/// <summary>
/// Computes the influence of all configuration options based on the measurements of the given result db. It uses linear programming (simplex) and is an exact algorithm.
/// </summary>
/// <param name="nfp">The non-funcitonal property for which the influences of configuration options are to be computed. If null, we use the property of the global model.</param>
/// <param name="infModel">The influence model containing options and interactions. The state of the model will be changed by the result of the process</param>
/// <param name="db">The result database containing the measurements.</param>
/// <returns>A map of binary options to their computed influences.</returns>
public Dictionary<BinaryOption, double> computeOptionInfluences(NFProperty nfp, InfluenceModel infModel, ResultDB db)
{
List<BinaryOption> variables = infModel.Vm.BinaryOptions;
List<double> results = new List<double>();
List<List<BinaryOption>> configurations = new List<List<BinaryOption>>();
foreach (Configuration c in db.Configurations)
{
configurations.Add(c.getBinaryOptions(BinaryOption.BinaryValue.Selected));
if (nfp != null)
results.Add(c.GetNFPValue(nfp));
else
results.Add(c.GetNFPValue());
}
List<String> errorEqs = new List<string>();
Dictionary<String, double> faultRates = new Dictionary<string, double>();
List<int> indexOfErrorMeasurements = new List<int>();
Dictionary<String, double> featureValuedAsStrings = solve(variables, results, configurations, infModel.InteractionInfluence.Keys.ToList());
foreach (String current in featureValuedAsStrings.Keys)
{
BinaryOption temp = infModel.Vm.getBinaryOption(current);
this.featureValues[temp] = featureValuedAsStrings[current];
InfluenceFunction influence = new InfluenceFunction(temp.Name + " + " + featureValuedAsStrings[current].ToString(),infModel.Vm);
if (infModel.BinaryOptionsInfluence.Keys.Contains(temp))
infModel.BinaryOptionsInfluence[temp] = influence;
else
infModel.BinaryOptionsInfluence.Add(temp, influence);
}
return this.featureValues;
}
public void clear()
{
this.nbBaggings = 0;
this.mLsettings = new ML_Settings();
this.metaModel = null;
this.models.Clear();
clearSampling();
}
/// <summary>
/// Clears the global state. This mehtod should be used after performing all experiments of one case study.
/// </summary>
public static void clear()
{
varModel = null;
currentNFP = null;
allMeasurements = new ResultDB();
evalutionSet = new ResultDB();
infModel = null;
nfProperties = new Dictionary<string,NFProperty>();
optionOrder = new List<ConfigurationOption>();
}
/// <summary>
/// This method searches for a corresponding methods in the dynamically loadeda assemblies and calls it if found. It prefers due to performance reasons the Microsoft Solver Foundation implementation.
/// </summary>
/// <param name="nfp">The non-funcitonal property for which the influences of configuration options are to be computed. If null, we use the property of the global model.</param>
/// <param name="infModel">The influence model containing options and interactions. The state of the model will be changed by the result of the process</param>
/// <param name="db">The result database containing the measurements.</param>
/// <returns>A map of binary options to their computed influences.</returns>
public Dictionary<BinaryOption, double> computeOptionInfluences(NFProperty nfp, InfluenceModel infModel, ResultDB db)
{
foreach (Lazy<ISolverLP, ISolverType> solver in solvers)
{
if (solver.Metadata.SolverType.Equals("MSSolverFoundation")) return solver.Value.computeOptionInfluences(nfp, infModel, db);
}
//If not MS Solver, take any solver. Should be changed when supporting more than 2 solvers here
foreach (Lazy<ISolverLP, ISolverType> solver in solvers)
{
return solver.Value.computeOptionInfluences(nfp, infModel, db);
}
return null;
}
/// <summary>
/// Performs the functionality of one command. If no functionality is found for the command, the command is retuned by this method.
/// </summary>
/// <param name="line">One command with its parameters.</param>
/// <returns>Returns an empty string if the command could be performed by the method. If the command could not be performed by the method, the original command is returned.</returns>
public string performOneCommand(string line)
{
GlobalState.logInfo.logLine(COMMAND + line);
// remove comment part of the line (the comment starts with an #)
line = line.Split(new Char[] { '#' }, 2)[0];
if (line.Length == 0)
return "";
// split line in command and parameters of the command
string[] components = line.Split(new Char[] { ' ' }, 2);
string command = components[0];
string task = "";
if (components.Length > 1)
task = components[1];
string[] taskAsParameter = task.Split(new Char[] { ' ' });
switch (command.ToLower())
{
case COMMAND_START_ALLMEASUREMENTS:
{
InfluenceModel infMod = new InfluenceModel(GlobalState.varModel, GlobalState.currentNFP);
List<Configuration> configurations_Learning = new List<Configuration>();
foreach (Configuration config in GlobalState.allMeasurements.Configurations)
{
if (config.nfpValues.ContainsKey(GlobalState.currentNFP))
configurations_Learning.Add(config);
}
if (configurations_Learning.Count == 0)
{
GlobalState.logInfo.logLine("The learning set is empty! Cannot start learning!");
break;
}
GlobalState.logInfo.logLine("Learning: " + "NumberOfConfigurationsLearning:" + configurations_Learning.Count);
// prepare the machine learning
exp = new MachineLearning.Learning.Regression.Learning(configurations_Learning, configurations_Learning);
exp.metaModel = infMod;
exp.mLsettings = this.mlSettings;
exp.learn();
}
break;
case COMMAND_TRUEMODEL:
StreamReader readModel = new StreamReader(task);
String model = readModel.ReadLine().Trim();
readModel.Close();
this.trueModel = new InfluenceFunction(model.Replace(',', '.'), GlobalState.varModel);
NFProperty artificalProp = new NFProperty("artificial");
GlobalState.currentNFP = artificalProp;
//computeEvaluationDataSetBasedOnTrueModel();
break;
case COMMAND_SUBSCRIPT:
{
FileInfo fi = new FileInfo(task);
StreamReader reader = null;
if (!fi.Exists)
throw new FileNotFoundException(@"Automation script not found. ", fi.ToString());
reader = fi.OpenText();
Commands co = new Commands();
co.exp = this.exp;
while (!reader.EndOfStream)
{
String oneLine = reader.ReadLine().Trim();
co.performOneCommand(oneLine);
}
}
break;
case COMMAND_EVALUATION_SET:
{
GlobalState.evalutionSet.Configurations = ConfigurationReader.readConfigurations(task, GlobalState.varModel);
GlobalState.logInfo.logLine("Evaluation set loaded.");
}
break;
case COMMAND_CLEAR_GLOBAL:
SPLConqueror_Core.GlobalState.clear();
toSample.Clear();
toSampleValidation.Clear();
break;
case COMMAND_CLEAR_SAMPLING:
exp.clearSampling();
toSample.Clear();
toSampleValidation.Clear();
break;
case COMMAND_CLEAR_LEARNING:
exp.clear();
toSample.Clear();
toSampleValidation.Clear();
break;
case COMMAND_LOAD_CONFIGURATIONS:
GlobalState.allMeasurements.Configurations = (GlobalState.allMeasurements.Configurations.Union(ConfigurationReader.readConfigurations(task, GlobalState.varModel))).ToList();
GlobalState.logInfo.logLine(GlobalState.allMeasurements.Configurations.Count + " configurations loaded.");
break;
case COMMAND_SAMPLE_ALLBINARY:
{
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
this.toSampleValidation.Add(SamplingStrategies.ALLBINARY);
this.exp.info.binarySamplings_Validation = "ALLBINARY";
}
else
{
this.toSample.Add(SamplingStrategies.ALLBINARY);
this.exp.info.binarySamplings_Learning = "ALLBINARY";
}
break;
}
case COMMAND_ANALYZE_LEARNING:
{//TODO: Analyzation is not supported in the case of bagging
GlobalState.logInfo.logLine("Models:");
if (this.mlSettings.bagging)
{
for (int i = 0; i < this.exp.models.Count; i++)
{
FeatureSubsetSelection learnedModel = exp.models[i];
if (learnedModel == null)
{
GlobalState.logError.logLine("Error... learning was not performed!");
break;
}
GlobalState.logInfo.logLine("Termination reason: " + learnedModel.LearningHistory.Last().terminationReason);
foreach (LearningRound lr in learnedModel.LearningHistory)
{
double relativeError = 0;
if (GlobalState.evalutionSet.Configurations.Count > 0)
{
double relativeErro2r = learnedModel.computeError(lr.FeatureSet, GlobalState.evalutionSet.Configurations, out relativeError);
}
else
{
double relativeErro2r = learnedModel.computeError(lr.FeatureSet, GlobalState.allMeasurements.Configurations, out relativeError);
}
GlobalState.logInfo.logLine(lr.ToString() + relativeError);
}
}
}
else
{
FeatureSubsetSelection learnedModel = exp.models[0];
if (learnedModel == null)
{
GlobalState.logError.logLine("Error... learning was not performed!");
break;
}
GlobalState.logInfo.logLine("Termination reason: " + learnedModel.LearningHistory.Last().terminationReason);
foreach (LearningRound lr in learnedModel.LearningHistory)
{
double relativeError = 0;
if (GlobalState.evalutionSet.Configurations.Count > 0)
{
double relativeErro2r = learnedModel.computeError(lr.FeatureSet, GlobalState.evalutionSet.Configurations, out relativeError);
}
else
{
double relativeErro2r = learnedModel.computeError(lr.FeatureSet, GlobalState.allMeasurements.Configurations, out relativeError);
}
GlobalState.logInfo.logLine(lr.ToString() + relativeError);
}
}
break;
}
case COMMAND_EXERIMENTALDESIGN:
performOneCommand_ExpDesign(task);
break;
case COMMAND_SAMPLING_OPTIONORDER:
parseOptionOrder(task);
break;
case COMMAND_VARIABILITYMODEL:
GlobalState.varModel = VariabilityModel.loadFromXML(task);
if (GlobalState.varModel == null)
GlobalState.logError.logLine("No variability model found at " + task);
break;
case COMMAND_SET_NFP:
GlobalState.currentNFP = GlobalState.getOrCreateProperty(task.Trim());
break;
case COMMAND_SAMPLE_OPTIONWISE:
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
this.toSampleValidation.Add(SamplingStrategies.OPTIONWISE);
this.exp.info.binarySamplings_Validation = "OPTIONSWISE";
}
else
{
this.toSample.Add(SamplingStrategies.OPTIONWISE);
this.exp.info.binarySamplings_Learning = "OPTIONSWISE";
}
break;
case COMMAND_LOG:
string location = task.Trim();
GlobalState.logInfo.close();
GlobalState.logInfo = new InfoLogger(location);
GlobalState.logError.close();
GlobalState.logError = new ErrorLogger(location + "_error");
break;
case COMMAND_SET_MLSETTING:
this.mlSettings = ML_Settings.readSettings(task);
break;
case COMMAND_LOAD_MLSETTINGS:
this.mlSettings = ML_Settings.readSettingsFromFile(task);
break;
case COMMAND_SAMPLE_PAIRWISE:
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
this.toSampleValidation.Add(SamplingStrategies.PAIRWISE);
this.exp.info.binarySamplings_Validation = "PAIRWISE";
}
else
{
this.toSample.Add(SamplingStrategies.PAIRWISE);
this.exp.info.binarySamplings_Learning = "PAIRWISE";
}
break;
case COMMAND_PRINT_MLSETTINGS:
GlobalState.logInfo.logLine(this.mlSettings.ToString());
break;
case COMMAND_PRINT_CONFIGURATIONS:
{
/* List<Dictionary<NumericOption, double>> numericSampling = exp.NumericSelection_Learning;
List<List<BinaryOption>> binarySampling = exp.BinarySelections_Learning;
List<Configuration> configurations = new List<Configuration>();
foreach (Dictionary<NumericOption, double> numeric in numericSampling)
{
foreach (List<BinaryOption> binary in binarySampling)
{
Configuration config = Configuration.getConfiguration(binary, numeric);
if (!configurations.Contains(config) && GlobalState.varModel.configurationIsValid(config))
{
configurations.Add(config);
}
}
}*/
var configs = ConfigurationBuilder.buildConfigs(GlobalState.varModel, this.toSample);
string[] para = task.Split(new char[] { ' ' });
// TODO very error prone..
ConfigurationPrinter printer = new ConfigurationPrinter(para[0], para[1], para[2], GlobalState.optionOrder);
printer.print(configs);
break;
}
case COMMAND_SAMPLE_BINARY_RANDOM:
{
string[] para = task.Split(new char[] { ' ' });
ConfigurationBuilder.binaryThreshold = Convert.ToInt32(para[0]);
ConfigurationBuilder.binaryModulu = Convert.ToInt32(para[1]);
VariantGenerator vg = new VariantGenerator(null);
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
this.toSampleValidation.Add(SamplingStrategies.BINARY_RANDOM);
this.exp.info.binarySamplings_Validation = "BINARY_RANDOM";
}
else
{
this.toSample.Add(SamplingStrategies.BINARY_RANDOM);
this.exp.info.binarySamplings_Learning = "BINARY_RANDOM " + task;
}
break;
}
case COMMAND_START_LEARNING:
{
InfluenceModel infMod = new InfluenceModel(GlobalState.varModel, GlobalState.currentNFP);
List<Configuration> configurationsLearning = buildSet(this.toSample);
List<Configuration> configurationsValidation = buildSet(this.toSampleValidation);
if (configurationsLearning.Count == 0)
{
configurationsLearning = configurationsValidation;
}
if (configurationsLearning.Count == 0)
{
GlobalState.logInfo.logLine("The learning set is empty! Cannot start learning!");
break;
}
if (configurationsValidation.Count == 0)
{
configurationsValidation = configurationsLearning;
}
GlobalState.logInfo.logLine("Learning: " + "NumberOfConfigurationsLearning:" + configurationsLearning.Count + " NumberOfConfigurationsValidation:" + configurationsValidation.Count);
//+ " UnionNumberOfConfigurations:" + (configurationsLearning.Union(configurationsValidation)).Count()); too costly to compute
// We have to reuse the list of models because of NotifyCollectionChangedEventHandlers that might be attached to the list of models.
exp.models.Clear();
var mod = exp.models;
exp = new MachineLearning.Learning.Regression.Learning(configurationsLearning, configurationsValidation);
exp.models = mod;
exp.metaModel = infMod;
exp.mLsettings = this.mlSettings;
exp.learn();
GlobalState.logInfo.logLine("Average model: \n" + exp.metaModel.printModelAsFunction());
double relativeError = 0;
if (GlobalState.evalutionSet.Configurations.Count > 0)
{
relativeError = FeatureSubsetSelection.computeError(exp.metaModel, GlobalState.evalutionSet.Configurations, ML_Settings.LossFunction.RELATIVE);
}
else
{
relativeError = FeatureSubsetSelection.computeError(exp.metaModel, GlobalState.allMeasurements.Configurations, ML_Settings.LossFunction.RELATIVE);
}
GlobalState.logInfo.logLine("Error :" + relativeError);
}
break;
case COMMAND_SAMPLE_NEGATIVE_OPTIONWISE:
// TODO there are two different variants in generating NegFW configurations.
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
this.toSampleValidation.Add(SamplingStrategies.NEGATIVE_OPTIONWISE);
this.exp.info.binarySamplings_Validation = "NEGATIVE_OPTIONWISE";
}
else
{
this.toSample.Add(SamplingStrategies.NEGATIVE_OPTIONWISE);
this.exp.info.binarySamplings_Learning = "NEGATIVE_OPTIONWISE";
}
break;
default:
return command;
}
return "";
}
/// <summary>
/// This method searches for a corresponding methods in the dynamically loadeda assemblies and calls it if found. It prefers due to performance reasons the Microsoft Solver Foundation implementation.
/// </summary>
/// <param name="nfp">The non-funcitonal property for which the influences of configuration options are to be computed. If null, we use the property of the global model.</param>
/// <param name="infModel">The influence model containing the variability model, all configuration options and interactions.</param>
/// <param name="db">The result database containing the measurements.</param>
/// <param name="evaluateFeatureInteractionsOnly">Only interactions are learned.</param>
/// <param name="withDeviation">(Not used) We can specifiy whether learned influences must be greater than a certain value (e.g., greater than measurement bias).</param>
/// <param name="deviation">(Not used) We can specifiy whether learned influences must be greater than a certain value (e.g., greater than measurement bias).</param>
/// <returns>Returns the learned infleunces of each option in a map whereas the String (Key) is the name of the option / interaction.</returns>
public Dictionary<string, double> computeOptionInfluences(NFProperty nfp, InfluenceModel infModel, ResultDB db, bool evaluateFeatureInteractionsOnly, bool withDeviation, double deviation)
{
foreach (Lazy<ISolverLP, ISolverType> solver in solvers)
{
if (solver.Metadata.SolverType.Equals("MSSolverFoundation")) return solver.Value.computeOptionInfluences(nfp, infModel, db, evaluateFeatureInteractionsOnly, withDeviation, deviation);
}
//If not MS Solver, take any solver. Should be changed when supporting more than 2 solvers here
foreach (Lazy<ISolverLP, ISolverType> solver in solvers)
{
return solver.Value.computeOptionInfluences(nfp, infModel, db, evaluateFeatureInteractionsOnly, withDeviation, deviation);
}
return null;
}
public void learn()
{
if (!hasNecessaryData())
return;
if (this.mLsettings.bagging)
{
//Get number of cores
int coreCount = 0;
foreach (var item in new System.Management.ManagementObjectSearcher("Select NumberOfCores from Win32_Processor").Get())
{
coreCount += int.Parse(item["NumberOfCores"].ToString());
}
createThreadPool(coreCount);
this.nbBaggings = this.mLsettings.baggingNumbers;
iCount = this.nbBaggings;
Random rand = new Random();
int nbOfConfigs = (testSet.Count * this.mLsettings.baggingTestDataFraction) / 100;
for (int i = 0; i < nbBaggings; i++)
{
InfluenceModel infMod = new InfluenceModel(GlobalState.varModel, GlobalState.currentNFP);
FeatureSubsetSelection sel = new FeatureSubsetSelection(infMod, this.mLsettings);
this.models.Add(sel);
List<int> selection = new List<int>();
for (int r = 0; r <= nbOfConfigs; r++)
{
selection.Add(rand.Next(nbOfConfigs));
}
List<Configuration> newTestSet = new List<Configuration>();
List<Configuration> newValidationSet = new List<Configuration>();
for (int r = 0; r <= selection.Count; r++)
{
if (selection.Contains(r))
newTestSet.Add(testSet[r]);
else
newValidationSet.Add(testSet[r]);
}
sel.setLearningSet(newTestSet);
sel.setValidationSet(newValidationSet);
Task task = EnqueueTask(() => sel.learn());
}
eventX.WaitOne(Timeout.Infinite, true);
averageModels();
}
else
{
InfluenceModel infMod = new InfluenceModel(GlobalState.varModel, GlobalState.currentNFP);
FeatureSubsetSelection sel = new FeatureSubsetSelection(infMod, this.mLsettings);
this.models.Add(sel);
sel.setLearningSet(testSet);
sel.setValidationSet(this.validationSet);
Stopwatch sw = new Stopwatch();
sw.Start();
sel.learn();
sw.Stop();
Console.WriteLine("Elapsed={0}", sw.Elapsed);
}
}
private void updateInfluenceModel(InfluenceModel influenceModel)
{
foreach (BinaryOption bin in influenceModel.BinaryOptionsInfluence.Keys)
{
if (this.metaModel.BinaryOptionsInfluence.Keys.Contains(bin))
{
((Feature)this.metaModel.BinaryOptionsInfluence[bin]).Constant += ((Feature)influenceModel.BinaryOptionsInfluence[bin]).Constant;
}
else
{
this.metaModel.BinaryOptionsInfluence.Add(bin, ((Feature)influenceModel.BinaryOptionsInfluence[bin]));
}
}
foreach (NumericOption num in influenceModel.NumericOptionsInfluence.Keys)
{
if (this.metaModel.NumericOptionsInfluence.Keys.Contains(num))
{
((Feature)this.metaModel.NumericOptionsInfluence[num]).Constant += ((Feature)influenceModel.NumericOptionsInfluence[num]).Constant;
}
else
{
this.metaModel.NumericOptionsInfluence.Add(num, ((Feature)influenceModel.NumericOptionsInfluence[num]));
}
}
foreach (Interaction interact in influenceModel.InteractionInfluence.Keys)
{
if (this.metaModel.InteractionInfluence.Keys.Contains(interact))
{
((Feature)this.metaModel.InteractionInfluence[interact]).Constant += ((Feature)influenceModel.InteractionInfluence[interact]).Constant;
}
else
{
this.metaModel.InteractionInfluence.Add(interact, ((Feature)influenceModel.InteractionInfluence[interact]));
}
}
}
/// <summary>
/// Computes the influence of all configuration options and interactions based on the measurements of the given result db. It uses linear programming (simplex) and is an exact algorithm.
/// </summary>
/// <param name="nfp">The non-funcitonal property for which the influences of configuration options are to be computed. If null, we use the property of the global model.</param>
/// <param name="infModel">The influence model containing the variability model, all configuration options and interactions.</param>
/// <param name="db">The result database containing the measurements.</param>
/// <param name="evaluateFeatureInteractionsOnly">Only interactions are learned.</param>
/// <param name="withDeviation">(Not used) We can specifiy whether learned influences must be greater than a certain value (e.g., greater than measurement bias).</param>
/// <param name="deviation">(Not used) We can specifiy whether learned influences must be greater than a certain value (e.g., greater than measurement bias).</param>
/// <returns>Returns the learned infleunces of each option in a map whereas the String (Key) is the name of the option / interaction.</returns>
public Dictionary<String, double> computeOptionInfluences(NFProperty nfp, InfluenceModel infModel, ResultDB db, bool evaluateFeatureInteractionsOnly, bool withDeviation, double deviation)
{
//Initialization
List<List<BinaryOption>> configurations = new List<List<BinaryOption>>();
this.evaluateInteractionsOnly = evaluateFeatureInteractionsOnly;
this.withStandardDeviation = withDeviation;
this.standardDeviation = deviation;
List<double> results = new List<double>();
foreach (Configuration c in db.Configurations)
{
configurations.Add(c.getBinaryOptions(BinaryOption.BinaryValue.Selected));
if (nfp != null)
results.Add(c.GetNFPValue(nfp));
else
results.Add(c.GetNFPValue());
}
List<BinaryOption> variables = new List<BinaryOption>();
Dictionary<String, double> featureValues = new Dictionary<string, double>();
Dictionary<String, double> faultRates = new Dictionary<string, double>();
List<int> indexOfErrorMeasurements = new List<int>();
if (configurations.Count == 0)
return null;
//For the case there is an empty base
if (configurations.Count != 0)
{
if (configurations[0].Count == 0)
{//Should never occur that we get a configuration with no option selected... at least the root must be there
BinaryOption root = infModel.Vm.Root;
//Element baseElement = new Element("base_gen", infModel.getID(), infModel);
//variables.Add(baseElement);
// featureValues.Add(baseElement.getName(), 0);
foreach (List<BinaryOption> config in configurations)
if(!config.Contains(root))
config.Insert(0, root);
}
}
//Building the variable list
foreach (var elem in infModel.Vm.BinaryOptions)
{
variables.Add(elem);
featureValues.Add(elem.Name, 0);
}
//First run
featureValues = solve(variables, results, configurations, null);
//if (evaluateFeatureInteractionsOnly == false)
return featureValues;
/*
//We might have some interactions here and cannot compute all values
//1. identify options that are only present in these equations
Dictionary<Element, int> featureCounter = new Dictionary<Element, int>();
for (int i = 0; i < indexOfErrorMeasurements.Count; i++)
{
}
*/
/*Todo: get compute interactins from deviations / errors of the LP results
if (errorEqs != null)
{
foreach (string eq in errorEqs)
{
double value = Double.Parse(eq.Substring(eq.IndexOf("==") + 2));
StringBuilder sb = new StringBuilder();
List<Element> derivativeParents = new List<Element>();
sb.Append("derivate_");
string[] splittedEQ = eq.Split('+');
foreach (string element in splittedEQ)
{
string name = element;
if (name.Contains("=="))
name = name.Substring(0, name.IndexOf("=="));
if (name.Contains("yp") && name.Contains("-yn"))
continue;
// string featureName = name.Substring(0, name.IndexOf("_p-"));
Element elem = infModel.getElementByNameUnsafe(name);
if (elem == null)
continue;
sb.Append("_" + name);
derivativeParents.Add(elem);
}
Element interaction = new Element(sb.ToString(), infModel.getID(), infModel);
interaction.setType("derivative");
interaction.addDerivativeParents(derivativeParents);
infModel.addElement(interaction);
this.featureValues.Add(interaction, value);
}
}
return featureValues;*/
}
/// <summary>
/// Performs the functionality of one command. If no functionality is found for the command, the command is retuned by this method.
/// </summary>
/// <param name="line">One command with its parameters.</param>
/// <returns>Returns an empty string if the command could be performed by the method. If the command could not be performed by the method, the original command is returned.</returns>
public string performOneCommand(string line)
{
GlobalState.logInfo.log(COMMAND + line);
// remove comment part of the line (the comment starts with an #)
line = line.Split(new Char[] { '#' }, 2)[0];
if (line.Length == 0)
return "";
// split line in command and parameters of the command
string[] components = line.Split(new Char[] { ' ' }, 2);
string command = components[0];
string task = "";
if (components.Length > 1)
task = components[1];
string[] taskAsParameter = task.Split(new Char[] { ' ' });
switch (command.ToLower())
{
case COMMAND_TRUEMODEL:
StreamReader readModel = new StreamReader(task);
String model = readModel.ReadLine().Trim();
readModel.Close();
exp.TrueModel = new InfluenceFunction(model.Replace(',','.'), GlobalState.varModel);
NFProperty artificalProp = new NFProperty("artificial");
GlobalState.currentNFP = artificalProp;
computeEvaluationDataSetBasedOnTrueModel();
break;
case COMMAND_SUBSCRIPT:
{
FileInfo fi = new FileInfo(task);
StreamReader reader = null;
if (!fi.Exists)
throw new FileNotFoundException(@"Automation script not found. ", fi.ToString());
reader = fi.OpenText();
Commands co = new Commands();
co.exp = this.exp;
while (!reader.EndOfStream)
{
String oneLine = reader.ReadLine().Trim();
co.performOneCommand(oneLine);
}
}
break;
case COMMAND_EVALUATION_SET:
{
GlobalState.evalutionSet.Configurations = ConfigurationReader.readConfigurations(task, GlobalState.varModel);
GlobalState.logInfo.log("Evaluation set loaded.");
}
break;
case COMMAND_CLEAR_GLOBAL:
SPLConqueror_Core.GlobalState.clear();
break;
case COMMAND_CLEAR_SAMPLING:
exp.clearSampling();
break;
case COMMAND_CLEAR_LEARNING:
exp.clear();
break;
case COMMAND_LOAD_CONFIGURATIONS:
GlobalState.allMeasurements.Configurations = (GlobalState.allMeasurements.Configurations.Union(ConfigurationReader.readConfigurations(task, GlobalState.varModel))).ToList();
GlobalState.logInfo.log(GlobalState.allMeasurements.Configurations.Count + " configurations loaded.");
break;
case COMMAND_SAMPLE_ALLBINARY:
{
VariantGenerator vg = new VariantGenerator(null);
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
exp.addBinarySelection_Validation(vg.generateAllVariantsFast(GlobalState.varModel));
exp.addBinarySampling_Validation(COMMAND_SAMPLE_ALLBINARY);
}
else
{
exp.addBinarySelection_Learning(vg.generateAllVariantsFast(GlobalState.varModel));
exp.addBinarySampling_Learning(COMMAND_SAMPLE_ALLBINARY);
}
break;
}
case COMMAND_ANALYZE_LEARNING:
{
GlobalState.logInfo.log("Models:");
FeatureSubsetSelection learning = exp.learning;
if (learning == null)
{
GlobalState.logError.log("Error... learning was not performed!");
break;
}
foreach (LearningRound lr in learning.LearningHistory)
{
double relativeError = 0;
if (GlobalState.evalutionSet.Configurations.Count > 0)
{
double relativeErro2r = exp.learning.computeError(lr.FeatureSet, GlobalState.evalutionSet.Configurations, out relativeError);
}
else
{
double relativeErro2r = exp.learning.computeError(lr.FeatureSet, GlobalState.allMeasurements.Configurations, out relativeError);
}
GlobalState.logInfo.log(lr.ToString() + relativeError);
}
break;
}
case COMMAND_EXERIMENTALDESIGN:
performOneCommand_ExpDesign(task);
break;
case COMMAND_SAMPLING_OPTIONORDER:
parseOptionOrder(task);
break;
case COMMAND_VARIABILITYMODEL:
GlobalState.varModel = VariabilityModel.loadFromXML(task);
if (GlobalState.varModel == null)
GlobalState.logError.log("No variability model found at " + task);
break;
case COMMAND_SET_NFP:
GlobalState.currentNFP = GlobalState.getOrCreateProperty(task.Trim());
break;
case COMMAND_SAMPLE_OPTIONWISE:
FeatureWise fw = new FeatureWise();
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
exp.addBinarySelection_Validation(fw.generateFeatureWiseConfigsCSP(GlobalState.varModel));
exp.addBinarySampling_Validation("FW");
}
else
{
//exp.addBinarySelection_Learning(fw.generateFeatureWiseConfigsCSP(GlobalState.varModel));
exp.addBinarySelection_Learning(fw.generateFeatureWiseConfigurations(GlobalState.varModel));
exp.addBinarySampling_Learning("FW");
}
break;
case COMMAND_LOG:
string location = task.Trim();
GlobalState.logInfo.close();
GlobalState.logInfo = new InfoLogger(location);
GlobalState.logError.close();
GlobalState.logError = new ErrorLogger(location + "_error");
break;
case COMMAND_SET_MLSETTING:
exp.mlSettings = ML_Settings.readSettings(task);
break;
case COMMAND_LOAD_MLSETTINGS:
exp.mlSettings = ML_Settings.readSettingsFromFile(task);
break;
case COMMAND_SAMPLE_PAIRWISE:
PairWise pw = new PairWise();
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
exp.addBinarySelection_Validation(pw.generatePairWiseVariants(GlobalState.varModel));
exp.addBinarySampling_Validation("PW");
}
else
{
exp.addBinarySelection_Learning(pw.generatePairWiseVariants(GlobalState.varModel));
exp.addBinarySampling_Learning("PW");
}
break;
case COMMAND_PRINT_MLSETTINGS:
GlobalState.logInfo.log(exp.mlSettings.ToString());
break;
case COMMAND_PRINT_CONFIGURATIONS:
{
List<Dictionary<NumericOption, double>> numericSampling = exp.NumericSelection_Learning;
List<List<BinaryOption>> binarySampling = exp.BinarySelections_Learning;
List<Configuration> configurations = new List<Configuration>();
foreach (Dictionary<NumericOption, double> numeric in numericSampling)
{
foreach (List<BinaryOption> binary in binarySampling)
{
Configuration config = Configuration.getConfiguration(binary, numeric);
if (!configurations.Contains(config) && GlobalState.varModel.configurationIsValid(config))
{
configurations.Add(config);
}
}
}
string[] para = task.Split(new char[] { ' ' });
// TODO very error prune..
ConfigurationPrinter printer = new ConfigurationPrinter(para[0], para[1], para[2], GlobalState.optionOrder);
printer.print(configurations);
break;
}
case COMMAND_SAMPLE_BINARY_RANDOM:
{
string[] para = task.Split(new char[] { ' ' });
int treshold = Convert.ToInt32(para[0]);
int modulu = Convert.ToInt32(para[1]);
VariantGenerator vg = new VariantGenerator(null);
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
exp.addBinarySelection_Validation(vg.generateRandomVariants(GlobalState.varModel, treshold, modulu));
exp.addBinarySampling_Validation("random " + task);
}
else
{
exp.addBinarySelection_Learning(vg.generateRandomVariants(GlobalState.varModel, treshold, modulu));
exp.addBinarySampling_Learning("random " + task);
}
break;
}
case COMMAND_START_LEARNING:
{
InfluenceModel infMod = new InfluenceModel(GlobalState.varModel, GlobalState.currentNFP);
List<Configuration> configurations_Learning = new List<Configuration>();
List<Configuration> configurations_Validation = new List<Configuration>();
if (exp.TrueModel == null)
{
//List<List<BinaryOption>> availableBinary
//configurations_Learning = GlobalState.getMeasuredConfigs(exp.BinarySelections_Learning, exp.NumericSelection_Learning);
configurations_Learning = GlobalState.getMeasuredConfigs(Configuration.getConfigurations(exp.BinarySelections_Learning, exp.NumericSelection_Learning));
configurations_Learning = configurations_Learning.Distinct().ToList();
configurations_Validation = GlobalState.getMeasuredConfigs(Configuration.getConfigurations(exp.BinarySelections_Validation, exp.NumericSelection_Validation));
configurations_Validation = configurations_Validation.Distinct().ToList();
//break;//todo only to get the configurations that we haven't measured
} else
{
foreach (List<BinaryOption> binConfig in exp.BinarySelections_Learning)
{
if (exp.NumericSelection_Learning.Count == 0)
{
Configuration c = new Configuration(binConfig);
c.setMeasuredValue(GlobalState.currentNFP, exp.TrueModel.eval(c));
if (!configurations_Learning.Contains(c))
configurations_Learning.Add(c);
continue;
}
foreach (Dictionary<NumericOption, double> numConf in exp.NumericSelection_Learning)
{
Configuration c = new Configuration(binConfig, numConf);
c.setMeasuredValue(GlobalState.currentNFP, exp.TrueModel.eval(c));
if(GlobalState.varModel.configurationIsValid(c))
// if (!configurations_Learning.Contains(c))
configurations_Learning.Add(c);
}
}
}
if (configurations_Learning.Count == 0)
{
configurations_Learning = configurations_Validation;
}
if (configurations_Learning.Count == 0)
{
GlobalState.logInfo.log("The learning set is empty! Cannot start learning!");
break;
}
if (configurations_Validation.Count == 0)
{
configurations_Validation = configurations_Learning;
}
//break;
GlobalState.logInfo.log("Learning: " + "NumberOfConfigurationsLearning:" + configurations_Learning.Count + " NumberOfConfigurationsValidation:" + configurations_Validation.Count
+ " UnionNumberOfConfigurations:" + (configurations_Learning.Union(configurations_Validation)).Count());
// prepare the machine learning
exp.learning.init(infMod, exp.mlSettings);
exp.learning.setLearningSet(configurations_Learning);
exp.learning.setValidationSet(configurations_Validation);
exp.learning.learn();
}
break;
case COMMAND_SAMPLE_NEGATIVE_OPTIONWISE:
// TODO there are two different variants in generating NegFW configurations.
NegFeatureWise neg = new NegFeatureWise();
if (taskAsParameter.Contains(COMMAND_VALIDATION))
{
exp.addBinarySelection_Validation(neg.generateNegativeFW(GlobalState.varModel));
exp.addBinarySampling_Validation("newFW");
}
else
{
exp.addBinarySelection_Learning(neg.generateNegativeFW(GlobalState.varModel));//neg.generateNegativeFWAllCombinations(GlobalState.varModel));
exp.addBinarySampling_Learning("newFW");
}
break;
default:
return command;
}
return "";
}
