/* (non-Javadoc)
* @see datamining.clustering.protoype.AbstractPrototypeClusteringAlgorithm#getObjectiveFunctionValue()
*/
@Override
public double getObjectiveFunctionValue() {
if (!this.initialized)
throw new AlgorithmNotInitializedException("Prototypes not initialized.");
int i, j;
// i: index for clusters
// j: index for data objects
// k: index for dimensions, others
double objectiveFunctionValue = 0.0d;
double distanceSum =
0.0d; // the sum_i dist[i][l]^{2/(1-fuzzifier)}: the sum of all parametrised distances for
// one cluster l
double doubleTMP = 0.0d; // a temporarly variable for multiple perpuses
double[] fuzzDistances = new double[this.getClusterCount()];
double[] distancesSq = new double[this.getClusterCount()];
double minDistValue = 0.0d;
for (j = 0; j < this.getDataCount(); j++) {
distanceSum = 0.0d;
minDistValue = Double.MAX_VALUE;
for (i = 0; i < this.getClusterCount(); i++) {
doubleTMP =
this.dist.distanceSq(this.data.get(j).element, this.prototypes.get(i).getPosition());
distancesSq[i] = doubleTMP;
if (minDistValue > doubleTMP) minDistValue = doubleTMP;
}
minDistValue *= this.distanceMultiplierConstant;
for (i = 0; i < this.getClusterCount(); i++) {
doubleTMP = distancesSq[i] - minDistValue;
if (doubleTMP <= 0.0d) {
fuzzDistances[i] = 1.0d;
} else {
doubleTMP = 1.0d / doubleTMP;
fuzzDistances[i] = doubleTMP;
distanceSum += doubleTMP;
}
}
// don't check for distance sum to be zero.. that would just be rediculus!!
for (i = 0; i < this.getClusterCount(); i++) {
doubleTMP = fuzzDistances[i] / distanceSum;
objectiveFunctionValue +=
MyMath.pow(doubleTMP, this.fuzzifier) * (distancesSq[i] - minDistValue);
}
}
return objectiveFunctionValue;
}
@Override
public void apply(int steps) {
if (!this.initialized)
throw new AlgorithmNotInitializedException("Prototypes not initialized.");
int i, j, k, t;
// i: index for clusters
// j: index for data objects
// k: index for dimensions, others
// t: index for iterations
double distanceSum =
0.0d; // the sum_i dist[i][l]^{2/(1-fuzzifier)}: the sum of all parametrised distances for
// one cluster l
double doubleTMP = 0.0d; // a temporarly variable for multiple perpuses
double maxPrototypeMovement = 0.0d;
ArrayList<T> newPrototypePosition = new ArrayList<T>(this.getClusterCount());
for (i = 0; i < this.getClusterCount(); i++)
newPrototypePosition.add(this.vs.getNewAddNeutralElement());
double[] fuzzDistances = new double[this.getClusterCount()];
double[] membershipValues = new double[this.getClusterCount()];
double[] membershipSum = new double[this.getClusterCount()];
T tmpX = this.vs.getNewAddNeutralElement();
double minDistValue = 0.0d;
int[] zeroDistanceIndexList = new int[this.getClusterCount()];
int zeroDistanceCount;
for (t = 0; t < steps; t++) {
// reset values
maxPrototypeMovement = 0.0d;
for (i = 0; i < this.getClusterCount(); i++) {
this.vs.resetToAddNeutralElement(newPrototypePosition.get(i));
membershipSum[i] = 0.0d;
}
// update membership values
for (j = 0; j < this.getDataCount(); j++) {
for (i = 0; i < this.getClusterCount(); i++) zeroDistanceIndexList[i] = -1;
zeroDistanceCount = 0;
distanceSum = 0.0d;
minDistValue = Double.MAX_VALUE;
for (i = 0; i < this.getClusterCount(); i++) {
doubleTMP =
this.dist.distanceSq(this.data.get(j).element, this.prototypes.get(i).getPosition());
fuzzDistances[i] = doubleTMP;
if (minDistValue > doubleTMP) minDistValue = doubleTMP;
}
minDistValue *= this.distanceMultiplierConstant;
for (i = 0; i < this.getClusterCount(); i++) {
doubleTMP = fuzzDistances[i] - minDistValue;
if (doubleTMP <= 0.0d) {
doubleTMP = 0.0d;
zeroDistanceIndexList[zeroDistanceCount] = i;
zeroDistanceCount++;
} else {
doubleTMP = 1.0d / doubleTMP;
fuzzDistances[i] = doubleTMP;
distanceSum += doubleTMP;
}
}
// special case handling: if one (or more) prototype sits on top of a data object
if (zeroDistanceCount > 0) {
for (i = 0; i < this.getClusterCount(); i++) {
membershipValues[i] = 0.0d;
}
doubleTMP = 1.0d / ((double) zeroDistanceCount);
for (k = 0; k < zeroDistanceCount; k++) {
membershipValues[zeroDistanceIndexList[k]] = doubleTMP;
}
} else {
for (i = 0; i < this.getClusterCount(); i++) {
doubleTMP = fuzzDistances[i] / distanceSum;
membershipValues[i] = doubleTMP;
}
}
for (i = 0; i < this.getClusterCount(); i++) {
doubleTMP = MyMath.pow(membershipValues[i], this.fuzzifier);
membershipSum[i] += doubleTMP;
this.vs.copy(tmpX, this.data.get(j).element);
this.vs.mul(tmpX, doubleTMP);
this.vs.add(newPrototypePosition.get(i), tmpX);
}
}
// update prototype positions
for (i = 0; i < this.getClusterCount(); i++) {
doubleTMP = 1.0d / membershipSum[i];
this.vs.mul(newPrototypePosition.get(i), doubleTMP);
}
// copy new prototype values into prototypes wrt. learning factor
for (i = 0; i < this.getClusterCount(); i++) {
if (Math.abs(this.learningFactor - 1.0d) > 0.01d) {
this.vs.sub(newPrototypePosition.get(i), this.prototypes.get(i).getPosition());
this.vs.mul(newPrototypePosition.get(i), this.learningFactor);
this.vs.add(newPrototypePosition.get(i), this.prototypes.get(i).getPosition());
}
doubleTMP =
this.dist.distanceSq(this.prototypes.get(i).getPosition(), newPrototypePosition.get(i));
maxPrototypeMovement =
(doubleTMP > maxPrototypeMovement) ? doubleTMP : maxPrototypeMovement;
this.prototypes.get(i).moveTo(newPrototypePosition.get(i));
}
this.iterationComplete();
if (maxPrototypeMovement < this.epsilon * this.epsilon) break;
}
}