeping track of the posterior log-probability of the simulated step.
*/
double logprob = 0;
int nfp = 0;
int nfn = 0;
int doOutput = 0;
while (edgeMap.size() > 0) {
/*
* Sample an event from the probability graph.
*/
DefaultWeightedComparableEdge event = sampleEdge(probabilityGraph, edgeMap);
edgeMap.remove(event);
double edgeWeight = probabilityGraph.getEdgeWeight(event);
double thisLogProb = edgeWeight;
logprob = logprob + thisLogProb;
Vertex v1 = probabilityGraph.getEdgeSource(event);
Vertex v2 = probabilityGraph.getEdgeTarget(event);
if (doOutput > 0) {
System.err.println("edgeWeight: " + edgeWeight);
doOutput = outputInterestingStuff(v1, v2, probabilityGraph);
}
if (v1.getClass().equals(ObsVertex.class)) {
assert (v2.getClass().equals(PathVertex.class));
Vertex tv = v1;
v1 = v2;
v2 = tv;
}
/*
* Update the probability graph.
*/
if (!v1.equals(falsePositive)) {
Set<DefaultWeightedComparableEdge> es = probabilityGraph.edgesOf(v1);
for (DefaultWeightedComparableEdge ed : es) edgeMap.remove(ed);
boolean tt = probabilityGraph.removeVertex(v1);
assert (tt);
} else {
nfp += 1;
thisParticle.currentFalsePositives.add(thisObservedPosMap.get(v2));
}
if (!v2.equals(falseNegative)) {
Set<DefaultWeightedComparableEdge> es = probabilityGraph.edgesOf(v2);
for (DefaultWeightedComparableEdge ed : es) edgeMap.remove(ed);
boolean tt = probabilityGraph.removeVertex(v2);
assert (tt);
} else nfn += 1;
// Utils.computeVertexSums(probabilityGraph,vertexSums);
/*
* Update AntPath trajectory.
*/
if (!v1.equals(falsePositive)) {
AntPath ap = thisAntPathMap.get(v1);
assert (ap != null);
Point obs = thisObservedPosMap.get(v2);
Point newPos = new Point();
double thislp = sampleConditionalPos(ap, obs, newPos) + thisLogProb;
ap.updatePosition(newPos, obs, thislp);
}
/*
* Compute likelihoods.
*/
}
}
/**
* Compute the log-probability that each observation arises for each ant. This is represented by a
* simple weighted graph, where the absence of an edge represents zero probability that the given
* observation was caused by the given ant. Currently the log-prob is set to be a truncated iid
* Gaussian around an ant location.
*
* @param probabilityGraph The probability graph being constructed.
* @param edgeMap A sorted map of the edges in probabilityGraph.
* @param thisObservedPosMap Observed points along with index.
* @param thisParticlePosMap Current particle locations.
* @param obsSums
* @param antSums
* @return Sum of edge weights
*/
private double computeLogProbabilityGraph(
SimpleWeightedGraph<Vertex, DefaultWeightedComparableEdge> probabilityGraph,
ValueSortedMap<DefaultWeightedComparableEdge, Double> edgeMap,
HashMap<Vertex, Point> thisObservedPosMap,
HashMap<Vertex, AntPath> thisParticlePathMap) {
/** Initialize vertices. */
edgeMap.clear();
probabilityGraph.addVertex(falseNegative);
probabilityGraph.addVertex(falsePositive);
for (Vertex v : thisObservedPosMap.keySet()) {
probabilityGraph.addVertex(v);
DefaultWeightedComparableEdge edge = probabilityGraph.addEdge(v, falsePositive);
probabilityGraph.setEdgeWeight(edge, falsePositiveLogProb);
edgeMap.put(edge, falsePositiveLogProb);
}
for (Vertex v : thisParticlePathMap.keySet()) {
probabilityGraph.addVertex(v);
DefaultWeightedComparableEdge edge = probabilityGraph.addEdge(v, falseNegative);
probabilityGraph.setEdgeWeight(edge, falseNegativeLogProb);
edgeMap.put(edge, falseNegativeLogProb);
}
/** Compute probability of each observation given each ant path */
for (Entry<Vertex, Point> obsEntry : thisObservedPosMap.entrySet()) {
for (Entry<Vertex, AntPath> parEntry : thisParticlePathMap.entrySet()) {
double logprob = observationLogProbGivenAntPath(obsEntry.getValue(), parEntry.getValue());
if (logprob > logProbThreshold) {
DefaultWeightedComparableEdge edge =
probabilityGraph.addEdge(obsEntry.getKey(), parEntry.getKey());
probabilityGraph.setEdgeWeight(edge, logprob);
edgeMap.put(edge, logprob);
}
}
}
/** Adjust probabilities according to log-probability of an AntPath existing. */
for (Entry<Vertex, AntPath> antEntry : thisParticlePathMap.entrySet()) {
Set<DefaultWeightedComparableEdge> edge = probabilityGraph.edgesOf(antEntry.getKey());
double antProb = antEntry.getValue().getCurrentLogProb();
for (DefaultWeightedComparableEdge e : edge) {
double ew = probabilityGraph.getEdgeWeight(e);
probabilityGraph.setEdgeWeight(e, ew + antProb);
// edgeMap.remove(edge);
edgeMap.put(e, ew + antProb);
}
}
/**
* ANG -- to do Interaction effects: -- probability of false negative higher when multiple ants
* in same area. however, there is a high probability that there will be SOME observation in the
* area. -- probability of a false positive higher when there is one ant in a region. this is
* because sometimes one ant is split into two.
*/
/** compute vertex sums */
// Utils.computeVertexSums(probabilityGraph);
double totalLogProb = Utils.maxstar(edgeMap);
return totalLogProb;
}