protected override ConditionalGaussianDistributionParams GetPlainConditionalGaussianDistribution(BranchOrLeaf branchOrLeaf, OptimizationParameterList gaussianParameters)
{
double variance = gaussianParameters["Variance"].Value;
//double alphaTimesVariance = gaussianParameters["AlphaVariance"].Value;
double mean = gaussianParameters["Mean"].Value;
double alpha = gaussianParameters["Alpha"].Value;
double fOfBranchLength = FOfBranchLength(branchOrLeaf.Length, alpha);
//ConditionalGaussianDistributionParams plainConditionalDistribution = ConditionalGaussianDistributionParams.GetInstance();
//double root1MinusAlphaTimesBranchLength = Math.Sqrt(1 - alpha * branchOrLeaf.Length);
double root1MinusFofBranchLength = Math.Sqrt(1 - FOfBranchLength(branchOrLeaf.Length, alpha));
// ax + b, Mean := b
double meanForGaussDistParams = mean * (1.0 - root1MinusFofBranchLength);
//double meanForGausDistParams = mean * (1.0 - root1MinusAlphaTimesBranchLength);
// ax + b = x, LinearCoefficent := a
double linearCoeffForGaussDistParams = root1MinusFofBranchLength;
//double linearCoeffForGausDistParams = root1MinusAlphaTimesBranchLength;
//double varForGausDistParams = alphaTimesVariance * branchOrLeaf.Length;
double varianceForGaussDistParams = fOfBranchLength * variance;
ConditionalGaussianDistributionParams plainConditionalDistribution =
ConditionalGaussianDistributionParams.GetInstance(meanForGaussDistParams, varianceForGaussDistParams, linearCoeffForGaussDistParams);
return(plainConditionalDistribution);
}
public ConditionalGaussianDistributionParams CreateDistributionGaussianOrNull(
Leaf leaf,
OptimizationParameterList gaussianParameters,
Converter <Leaf, SufficientStatistics> predictorLeafToBoolStats)
{
if (leaf.Length == 0)
{
Debug.WriteLine("Branch length of zero observed");
}
// TODO: make continuous
BooleanStatistics hasPredictor = (BooleanStatistics)predictorLeafToBoolStats(leaf);
if (hasPredictor.IsMissing())
{
ConditionalGaussianDistributionParams.GetNullInstance();
//return null; // Predictor data is missing, so skip this leaf.
}
ConditionalGaussianDistributionParams plainConditionalDistribution = GetPlainConditionalGaussianDistribution(leaf, gaussianParameters);
if (!hasPredictor)
{
return(plainConditionalDistribution);
}
else
{
double delta = GetOffset(gaussianParameters);
ConditionalGaussianDistributionParams offsetConditionalDistribution = plainConditionalDistribution.AddOffsetToMean(delta);
return(offsetConditionalDistribution);
}
//return CreateDistributionGaussianOrNull(plainConditionalDistribution, hasPredictor, delta);
}
protected override ConditionalGaussianDistributionParams GetPlainConditionalGaussianDistribution(BranchOrLeaf branchOrLeaf, OptimizationParameterList gaussianParameters)
{
double variance = gaussianParameters["Variance"].Value;
ConditionalGaussianDistributionParams plainConditionalDistribution =
ConditionalGaussianDistributionParams.GetInstance(0, variance, 1);
return(plainConditionalDistribution);
}
public ConditionalGaussianDistributionParams CreateDistributionGaussianOrNull(Branch branch, OptimizationParameterList gaussianParameters)
{
if (branch.Length == 0)
{
Debug.WriteLine("Branch length of zero observed");
}
ConditionalGaussianDistributionParams plainConditionalDistribution = GetPlainConditionalGaussianDistribution(branch, gaussianParameters);
return(plainConditionalDistribution);
}
public override IMessage InitializeMessage(Leaf leaf, OptimizationParameterList gaussianParameters)
{
if (IsMissing(leaf))
{
return(null);
}
//ConditionalGaussianDistribution distOrNull = GaussianDistribution.CreateDistributionGaussianOrNull(leaf, gaussianParameters, _caseNameToHasPredictor);
ConditionalGaussianDistributionParams distOrNull = GaussianDistribution.CreateDistributionGaussianOrNull(leaf, gaussianParameters, LeafToPredictorStatistics);
if (distOrNull.IsNull)
{
return(null);
}
MessageGaussian message = Initalize(distOrNull, leaf, gaussianParameters);
return(message);
}
protected override ConditionalGaussianDistributionParams GetPlainConditionalGaussianDistribution(BranchOrLeaf branchOrLeaf, OptimizationParameterList gaussianParameters)
{
// NOTE: For BM, interpret Alpha as AlphaTimesVariance. Alpha is only used to multiply against variance, and searching is easier we we do
// it this way, as otherwise alphaTimesVariance is affected by both alpha and variance when we're searching.
double alphaTimesVariance = gaussianParameters["Alpha"].Value;
//double alpha = gaussianParameters["Alpha"].Value;
//double variance = gaussianParameters["Variance"].Value;
// 0 (const part of mean) ax + b, Mean := b
double mean = 0;
// 1 (ax + b = x), linearCoefficent := a.
double linearCoefficient = 1;
//double conditionalVariance = alpha * variance * branchOrLeaf.Length;
double conditionalVariance = alphaTimesVariance * branchOrLeaf.Length;
ConditionalGaussianDistributionParams plainConditionalDistribution =
ConditionalGaussianDistributionParams.GetInstance(mean, conditionalVariance, linearCoefficient);
return(plainConditionalDistribution);
}
private MessageGaussian Initalize(ConditionalGaussianDistributionParams dist, Leaf leaf, OptimizationParameterList gaussianParameters)
{
Debug.Assert(!LeafToTargetStatistics(leaf).IsMissing());
GaussianStatistics gaussianStatistics = (GaussianStatistics)LeafToTargetStatistics(leaf);
SpecialFunctions.CheckCondition(gaussianStatistics != null, "why is caseNameToTargetOrNull unknown?");
if (gaussianStatistics.SampleSize == 1)
{
double z = gaussianStatistics.Mean;
double logK = -Math.Log(dist.LinearCoefficient);
Debug.Assert(!double.IsNaN(logK)); // real assert - in release mode, the NaN is OK.
double a = (z - dist.Mean) / dist.LinearCoefficient;
double v = dist.Variance / Math.Pow(dist.LinearCoefficient, 2);
MessageGaussian message = MessageGaussian.GetInstance(logK, a, v);
return(message);
}
else
{
double vNoise = GaussianDistribution.GetSamplingVariance(gaussianParameters);
double logKMult = (
-Math.Log(gaussianStatistics.SampleSize)
- gaussianStatistics.SampleSize
* (gaussianStatistics.Variance + vNoise * Log2PI + vNoise * Math.Log(vNoise))
/ vNoise
+ Math.Log(2 * Math.PI * vNoise)
) / 2.0;
double aMult = gaussianStatistics.Mean;
double vMult = vNoise / (double)gaussianStatistics.SampleSize;
double logK = logKMult - Math.Log(dist.LinearCoefficient);
double a = (aMult - dist.Mean) / dist.LinearCoefficient;
double v = (vMult + dist.Variance) / Math.Pow(dist.LinearCoefficient, 2);
MessageGaussian message = MessageGaussian.GetInstance(logK, a, v);
return(message);
}
}
internal ConditionalGaussianDistributionParams AddOffsetToMean(double delta)
{
return(ConditionalGaussianDistributionParams.GetInstance(Mean + delta, Variance, LinearCoefficient));
}
