@Override
public EF_LearningBayesianNetwork initModel(
EF_LearningBayesianNetwork bayesianNetwork, PlateuStructure plateuStructure) {
for (Variable paramVariable :
bayesianNetwork.getParametersVariables().getListOfParamaterVariables()) {
if (!paramVariable.isNormalParameter()) continue;
// if (paramVariable.getName().contains("_Beta0_"))
// continue;
EF_Normal prior = bayesianNetwork.getDistribution(paramVariable);
double varPrior = 1;
double precisionPrior = 1 / varPrior;
double meanPrior = 0;
prior.getNaturalParameters().set(0, precisionPrior * meanPrior);
prior.getNaturalParameters().set(1, -0.5 * precisionPrior);
prior.fixNumericalInstability();
prior.updateMomentFromNaturalParameters();
}
for (Variable localVar : this.localHiddenVars) {
EF_NormalGamma normal = bayesianNetwork.getDistribution(localVar);
Variable gammaVar = normal.getGammaParameterVariable();
EF_Gamma gamma = bayesianNetwork.getDistribution(gammaVar);
int initVariance = 1;
double alpha = 1000;
double beta = alpha * initVariance;
gamma.getNaturalParameters().set(0, alpha - 1);
gamma.getNaturalParameters().set(1, -beta);
gamma.fixNumericalInstability();
gamma.updateMomentFromNaturalParameters();
Variable meanVar = normal.getMeanParameterVariable();
EF_Normal meanDist = bayesianNetwork.getDistribution(meanVar);
double mean = meanStart;
double var = initVariance;
meanDist.getNaturalParameters().set(0, mean / (var));
meanDist.getNaturalParameters().set(1, -1 / (2 * var));
meanDist.fixNumericalInstability();
meanDist.updateMomentFromNaturalParameters();
}
return bayesianNetwork;
}
@Override
public EF_LearningBayesianNetwork transitionModel(
EF_LearningBayesianNetwork bayesianNetwork, PlateuStructure plateuStructure) {
for (Variable localVar : this.localHiddenVars) {
Normal normalGlobalHiddenPreviousTimeStep =
plateuStructure.getEFVariablePosterior(localVar, 0).toUnivariateDistribution();
EF_NormalGamma normal = bayesianNetwork.getDistribution(localVar);
Variable gammaVar = normal.getGammaParameterVariable();
EF_Gamma gamma = bayesianNetwork.getDistribution(gammaVar);
double variance = normalGlobalHiddenPreviousTimeStep.getVariance() + this.transtionVariance;
double alpha = 1000;
double beta = alpha * variance;
gamma.getNaturalParameters().set(0, alpha - 1);
gamma.getNaturalParameters().set(1, -beta);
gamma.fixNumericalInstability();
gamma.updateMomentFromNaturalParameters();
Variable meanVar = normal.getMeanParameterVariable();
EF_Normal meanDist = bayesianNetwork.getDistribution(meanVar);
double mean = normalGlobalHiddenPreviousTimeStep.getMean();
meanDist.getNaturalParameters().set(0, mean / (variance));
meanDist.getNaturalParameters().set(1, -1 / (2 * variance));
meanDist.fixNumericalInstability();
meanDist.updateMomentFromNaturalParameters();
}
/** *** FADING *** */
if (fading < 1.0) {
bayesianNetwork
.getParametersVariables()
.getListOfParamaterVariables()
.stream()
.forEach(
var -> {
EF_BaseDistribution_MultinomialParents dist = bayesianNetwork.getDistribution(var);
EF_UnivariateDistribution prior = dist.getBaseEFUnivariateDistribution(0);
NaturalParameters naturalParameters = prior.getNaturalParameters();
naturalParameters.multiplyBy(fading);
prior.setNaturalParameters(naturalParameters);
dist.setBaseEFDistribution(0, prior);
});
}
return bayesianNetwork;
}