protected double[][] EstimateGamma(double[][][] xi,
ForwardBackwardCalculator fbc)
{
double[][] gamma = EngineArray.AllocateDouble2D(xi.Length + 1, xi[0].Length);
for (int t = 0; t < (xi.Length + 1); t++)
{
EngineArray.Fill(gamma[t], 0.0);
}
for (int t = 0; t < xi.Length; t++)
{
for (int i = 0; i < xi[0].Length; i++)
{
for (int j = 0; j < xi[0].Length; j++)
{
gamma[t][i] += xi[t][i][j];
}
}
}
for (int j = 0; j < xi[0].Length; j++)
{
for (int i = 0; i < xi[0].Length; i++)
{
gamma[xi.Length][j] += xi[xi.Length - 1][i][j];
}
}
return(gamma);
}
public override double[][][] EstimateXi(IMLDataSet sequence,
ForwardBackwardCalculator fbc, HiddenMarkovModel hmm)
{
if (sequence.Count <= 1)
{
throw new EncogError(
"Must have more than one observation");
}
double[][][] xi = EngineArray.AllocDouble3D((int)sequence.Count - 1, hmm.StateCount, hmm.StateCount);
double probability = fbc.Probability();
for (int t = 0; t < (sequence.Count - 1); t++)
{
IMLDataPair o = sequence[t + 1];
for (int i = 0; i < hmm.StateCount; i++)
{
for (int j = 0; j < hmm.StateCount; j++)
{
xi[t][i][j] = (fbc.AlphaElement(t, i)
* hmm.TransitionProbability[i][j]
* hmm.StateDistributions[j].Probability(o) * fbc
.BetaElement(t + 1, j)) / probability;
}
}
}
return(xi);
}
public abstract double[][][] EstimateXi(IMLDataSet sequence,
ForwardBackwardCalculator fbc, HiddenMarkovModel hmm);
public void Iteration()
{
HiddenMarkovModel nhmm;
nhmm = _method.Clone();
var allGamma = new double[_training.SequenceCount][][];
double[][] aijNum = EngineArray.AllocateDouble2D(_method.StateCount, _method.StateCount);
var aijDen = new double[_method.StateCount];
EngineArray.Fill(aijDen, 0.0);
for (int i = 0; i < _method.StateCount; i++)
{
EngineArray.Fill(aijNum[i], 0.0);
}
int g = 0;
foreach (IMLDataSet obsSeq in _training.Sequences)
{
ForwardBackwardCalculator fbc = GenerateForwardBackwardCalculator(
obsSeq, _method);
double[][][] xi = EstimateXi(obsSeq, fbc, _method);
double[][] gamma = allGamma[g++] = EstimateGamma(xi, fbc);
for (int i = 0; i < _method.StateCount; i++)
{
for (int t = 0; t < (obsSeq.Count - 1); t++)
{
aijDen[i] += gamma[t][i];
for (int j = 0; j < _method.StateCount; j++)
{
aijNum[i][j] += xi[t][i][j];
}
}
}
}
for (int i = 0; i < _method.StateCount; i++)
{
if (aijDen[i] == 0.0)
{
for (int j = 0; j < _method.StateCount; j++)
{
nhmm.TransitionProbability[i][j] =
_method.TransitionProbability[i][j];
}
}
else
{
for (int j = 0; j < _method.StateCount; j++)
{
nhmm.TransitionProbability[i][j] = aijNum[i][j]
/ aijDen[i];
}
}
}
/* compute pi */
for (int i = 0; i < _method.StateCount; i++)
{
nhmm.Pi[i] = 0.0;
}
for (int o = 0; o < _training.SequenceCount; o++)
{
for (int i = 0; i < _method.StateCount; i++)
{
nhmm.Pi[i] += (allGamma[o][0][i] / _training
.SequenceCount);
}
}
/* compute pdfs */
for (int i = 0; i < _method.StateCount; i++)
{
var weights = new double[_training.Count];
double sum = 0.0;
int j = 0;
int o = 0;
foreach (IMLDataSet obsSeq in _training.Sequences)
{
for (int t = 0; t < obsSeq.Count; t++, j++)
{
sum += weights[j] = allGamma[o][t][i];
}
o++;
}
for (j--; j >= 0; j--)
{
weights[j] /= sum;
}
IStateDistribution opdf = nhmm.StateDistributions[i];
opdf.Fit(_training, weights);
}
_method = nhmm;
Iterations++;
}
