private double[] PredictNextValue(IGaussianMixtureModelState model, IPredictionRequest request, double[][] trainingSet)
{
var N = trainingSet.Length;
var K = trainingSet[0].Length;
var result = new double[K];
var yesterday = trainingSet[N - 1];
var yesterdayLikelihood = LogLikelihood.Calculate(new[] { yesterday }, model.Mixture.Coefficients, GetCovariances(model.Mixture.Components), GetMeans(model.Mixture.Components));
Debug.WriteLine("Yesterday Likelihood : " + new Vector(yesterday) + " : " + yesterdayLikelihood + " ");
var guessess = FindMostSimilarObservations(model, trainingSet, yesterdayLikelihood, request.Tolerance);
var bestGuessPlace = FindBestGuess(request, guessess);
var tomorrow = trainingSet[bestGuessPlace.PlaceInSequence + 1];
var mostSimilar = trainingSet[bestGuessPlace.PlaceInSequence];
for (var k = 0; k < K; k++)
{
if (bestGuessPlace.PlaceInSequence != trainingSet.Length)
{
result[k] = yesterday[k] + (tomorrow[k] - mostSimilar[k]);
}
}
Debug.WriteLine("Predicted : " + new Vector(result) + " : " + LogLikelihood.Calculate(new[] { result }, model.Mixture.Coefficients, GetCovariances(model.Mixture.Components), GetMeans(model.Mixture.Components)));
return(result);
}
/// <summary>
/// Fits Distribution parameters to given observation vector
/// </summary>
/// <param name="observations">Vector</param>
/// <param name="weights">Vector</param>
/// <param name="likelihood">Output value</param>
/// <returns>IDistribution</returns>
public override IDistribution Evaluate(double[][] observations, double[] weights, out double likelihood)
{
var mean = Utils.Mean(observations, weights);
var covariance = Utils.Covariance(observations, mean, weights);
var result = new NormalDistribution(mean, covariance);
likelihood = LogLikelihood.Calculate(observations, result.Covariance, result.Mean);
return(result);
}
public IDistribution Evaluate(double[][] observations, double[] weights, double treshold, out double likelihood)
{
// Calculate Maximization of the Likelihood
_stepsTillConvirgence = 0;
var convirged = false;
var pdf = (IMultivariateDistribution[])_components.Clone();
var K = _coefficients.Length;
var N = observations.Length;
//// Initialize pi, covariance matrix and mean
var pi = (double[])_coefficients.Clone();
weights = weights.Product(N);
likelihood = LogLikelihood.Calculate(observations, _coefficients, GetCovariances(pdf), GetMeans(pdf));
while (!convirged)
{
//// Expectation
var gamma = CalculateGamma(observations, weights, pdf, pi);
//// Maximization
pi = CalculatePi(gamma);
//// For each component train new distribution function
for (var k = 0; k < K; k++)
{
var mean = CalculateMean(gamma, observations, k);
var covariance = CalculateCovariance(gamma, observations, mean, k);
pdf[k] = new NormalDistribution(mean, covariance);
}
//// Check treshold
var newLikelihood = LogLikelihood.Calculate(observations, _coefficients, GetCovariances(pdf), GetMeans(pdf));
if (double.IsNaN(newLikelihood) || double.IsInfinity(newLikelihood))
{
throw new ApplicationException("EM algorithm does not convirged");
}
convirged = (newLikelihood - likelihood) <= treshold;
likelihood = newLikelihood;
_stepsTillConvirgence++;
}
return(new Mixture <IMultivariateDistribution>(pi, pdf));
}
private IList <ObservationWithLikelihood <double[]> > FindMostSimilarObservations(IGaussianMixtureModelState model, double[][] trainingSet, double yesterdayLikelihood, double tolerance)
{
var N = trainingSet.Length;
var guessess = new List <ObservationWithLikelihood <double[]> >();
for (var n = N - 2; n > 0; n--)
{
var x = new[] { trainingSet[n] };
var likelihood = LogLikelihood.Calculate(x, model.Mixture.Coefficients, GetCovariances(model.Mixture.Components), GetMeans(model.Mixture.Components));
//Debug.Write((new Vector(observations[n])).ToString() + " : " + likelihood + " " + Environment.NewLine);
if (Math.Abs(yesterdayLikelihood) - tolerance < Math.Abs(likelihood) && Math.Abs(yesterdayLikelihood) + tolerance > Math.Abs(likelihood))
{
guessess.Add(new ObservationWithLikelihood <double[]>()
{
LogLikelihood = likelihood, Observation = trainingSet[n], PlaceInSequence = n - 1
});
}
}
return(guessess);
}