/**
* Constructor. The Gaussian mixture model will be learned from the given data with the EM
* algorithm. The number of components will be selected by BIC.
*
* @param data the training data.
*/
@SuppressWarnings("unchecked")
public GaussianMixture(double[] data) {
if (data.length < 20) throw new IllegalArgumentException("Too few samples.");
ArrayList<Component> mixture = new ArrayList<Component>();
Component c = new Component();
c.priori = 1.0;
c.distribution = new GaussianDistribution(data);
mixture.add(c);
int freedom = 0;
for (int i = 0; i < mixture.size(); i++) freedom += mixture.get(i).distribution.npara();
double bic = 0.0;
for (double x : data) {
double p = c.distribution.p(x);
if (p > 0) bic += Math.log(p);
}
bic -= 0.5 * freedom * Math.log(data.length);
double b = Double.NEGATIVE_INFINITY;
while (bic > b) {
b = bic;
components = (ArrayList<Component>) mixture.clone();
split(mixture);
bic = EM(mixture, data);
freedom = 0;
for (int i = 0; i < mixture.size(); i++) freedom += mixture.get(i).distribution.npara();
bic -= 0.5 * freedom * Math.log(data.length);
}
}
/** Round the bounds for axis i. */
public void extendBound(int i) {
if (i < 0 || i >= dimension) {
throw new IllegalArgumentException("Invalid bound index: " + i);
}
extendBound[i] = true;
lowerBound[i] = precisionUnit[i] * (Math.floor(originalLowerBound[i] / precisionUnit[i]));
upperBound[i] = precisionUnit[i] * (Math.ceil(originalUpperBound[i] / precisionUnit[i]));
}
/**
* Constructor. The Gaussian mixture model will be learned from the given data with the EM
* algorithm.
*
* @param data the training data.
* @param k the number of components.
*/
public GaussianMixture(double[] data, int k) {
if (k < 2) throw new IllegalArgumentException("Invalid number of components in the mixture.");
double min = Math.min(data);
double max = Math.max(data);
double step = (max - min) / (k + 1);
for (int i = 0; i < k; i++) {
Component c = new Component();
c.priori = 1.0 / k;
c.distribution = new GaussianDistribution(min += step, step);
components.add(c);
}
EM(components, data);
}
public ExponentialDistributionDemo() {
super(new BorderLayout());
Hashtable<Integer, JLabel> labelTable = new Hashtable<Integer, JLabel>();
for (int i = 0; i <= 50; i += 10) {
labelTable.put(new Integer(i), new JLabel(String.valueOf(i / 10)));
}
lambdaSlider = new JSlider(0, 50, (int) Math.round(lambda * 10));
lambdaSlider.addChangeListener(this);
lambdaSlider.setLabelTable(labelTable);
lambdaSlider.setMajorTickSpacing(10);
lambdaSlider.setMinorTickSpacing(2);
lambdaSlider.setPaintTicks(true);
lambdaSlider.setPaintLabels(true);
optionPane = new JPanel(new FlowLayout(FlowLayout.LEFT));
optionPane.setBorder(BorderFactory.createRaisedBevelBorder());
optionPane.add(new JLabel("\u03BB:"));
optionPane.add(lambdaSlider);
add(optionPane, BorderLayout.NORTH);
canvas = new JPanel(new GridLayout(2, 2));
add(canvas, BorderLayout.CENTER);
ExponentialDistribution dist = new ExponentialDistribution(lambda);
double[][] p = new double[100][2];
double[][] q = new double[100][2];
for (int i = 0; i < p.length; i++) {
p[i][0] = i / 20.0;
p[i][1] = dist.p(p[i][0]);
q[i][0] = i / 20.0;
q[i][1] = dist.cdf(p[i][0]);
}
pdf = LinePlot.plot(p, Line.Style.SOLID, Color.BLUE);
pdf.setTitle("PDF");
canvas.add(pdf);
cdf = LinePlot.plot(q, Line.Style.SOLID, Color.BLUE);
cdf.setTitle("CDF");
canvas.add(cdf);
double[] data = new double[500];
for (int i = 0; i < data.length; i++) {
data[i] = dist.rand();
}
histogram = Histogram.plot(data, 20);
histogram.setTitle("Histogram");
canvas.add(histogram);
qqplot = QQPlot.plot(data, dist);
qqplot.setTitle("Q-Q Plot");
canvas.add(qqplot);
}
/** Set the precision unit for axis i. */
void setPrecisionUnit(int i) {
if (upperBound[i] > lowerBound[i]) {
double digits = Math.log10(Math.abs(upperBound[i] - lowerBound[i]));
double residual = digits - Math.floor(digits);
if (residual < 0.2) {
// If the range is less than 15 units, we reduce one level.
digits -= 1.0;
}
precisionDigits[i] = (int) Math.floor(digits);
precisionUnit[i] = Math.pow(10, precisionDigits[i]);
if (residual >= 0.2 && residual <= 0.7) {
// In case of too few grids, we use a half of precision unit.
precisionUnit[i] /= 2;
precisionDigits[i] -= 1;
}
} else {
precisionUnit[i] = 0.1;
}
}