public static GradientDescentResult Mlat(Matrix <double> anchorsIn, Vector <double> rangesIn,
Matrix <double> boundsIn = null, int nTrial = 200, double alpha = 0.01, double timeThreshold = 0)
{
GradientDescentResult gradientDescentResult = GradientDescent(anchorsIn, rangesIn, boundsIn, nTrial, alpha, timeThreshold);
int idx = -1;
double error = double.MaxValue;
for (int i = 0; i < gradientDescentResult.Error.Count; i++)
{
if (gradientDescentResult.Error[i] < error)
{
idx = i;
error = gradientDescentResult.Error[i];
}
}
gradientDescentResult.Estimator = gradientDescentResult.EstimatorCandidate.Row(idx);
return(gradientDescentResult);
}
private static GradientDescentResult GradientDescent(Matrix <double> anchorsIn, Vector <double> rangesIn,
Matrix <double> boundsIn = null, int nTrial = 100, double alpha = 0.001, double timeThreshold = 0)
{
System.Random random = new SystemRandomSource();
int n = anchorsIn.RowCount;
int dim = anchorsIn.ColumnCount;
var gradientDescentResult = new GradientDescentResult(nTrial, dim);
if (boundsIn == null)
{
boundsIn = DenseMatrix.Build.Dense(1, dim);
}
Matrix <double> boundsTemp = anchorsIn.Stack(boundsIn);
Matrix <double> bounds = DenseMatrix.Build.Dense(2, dim);
for (int i = 0; i < dim; i++)
{
bounds[0, i] = boundsTemp.Column(i).Min();
bounds[1, i] = boundsTemp.Column(i).Max();
}
if (timeThreshold == 0)
{
timeThreshold = 1.0 / nTrial;
}
Vector <double> ranges = DenseVector.Build.Dense(n);
for (int i = 0; i < nTrial; i++)
{
Vector <double> estimator0 = DenseVector.Build.Dense(dim);
for (int j = 0; j < dim; j++)
{
estimator0[j] = random.NextDouble() * (bounds[1, j] - bounds[0, j]) + bounds[0, j];
}
Vector <double> estimator = DenseVector.OfVector(estimator0);
Stopwatch stopwatch = new Stopwatch();
while (true)
{
for (int j = 0; j < n; j++)
{
ranges[j] = GetEuclidean(anchorsIn.Row(j), estimator);
}
double error = GetEuclidean(rangesIn, ranges);
Vector <double> delta = DenseVector.Build.Dense(dim);
for (int j = 0; j < n; j++)
{
delta += (rangesIn[j] - ranges[j]) / ranges[j] * (estimator - anchorsIn.Row(j));
}
delta *= 2 * alpha;
Vector <double> estimatorNext = estimator - delta;
for (int j = 0; j < n; j++)
{
ranges[j] = MLAT.GetEuclidean(anchorsIn.Row(j), estimatorNext);
}
double errorNext = MLAT.GetEuclidean(rangesIn, ranges);
if (errorNext < error)
{
estimator = estimatorNext;
}
else
{
gradientDescentResult.EstimatorCandidate.SetRow(i, estimator);
gradientDescentResult.Error[i] = error;
break;
}
if (stopwatch.ElapsedMilliseconds > timeThreshold)
{
gradientDescentResult.Error[i] = double.MaxValue;
break;
}
}
}
return(gradientDescentResult);
}