/// <summary>
/// Initializes a new instance of the
/// <see cref="CategoricalEntailmentEnsembleClassifier" /> class
/// by exploiting the specified <paramref name="trainer"/>
/// to select the given
/// <paramref name="numberOfTrainedCategoricalEntailments"/>.
/// </summary>
/// <param name="trainer">
/// An object whose state contains the information needed to
/// train the classifier.
/// </param>
/// <param name="numberOfTrainedCategoricalEntailments">
/// The number of categorical entailments to be trained.
/// </param>
/// <returns>
/// A classifier whose ensemble contains the trained
/// categorical entailments.
/// </returns>
/// <remarks>
/// <para>
/// The <paramref name="trainer"/> can be equipped with a
/// nonempty initial ensemble of categorical entailments.
/// This method always trains a classifier by adding
/// to such ensemble further entailments, whose number
/// is specified by parameter
/// <paramref name="numberOfTrainedCategoricalEntailments"/>.
/// However, the method returns the partial classifier whose
/// ensemble contains the additional trained entailments only
/// (no initial entailments).
/// </para>
/// </remarks>
private static CategoricalEntailmentEnsembleClassifier Train(
CategoricalEntailmentEnsembleTrainer trainer,
int numberOfTrainedCategoricalEntailments)
{
var optimizer = new SystemPerformanceOptimizer();
int numberOfResponseCategories =
trainer.ResponseVariable.NumberOfCategories;
var context =
new CategoricalEntailmentEnsembleOptimizationContext(
objectiveFunction: trainer.Performance,
trainer.featureCategoryCounts,
numberOfResponseCategories,
numberOfTrainedCategoricalEntailments,
trainer.allowEntailmentPartialTruthValues,
probabilitySmoothingCoefficient: .9,
optimizationGoal: OptimizationGoal.Maximization,
minimumNumberOfIterations: 10,
maximumNumberOfIterations: 1000);
int numberOfParameters = numberOfTrainedCategoricalEntailments * (
trainer.featureCategoryCounts.Sum() + numberOfResponseCategories);
int sampleSize = 100 * numberOfParameters;
double rarity = .01;
var results = optimizer.Optimize(
context,
rarity,
sampleSize);
var partialClassifier = context.GetCategoricalEntailmentEnsembleClassifier(
results.OptimalState,
new List <CategoricalVariable>(trainer.FeatureVariables),
trainer.ResponseVariable);
return(partialClassifier);
}
public void GetOptimalStateTest()
{
// valid input - random ties resolution
{
var context = new CategoricalEntailmentEnsembleOptimizationContext(
objectiveFunction:
(DoubleMatrix state) => { return(Double.PositiveInfinity); },
featureCategoryCounts: new List <int>(1)
{
6
},
numberOfResponseCategories: 4,
numberOfCategoricalEntailments: 1,
allowEntailmentPartialTruthValues: true,
probabilitySmoothingCoefficient: .9,
optimizationGoal: OptimizationGoal.Maximization,
minimumNumberOfIterations: 5,
maximumNumberOfIterations: 1000);
int numberOfEvaluations = 10000;
double delta = .01;
var parameter = DoubleMatrix.Dense(1, 10, new double[10] {
.5, .5, .5, .5, .5, .5, .25, .25, .25, .25
});
// Generate states
var states = new int[numberOfEvaluations];
var responseIndexes = IndexCollection.Range(6, 9);
for (int i = 0; i < numberOfEvaluations; i++)
{
var state = context.GetOptimalState(parameter);
states[i] = state.Vec(responseIndexes).FindNonzero()[0];
}
// Compute the actual inclusion probabilities
DoubleMatrix actualInclusionProbabilities =
DoubleMatrix.Dense(context.NumberOfResponseCategories, 1);
var stateIndexes = IndexCollection.Default(numberOfEvaluations - 1);
for (int j = 0; j < context.NumberOfResponseCategories; j++)
{
var samplesContainingCurrentUnit =
IndexPartition.Create(
stateIndexes,
(i) => { return(states[i] == j); });
actualInclusionProbabilities[j] =
(double)samplesContainingCurrentUnit[true].Count
/
(double)numberOfEvaluations;
}
// Check the number of distinct generated states
var distinctStates =
IndexPartition.Create(
states);
int numberOfDistinctStates =
distinctStates.Count;
Assert.AreEqual(
expected: context.NumberOfResponseCategories,
actual: numberOfDistinctStates);
// Check that the Chebyshev Inequality holds true
// for each inclusion probability
var expectedInclusionProbabilities =
DoubleMatrix.Dense(context.NumberOfResponseCategories, 1,
1.0 / context.NumberOfResponseCategories);
for (int j = 0; j < context.NumberOfResponseCategories; j++)
{
ProbabilityDistributionTest.CheckChebyshevInequality(
new BernoulliDistribution(expectedInclusionProbabilities[j]),
actualInclusionProbabilities[j],
numberOfEvaluations,
delta);
}
// Check how good the actual inclusion probabilities fit
// the expected ones
// The following assumes a number of response
// categories equal to 4.
//
// The quantile of order .9 for
// the chi-squared distribution having 4-1
// degrees of freedom is 6.251389
// (as from R function qchisq(.9, 3))
var goodnessOfFitCriticalValue = 6.251389;
ProbabilityDistributionTest.CheckGoodnessOfFit(
expectedInclusionProbabilities,
actualInclusionProbabilities,
goodnessOfFitCriticalValue);
}
}
