public void F64Matrix_GetColumns()
{
// matrix is created row wise
var matrix = new F64Matrix(new double[] { 1, 2, 3,
4, 5, 6,
7, 8, 9 }, 3, 3);
// returns selected columns a a new matrix
var cols = matrix.Columns(new int[] { 0, 2 }); // [1, 3,
// 4, 6
// 7, 9]
}
double SelectNextModelToRemove(F64Matrix crossValidatedModelPredictions,
double[] targets,
double currentBestError)
{
var candidateModelMatrix = new F64Matrix(crossValidatedModelPredictions.RowCount,
m_remainingModelIndices.Count - 1);
var candidatePredictions = new double[crossValidatedModelPredictions.RowCount];
var candidateModelIndices = new int[m_remainingModelIndices.Count - 1];
var bestError = double.MaxValue;
var bestIndex = -1;
foreach (var index in m_remainingModelIndices)
{
var candidateIndex = 0;
for (int i = 0; i < m_remainingModelIndices.Count; i++)
{
var curIndex = m_remainingModelIndices[i];
if (curIndex != index)
{
candidateModelIndices[candidateIndex++] = m_remainingModelIndices[i];
}
}
crossValidatedModelPredictions.Columns(candidateModelIndices, candidateModelMatrix);
m_ensembleStrategy.Combine(candidateModelMatrix, candidatePredictions);
var error = m_metric.Error(targets, candidatePredictions);
if (error < bestError)
{
bestError = error;
bestIndex = index;
}
}
m_remainingModelIndices.Remove(bestIndex);
return(bestError);
}
double SelectNextModelToAdd(F64Matrix crossValidatedModelPredictions, double[] targets, double currentBestError)
{
var candidateModelMatrix = new F64Matrix(crossValidatedModelPredictions.RowCount, m_selectedModelIndices.Count + 1);
var candidatePredictions = new double[crossValidatedModelPredictions.RowCount];
var candidateModelIndices = new int[m_selectedModelIndices.Count + 1];
var bestError = currentBestError;
var bestIndex = -1;
foreach (var index in m_remainingModelIndices)
{
m_selectedModelIndices.CopyTo(candidateModelIndices);
candidateModelIndices[candidateModelIndices.Length - 1] = index;
crossValidatedModelPredictions.Columns(candidateModelIndices, candidateModelMatrix);
m_ensembleStrategy.Combine(candidateModelMatrix, candidatePredictions);
var error = m_metric.Error(targets, candidatePredictions);
if (error < bestError)
{
bestError = error;
bestIndex = index;
}
}
if (bestIndex != -1)
{
m_selectedModelIndices.Add(bestIndex);
if (!m_selectWithReplacement)
{
m_remainingModelIndices.Remove(bestIndex);
}
}
return(bestError);
}
/// <summary>
/// Iterative random selection of ensemble models.
/// </summary>
/// <param name="crossValidatedModelPredictions">cross validated predictions from multiple models.
/// Each column in the matrix corresponds to predictions from a separate model</param>
/// <param name="targets">Corresponding targets</param>
/// <returns>The indices of the selected model</returns>
public int[] Select(F64Matrix crossValidatedModelPredictions, double[] targets)
{
if (crossValidatedModelPredictions.ColumnCount < m_numberOfModelsToSelect)
{
throw new ArgumentException("Available models: " + crossValidatedModelPredictions.ColumnCount +
" is smaller than number of models to select: " + m_numberOfModelsToSelect);
}
m_allIndices = Enumerable.Range(0, crossValidatedModelPredictions.ColumnCount).ToArray();
var bestModelIndices = new int[m_numberOfModelsToSelect];
var candidateModelIndices = new int[m_numberOfModelsToSelect];
var candidateModelMatrix = new F64Matrix(crossValidatedModelPredictions.RowCount, m_numberOfModelsToSelect);
var candidatePredictions = new double[crossValidatedModelPredictions.RowCount];
var bestError = double.MaxValue;
for (int i = 0; i < m_iterations; i++)
{
SelectNextRandomIndices(candidateModelIndices);
crossValidatedModelPredictions.Columns(candidateModelIndices, candidateModelMatrix);
m_ensembleStrategy.Combine(candidateModelMatrix, candidatePredictions);
var error = m_metric.Error(targets, candidatePredictions);
if (error < bestError)
{
bestError = error;
candidateModelIndices.CopyTo(bestModelIndices, 0);
Trace.WriteLine("Models selected: " + bestModelIndices.Length + ": " + error);
}
}
Trace.WriteLine("Selected model indices: " + string.Join(", ", bestModelIndices.ToArray()));
return(bestModelIndices);
}