// Compute scores for new leaf0 and leaf1, and add them to scoreList.
private void computeNewLeafScores(SelectedSet selectedSet, int i, int j) {
int NS = selectedSet.getN();
for (int a = 0; a < NS; a++) {
Individual individual = selectedSet.getIndividual(a);
IGraph iterator = decisionGraphs[i].getGraph();
while (!(iterator
instanceof Leaf)) { // Iterator is a variable because we are still traversing the DG.
int x = ((Variable) iterator).getVariable();
char alleleX = individual.getAllele(x);
if (alleleX == '0') {
iterator = ((Variable) iterator).getZero();
} else {
iterator = ((Variable) iterator).getOne();
}
}
int itrPosition = decisionGraphs[i].getLeafs().indexOf((Leaf) iterator);
if (itrPosition == j || itrPosition == j + 1) { // We've reached one of the new leafs.
int mZero = ((Leaf) iterator).getMZero();
int mOne = ((Leaf) iterator).getMOne();
if (mZero > 0 && mOne > 0) { // It's still "interesting" to split.
char alleleI = individual.getAllele(i); // Value of Xi in individual a.
for (int split : splitList[i]) {
char alleleS = individual.getAllele(split); // Value of Xsplit in individual a.
if (alleleI == '0') {
if (alleleS == '0') {
((Leaf) iterator).addPossibleSplitFrequency(0, split); // m00[split]++;
} else {
((Leaf) iterator).addPossibleSplitFrequency(1, split); // m01[split]++;
}
} else if (alleleS == '0') {
((Leaf) iterator).addPossibleSplitFrequency(2, split); // m10[split]++;
} else {
((Leaf) iterator).addPossibleSplitFrequency(3, split); // m11[split]++;
}
}
}
}
}
for (int a = 0; a <= 1; a++) {
Leaf newLeaf =
decisionGraphs[i].getLeaf(j + a); // The two new leafs are at positions 'j' and 'j+1'.
int mZero = newLeaf.getMZero();
int mOne = newLeaf.getMOne();
if (mZero > 0 && mOne > 0) {
for (int s : splitList[i]) {
int m00 = newLeaf.getPossibleSplitFrequency(0, s);
int m01 = newLeaf.getPossibleSplitFrequency(1, s);
int m10 = newLeaf.getPossibleSplitFrequency(2, s);
int m11 = newLeaf.getPossibleSplitFrequency(3, s);
double scoreGain = bayesianMetric.computeScoreGain(mZero, mOne, m00, m01, m10, m11);
newLeaf.setScoreGain(s, scoreGain);
newLeaf.updateBestSplit(
s,
scoreGain); // Responsible for updating the value of the best split score gain in this
// leaf.
}
}
} // END: for(int a = 0 ...)
} // END: computeNewLeafScores(...)
private void generateDecisionGraphs(SelectedSet selectedSet) {
this.initializeBN(
selectedSet); // Refresh the BN. Compute and store in decreasing order the first score
// gains, corresponding to adding a first edge to the empty BN.
int SIZE_GENOME = selectedSet.getIndividual(0).getIndividual().length;
while (bestScoreGain
> 0) { // Search for the best split and store all the necessary information to effectively
// perform it.
DecisionGraph bestDecisionGraph = decisionGraphs[bestDecisionGraphPos];
int bestLeafPos = bestDecisionGraph.getBestLeafPos();
int bestSplitPos =
bestDecisionGraph
.getBestLeaf()
.getBestSplit(); // Split Variable k. We are adding the edge (Xk)-->(Xi) to the BN.
if (parentList[bestDecisionGraphPos].size()
>= maxVertexDegree) // Do NOT perform this split because Xi has reached the maximum number
// of parents.
{
bestDecisionGraph.setBestLeafScoreGain(
Double
.NEGATIVE_INFINITY); // Ensure that this decision graph will no longer allow splits.
} else {
if (!(splitList[bestDecisionGraphPos].contains(
bestSplitPos))) { // Do NOT perform this split because it creates a cycle in the BN.
Leaf bestLeaf = bestDecisionGraph.getLeaf(bestLeafPos);
bestLeaf.setScoreGain(bestSplitPos, -1);
bestLeaf.resetBestSplit(splitList[bestDecisionGraphPos]);
} else { // Perform the best split.
performBestSplit(
bestDecisionGraphPos,
bestLeafPos,
bestSplitPos,
SIZE_GENOME); // Update the BN and remove non-valid splits.
computeNewLeafScores(selectedSet, bestDecisionGraphPos, bestLeafPos);
}
decisionGraphs[bestDecisionGraphPos].updateBestLeaf();
}
updateScoreGain();
} // END: while(bestScoreGain > 0)
} // END: generateDecisionGraphs(...)
private void initializeBN(SelectedSet selectedSet) {
int NS = selectedSet.getN();
int SIZE = selectedSet.getIndividual(0).getIndividual().length;
this.bestScoreGain = Double.NEGATIVE_INFINITY;
Individual[] individuals = selectedSet.getIndividuals();
for (int i = 0; i < Problem.n; i++) { // Refresh the Bayesian Network structure.
parentList[i] =
new HashSet<Integer>(
Problem.n); // Initial Capacity = stringSize; Default Load Factor = 0.75
adjacencyList[i] =
new HashSet<Integer>(
Problem.n); // Initial Capacity = stringSize; Default Load Factor = 0.75
splitList[i] =
new HashSet<Integer>(
2 * Problem.n); // Initial Capacity = 2*stringSize; Default Load Factor = 0.75
for (int n = 0; n < Problem.n; n++) {
if (n != i) {
splitList[i].add(n);
}
}
int mOne =
selectedSet.getUniFrequencies(i); // Construct the initial single leaf decision graphs.
int mZero = NS - mOne;
Leaf newLeaf =
new Leaf(
0, -1, mZero, mOne,
SIZE); // There is only a single leaf per variable, at depth 0, with side -1.
if (mZero > 0
&& mOne
> 0) { // If mZero = 0 or mOne = 0 there is no need to try any split with this leaf.
for (int j = 0; j < NS; j++) {
char alleleJI = individuals[j].getAllele(i); // Value of Xi in individual j.
for (int s : splitList[i]) {
char alleleS = individuals[j].getAllele(s); // Value of Xs in individual j.
if (alleleJI == '0') {
if (alleleS == '0') {
newLeaf.addPossibleSplitFrequency(0, s); // m00[s]++;
} else {
newLeaf.addPossibleSplitFrequency(1, s); // m01[s]++;
}
} else if (alleleS == '0') {
newLeaf.addPossibleSplitFrequency(2, s); // m10[s]++;
} else {
newLeaf.addPossibleSplitFrequency(3, s); // m11[s]++;
}
}
}
for (int s : splitList[i]) {
int m00 = newLeaf.getPossibleSplitFrequency(0, s);
int m01 = newLeaf.getPossibleSplitFrequency(1, s);
int m10 = newLeaf.getPossibleSplitFrequency(2, s);
int m11 = newLeaf.getPossibleSplitFrequency(3, s);
double scoreGain = bayesianMetric.computeScoreGain(mZero, mOne, m00, m01, m10, m11);
newLeaf.setScoreGain(s, scoreGain);
newLeaf.updateBestSplit(
s,
scoreGain); // Responsible for updating the value of the best split score gain in this
// leaf.
}
} // END: if(mZero > 0 ...)
decisionGraphs[i] =
new DecisionGraph(
newLeaf); // Initially each graph as a single leaf and there are n-1 possible splits.
decisionGraphs[i].updateBestLeaf();
double scoreGain = decisionGraphs[i].getBestLeafScoreGain();
if (scoreGain
> this
.bestScoreGain) { // Update the value of the global best score gain and position (DG).
this.bestDecisionGraphPos = i;
this.bestScoreGain = scoreGain;
}
} // END: for(int i = 0 ...)
} // END: initializeBN(...)