/// <param name="dataSet"> training set used for error estimation </param>
/// <returns> neural network model with optimized architecture for provided data set </returns>
public virtual NeuralNetwork createOptimalModel(DataSet dataSet)
{
List <int> neurons = new List <int>();
neurons.Add(minNeuronsPerLayer);
findArchitectures(1, minNeuronsPerLayer, neurons);
LOG.info("Total [{}] different network topologies found", allArchitectures.Count);
foreach (List <int> architecture in allArchitectures)
{
architecture.Insert(0, dataSet.InputSize);
architecture.Add(dataSet.OutputSize);
LOG.info("Architecture: [{}]", architecture);
MultiLayerPerceptron network = new MultiLayerPerceptron(architecture);
LearningListener listener = new LearningListener(10, learningRule.MaxIterations);
learningRule.addListener(listener);
network.LearningRule = learningRule;
errorEstimationMethod = new CrossValidation(network, dataSet, 10);
errorEstimationMethod.run();
// FIX
var evaluator = errorEstimationMethod.getEvaluator <ClassifierEvaluator.MultiClass>(typeof(ClassifierEvaluator.MultiClass));
ClassificationMetrics[] result = ClassificationMetrics.createFromMatrix(evaluator.Result);
// nadji onaj sa najmanjim f measure
if (optimalResult == null || optimalResult.FMeasure < result[0].FMeasure)
{
LOG.info("Architecture [{}] became optimal architecture with metrics {}", architecture, result);
optimalResult = result[0];
optimalClassifier = network;
optimalArchitecure = architecture;
}
LOG.info("#################################################################");
}
LOG.info("Optimal Architecture: {}", optimalArchitecure);
return(optimalClassifier);
}
