// step 1: extraction of data from traces
public static MLDataset FromExecutionTraces(Individual /* unused */ individual, IList <ExecutionTrace> traces)
{
MLDataset dataset = new MLDataset();
int count = traces.Count;
dataset.Count = count;
for (int i = 0; i < count; i++)
{
var trace = traces[i];
IList <ulong> featureOrder = new List <ulong>(trace.States.Count);
foreach (var record in trace.States)
{
object[] featureValues;
if (!dataset._features.TryGetValue(record.NodeId, out featureValues))
{
featureValues = new object[count];
dataset._features.Add(record.NodeId, featureValues);
dataset._featureInfo.Add(record.NodeId,
new MLDatasetFeature(record.NodeId, record.OperatorId, record.Value.GetType()));
}
featureValues[i] = record.Value;
if (record.OperatorId > 0) // only use operators (no variables, constants), because constant values get dropped anyway
{
featureOrder.Add(record.NodeId);
}
}
dataset._featureOrder.Add(featureOrder);
}
return(dataset);
}
private int[] GetExpectedOutputDataset(TestSuite testSuite)
{
int[] dataset = new int[testSuite.TestCases.Count];
for (int i = 0; i < dataset.Length; i++)
{
dataset[i] = MLDataset.ToDatasetValue(testSuite.TestCases[i].Result).GetValueOrDefault();
}
return(dataset);
}
private void LogDatasetFeatures(MLDataset dataset)
{
if (dataset != null)
{
Logger.Write(4, "Features: ");
foreach (var feature in dataset.Features)
{
Logger.Write(4, feature.ToString() + ",");
}
Logger.WriteLine(4, "");
}
}
public MDLFitnessResult(double fitness, MLDataset dataset = null)
: base(fitness)
{
Dataset = dataset;
}
public FitnessResult CalculateFitness(Individual individual, TestSuite testSuite, object[] results)
{
double fitness = StandardFitnessCalculator.CalculateFitness(individual, testSuite, results).Fitness; // standard fitness f0
MLDataset dataset = MLDataset.FromExecutionTraces(individual, Singleton <ExecutionRecord> .Instance.Traces);
LogDatasetFeatures(dataset);
var fitnessResult = new MDLFitnessResult(fitness, dataset);
if (dataset.Features.Count() > 0)
{
// Variant 1 (results of all nodes):
//var input = dataset.ToRawInputDataset();
// Variant 2 (results and operation types of first n and last n operations in chronological order):
//var input = dataset.ToRawFirstNLastNInputDataset(5, 5);
// Veriant 3 (results only consider values of bool and int expressions and operations)
int startContinuousFeatures;
var input = MLDataset.ConvertTracesToTypedSteps(individual, Singleton <ExecutionRecord> .Instance.Traces,
Steps, true, out startContinuousFeatures);
if (input.Length > 0 && input[0].Length > 0)
{
var expected = GetExpectedOutputDataset(testSuite);
LogDataset(input, expected);
var decisionTree = CreateDecisionTree(input, expected, startContinuousFeatures);
if (decisionTree != null)
{
int error = GetClassificationError(decisionTree, input, expected);
int treeSize = GetTreeSize(decisionTree);
double mdlFitness = CalculateMDLFitness(error, treeSize, results.Length);
double stdFitness = fitness;
fitness *= mdlFitness;
var rules = decisionTree.ToRules();
LogResult(fitness, error, treeSize, mdlFitness, rules);
int usedAttributes = GetUsedAttributes(decisionTree).Count;
fitnessResult.Fitness = fitness;
fitnessResult.ClassificationError = error;
fitnessResult.TreeSize = treeSize;
fitnessResult.StandardFitness = stdFitness;
fitnessResult.UsedAttributes = usedAttributes;
// log MDL result details
if (fitness == 0)
{
Logger.WriteLine(1, "Std fitness: " + stdFitness);
var predicted = decisionTree.Decide(input);
var loss = Math.Round(new ZeroOneLoss(expected).Loss(predicted) * expected.Length);
var classificationErrorFactor = (((double)error + 1) / (results.Length + 1));
Logger.WriteLine(1, "Program error: " + error + "; loss: " + loss + "; factor: " + classificationErrorFactor);
var treeSizeFactor = Math.Log(treeSize + 1, 2);
Logger.WriteLine(1, "Tree size: " + treeSize + "; factor: " + treeSizeFactor);
Logger.WriteLine(1, "MDL fitness: " + mdlFitness);
}
}
}
}
else if (fitness == 0)
{
Logger.WriteLine(1, "Standard fitness: 0");
}
return(fitnessResult);
}
