@Override
public double[] getLogLikelihoods(List<Beacon> beacons, State[] locations) {
List<Beacon> beaconsCleansed = Beacon.filterBeacons(beacons, minRssi, maxRssi);
beaconsCleansed = beaconFilter.setBLEBeacon(bleBeacons).filter(beaconsCleansed, locations);
BLEBeacon.setBLEBeaconIdsToMeasuredBeacons(bleBeacons, beaconsCleansed);
int[] activeBeaconList = ModelAdaptUtils.beaconsToActiveBeaconArray(beaconsCleansed);
final double[] ySub = ModelAdaptUtils.beaconsToVecSubset(beaconsCleansed);
final double[][] X = ModelAdaptUtils.locationsToMat(locations);
// Adjust bias by average bias
final double[] rssiBiases = ModelAdaptUtils.biasesToVec(locations);
double logLLs[] = null;
try {
Class<?> cls = gpLDPL.getClass();
Constructor<?> cst = cls.getConstructor(gpLDPL.getClass());
final GaussianProcessLDPLMean gpLDPLtmp = (GaussianProcessLDPLMean) cst.newInstance(gpLDPL);
gpLDPLtmp.updateByActiveBeaconList(activeBeaconList);
int n = X.length;
final double logpro[] = new double[n];
Future<?>[] futures = new Future<?>[n];
for (int i = 0; i < n; i++) {
final int idx = i;
futures[idx] =
ExecutorServiceHolder.getExecutorService()
.submit(
new Runnable() {
public void run() {
// Subtract bias from an observation vector.
double[] ySubAdjusted = ArrayUtils.addScalar(ySub, -rssiBiases[idx]);
logpro[idx] = gpLDPLtmp.logProbaygivenx(X[idx], ySubAdjusted);
}
});
}
for (int i = 0; i < n; i++) {
futures[i].get();
}
logLLs = logpro;
} catch (InstantiationException
| IllegalAccessException
| IllegalArgumentException
| InvocationTargetException
| NoSuchMethodException
| SecurityException e) {
e.printStackTrace();
} catch (InterruptedException | ExecutionException e) {
e.printStackTrace();
}
return logLLs;
}
protected void trainByAveragedSamples(List<Sample> samples) {
int nBeacons = bleBeacons.size();
BLEBeacon.setBLEBeaconIdsToSamples(bleBeacons, samples);
setupGPLDPL();
this.samples = samples;
double[][] X = ModelAdaptUtils.samplesToLocationsMat(samples);
double[][] Y = ModelAdaptUtils.samplesToBeaconsListMat(samples, minRssi, nBeacons);
double[][] bLocs = ModelAdaptUtils.BLEBeaconsToMat(bleBeacons, nBeacons);
gpLDPL.setSourceLocs(bLocs);
int[] actBecList = new int[nBeacons];
for (int i = 0; i < nBeacons; i++) actBecList[i] = i;
gpLDPL.setActiveBeaconList(actBecList);
gpLDPL.fit(X, Y);
}
public List<Beacon> predictBeacons(State state, List<Beacon> beacons) {
double[] x = ModelAdaptUtils.locationToVec(state);
int[] activeBeaconList = ModelAdaptUtils.beaconsToActiveBeaconArray(beacons);
double[] ypred;
synchronized (gpLDPL) {
gpLDPL.updateByActiveBeaconList(activeBeaconList);
ypred = gpLDPL.predictygivenx(x);
}
List<Beacon> beaconsPred = new ArrayList<Beacon>();
for (int i = 0; i < activeBeaconList.length; i++) {
int major = -1;
int minor = activeBeaconList[i];
double rssi = ypred[i];
Beacon beacon = new Beacon(major, minor, (float) rssi);
beaconsPred.add(beacon);
}
return beaconsPred;
}
protected void trainWithHyperParameters(List<List<Sample>> samplesListTrainParent) {
List<Sample> samplesAvaragedTrain = Sample.meanTrimList(samplesListTrainParent, minRssi);
this.train(samplesAvaragedTrain);
// Update active beacon list about actually observed beacons
Sample samplesAllAverage = Sample.mean(samplesAvaragedTrain, minRssi);
List<Beacon> observedBeacons = samplesAllAverage.getBeacons();
this.setActiveBeaconList(ModelAdaptUtils.beaconsToActiveBeaconArray(observedBeacons));
double stdevTmp = stdev;
this.setStdev(0.0); // stdev is set to be zero before optimization of LDPL model
this.optimizeForLDPLModel();
this.setStdev(stdevTmp); // stdev must be updated before optimization of GP
this.optimizeForGPIsoScaleLOOMSE(); // optimize for lx(=ly) and stdev
this.optimizeEstimateSigmaN(
Sample.samplesListToSamples(samplesListTrainParent)); // estimate sigmaN
this.train(samplesAvaragedTrain);
}