Examples with GaussianDistribution jskills.numerics.GaussianDistribution used on opensource projects

Example 1 with GaussianDistribution

use of jskills.numerics.GaussianDistribution in project ACManager by kun368.

the class PlayerPriorValuesToSkillsLayer method buildLayer.

@Override
public void buildLayer() {
    for (ITeam currentTeam : teams) {
        List<KeyedVariable<IPlayer, GaussianDistribution>> currentTeamSkills = new ArrayList<>();
        for (Entry<IPlayer, Rating> currentTeamPlayer : currentTeam.entrySet()) {
            KeyedVariable<IPlayer, GaussianDistribution> playerSkill = createSkillOutputVariable(currentTeamPlayer.getKey());
            AddLayerFactor(createPriorFactor(currentTeamPlayer.getKey(), currentTeamPlayer.getValue(), playerSkill));
            currentTeamSkills.add(playerSkill);
        }
        addOutputVariableGroup(currentTeamSkills);
    }
}

Example 2 with GaussianDistribution

use of jskills.numerics.GaussianDistribution in project ACManager by kun368.

the class TeamDifferencesComparisonLayer method buildLayer.

@Override
public void buildLayer() {
    for (int i = 0; i < getInputVariablesGroups().size(); i++) {
        boolean isDraw = (teamRanks[i] == teamRanks[i + 1]);
        Variable<GaussianDistribution> teamDifference = getInputVariablesGroups().get(i).get(0);
        GaussianFactor factor = isDraw ? (GaussianFactor) new GaussianWithinFactor(epsilon, teamDifference) : new GaussianGreaterThanFactor(epsilon, teamDifference);
        AddLayerFactor(factor);
    }
}

Example 3 with GaussianDistribution

use of jskills.numerics.GaussianDistribution in project ACManager by kun368.

the class GaussianGreaterThanFactor method getLogNormalization.

@Override
public double getLogNormalization() {
    GaussianDistribution marginal = getVariables().get(0).getValue();
    GaussianDistribution message = getMessages().get(0).getValue();
    GaussianDistribution messageFromVariable = divide(marginal, message);
    return -logProductNormalization(messageFromVariable, message) + Math.log(cumulativeTo((messageFromVariable.getMean() - epsilon) / messageFromVariable.getStandardDeviation()));
}

Example 4 with GaussianDistribution

use of jskills.numerics.GaussianDistribution in project ACManager by kun368.

the class GaussianPriorFactor method updateMessage.

@Override
protected double updateMessage(Message<GaussianDistribution> message, Variable<GaussianDistribution> variable) {
    GaussianDistribution oldMarginal = new GaussianDistribution(variable.getValue());
    Message<GaussianDistribution> oldMessage = message;
    GaussianDistribution newMarginal = GaussianDistribution.fromPrecisionMean(oldMarginal.getPrecisionMean() + newMessage.getPrecisionMean() - oldMessage.getValue().getPrecisionMean(), oldMarginal.getPrecision() + newMessage.getPrecision() - oldMessage.getValue().getPrecision());
    variable.setValue(newMarginal);
    message.setValue(newMessage);
    return sub(oldMarginal, newMarginal);
}

Example 5 with GaussianDistribution

use of jskills.numerics.GaussianDistribution in project ACManager by kun368.

the class GaussianWeightedSumFactor method UpdateHelper.

private double UpdateHelper(double[] weights, double[] weightsSquared, List<Message<GaussianDistribution>> messages, List<Variable<GaussianDistribution>> variables) {
    // Potentially look at http://mathworld.wolfram.com/NormalSumDistribution.html for clues as
    // to what it's doing
    GaussianDistribution message0 = new GaussianDistribution(messages.get(0).getValue());
    GaussianDistribution marginal0 = new GaussianDistribution(variables.get(0).getValue());
    // The math works out so that 1/newPrecision = sum of a_i^2 /marginalsWithoutMessages[i]
    double inverseOfNewPrecisionSum = 0.0;
    double anotherInverseOfNewPrecisionSum = 0.0;
    double weightedMeanSum = 0.0;
    double anotherWeightedMeanSum = 0.0;
    for (int i = 0; i < weightsSquared.length; i++) {
        // These flow directly from the paper
        inverseOfNewPrecisionSum += weightsSquared[i] / (variables.get(i + 1).getValue().getPrecision() - messages.get(i + 1).getValue().getPrecision());
        GaussianDistribution diff = divide(variables.get(i + 1).getValue(), messages.get(i + 1).getValue());
        anotherInverseOfNewPrecisionSum += weightsSquared[i] / diff.getPrecision();
        weightedMeanSum += weights[i] * (variables.get(i + 1).getValue().getPrecisionMean() - messages.get(i + 1).getValue().getPrecisionMean()) / (variables.get(i + 1).getValue().getPrecision() - messages.get(i + 1).getValue().getPrecision());
        anotherWeightedMeanSum += weights[i] * diff.getPrecisionMean() / diff.getPrecision();
    }
    double newPrecision = 1.0 / inverseOfNewPrecisionSum;
    double anotherNewPrecision = 1.0 / anotherInverseOfNewPrecisionSum;
    double newPrecisionMean = newPrecision * weightedMeanSum;
    double anotherNewPrecisionMean = anotherNewPrecision * anotherWeightedMeanSum;
    GaussianDistribution oldMarginalWithoutMessage = divide(marginal0, message0);
    GaussianDistribution newMessage = GaussianDistribution.fromPrecisionMean(newPrecisionMean, newPrecision);
    GaussianDistribution anotherNewMessage = GaussianDistribution.fromPrecisionMean(anotherNewPrecisionMean, anotherNewPrecision);
    if (!newMessage.equals(anotherNewMessage))
        throw new RuntimeException("newMessage and anotherNewMessage aren't the same");
    GaussianDistribution newMarginal = mult(oldMarginalWithoutMessage, newMessage);
    // Update the message and marginal
    messages.get(0).setValue(newMessage);
    variables.get(0).setValue(newMarginal);
    // Return the difference in the new marginal
    return sub(newMarginal, marginal0);
}