Changeset 2977

Timestamp:

03/09/10 09:40:52 (14 years ago)

Author:

gkronber

Message:

Moved linear scaling functionality out of tree evaluator into a separate operator. #823 (Implement tree evaluator with linear scaling to improve convergence in symbolic regression.)

Location:

trunk/sources

Files:

• HeuristicLab.GP.StructureIdentification/3.3/DefaultStructureIdentificationOperators.cs (modified) (8 diffs)
• HeuristicLab.GP.StructureIdentification/3.3/HeuristicLab.GP.StructureIdentification-3.3.csproj (modified) (1 diff)
• HeuristicLab.GP.StructureIdentification/3.3/LinearScalingPredictorBuilder.cs (modified) (3 diffs)
• HeuristicLab.GP.StructureIdentification/3.3/OffspringSelectionGPRegression.cs (modified) (1 diff)
• HeuristicLab.GP.StructureIdentification/3.3/ScalingTreeEvaluator.cs (deleted)
• HeuristicLab.GP.StructureIdentification/3.3/ScalingTreeEvaluatorInjector.cs (deleted)
• HeuristicLab.GP.StructureIdentification/3.3/StandardGPRegression.cs (modified) (4 diffs)
• HeuristicLab.Modeling/3.2/HeuristicLab.Modeling-3.2.csproj (modified) (1 diff)
• HeuristicLab.Modeling/3.2/LinearScaler.cs (added)
• HeuristicLab.Modeling/3.2/SimpleEvaluatorBase.cs (modified) (1 diff)

trunk/sources/HeuristicLab.GP.StructureIdentification/3.3/DefaultStructureIdentificationOperators.cs

@@ -39 +39 @@
       SequentialProcessor seq = new SequentialProcessor();
       seq.AddSubOperator(new DefaultFunctionLibraryInjector());
-      seq.AddSubOperator(new ScalingTreeEvaluatorInjector());
+      seq.AddSubOperator(new HL3TreeEvaluatorInjector());
       op.OperatorGraph.AddOperator(seq);
       op.OperatorGraph.InitialOperator = seq;
@@ -46 +46 @@
 
     public static IOperator CreateInitialPopulationEvaluator() {
-      MeanSquaredErrorEvaluator eval = new MeanSquaredErrorEvaluator();
-      eval.Name = "Evaluator";
-      eval.GetVariableInfo("MSE").ActualName = "Quality";
+      CombinedOperator combinedOp = new CombinedOperator();
+      combinedOp.Name = "Evaluator";
+      SequentialProcessor seq = new SequentialProcessor();
+
+      SimpleEvaluator eval = new SimpleEvaluator();
+      eval.GetVariableInfo("Values").ActualName = "TrainingValues";
       eval.GetVariableInfo("SamplesStart").ActualName = "ActualTrainingSamplesStart";
       eval.GetVariableInfo("SamplesEnd").ActualName = "ActualTrainingSamplesEnd";
-      return eval;
+
+      LinearScaler scaler = new LinearScaler();
+      scaler.GetVariableInfo("Values").ActualName = "TrainingValues";
+
+      SimpleMSEEvaluator mseEval = new SimpleMSEEvaluator();
+      mseEval.GetVariableInfo("Values").ActualName = "TrainingValues";
+      mseEval.GetVariableInfo("MSE").ActualName = "Quality";
+
+      seq.AddSubOperator(eval);
+      seq.AddSubOperator(scaler);
+      seq.AddSubOperator(mseEval);
+      combinedOp.OperatorGraph.AddOperator(seq);
+      combinedOp.OperatorGraph.InitialOperator = seq;
+      return combinedOp;
     }
 
@@ -70 +86 @@
       UniformSequentialSubScopesProcessor subScopesProc = new UniformSequentialSubScopesProcessor();
       SequentialProcessor individualProc = new SequentialProcessor();
-      MeanSquaredErrorEvaluator validationEvaluator = new MeanSquaredErrorEvaluator();
+      SimpleEvaluator validationEvaluator = new SimpleEvaluator();
       validationEvaluator.Name = "ValidationEvaluator";
-      validationEvaluator.GetVariableInfo("MSE").ActualName = "ValidationQuality";
+      validationEvaluator.GetVariableInfo("Values").ActualName = "ValidationValues";
       validationEvaluator.GetVariableInfo("SamplesStart").ActualName = "ValidationSamplesStart";
       validationEvaluator.GetVariableInfo("SamplesEnd").ActualName = "ValidationSamplesEnd";
-
+      LinearScaler validationScaler = new LinearScaler();
+      validationScaler.GetVariableInfo("Values").ActualName = "ValidationValues";
+      SimpleMSEEvaluator mseEval = new SimpleMSEEvaluator();
+      mseEval.GetVariableInfo("Values").ActualName = "ValidationValues";
+      mseEval.GetVariableInfo("MSE").ActualName = "ValidationQuality";
       individualProc.AddSubOperator(validationEvaluator);
+      individualProc.AddSubOperator(validationScaler);
+      individualProc.AddSubOperator(mseEval);
 
       Counter bestValidationSolutionAgeCounter = new Counter();
@@ -127 +149 @@
       seqSubScopeProc.AddSubOperator(solutionProc);
 
-      ScalingTreeEvaluatorInjector evaluatorInjector = new ScalingTreeEvaluatorInjector();
-      evaluatorInjector.AddVariable(new HeuristicLab.Core.Variable("PunishmentFactor", new DoubleData(1000.0)));
-      evaluatorInjector.GetVariableInfo("TreeEvaluator").ActualName = "ModelAnalysisTreeEvaluator";
-
       #region simple evaluators
       SimpleEvaluator trainingEvaluator = new SimpleEvaluator();
@@ -137 +155 @@
       trainingEvaluator.GetVariableInfo("SamplesEnd").ActualName = "TrainingSamplesEnd";
       trainingEvaluator.GetVariableInfo("Values").ActualName = "TrainingValues";
-      trainingEvaluator.GetVariableInfo("TreeEvaluator").ActualName = "ModelAnalysisTreeEvaluator";
       SimpleEvaluator validationEvaluator = new SimpleEvaluator();
       validationEvaluator.Name = "ValidationEvaluator";
@@ -143 +160 @@
       validationEvaluator.GetVariableInfo("SamplesEnd").ActualName = "ValidationSamplesEnd";
       validationEvaluator.GetVariableInfo("Values").ActualName = "ValidationValues";
-      validationEvaluator.GetVariableInfo("TreeEvaluator").ActualName = "ModelAnalysisTreeEvaluator";
       SimpleEvaluator testEvaluator = new SimpleEvaluator();
       testEvaluator.Name = "TestEvaluator";
@@ -149 +165 @@
       testEvaluator.GetVariableInfo("SamplesEnd").ActualName = "TestSamplesEnd";
       testEvaluator.GetVariableInfo("Values").ActualName = "TestValues";
-      testEvaluator.GetVariableInfo("TreeEvaluator").ActualName = "ModelAnalysisTreeEvaluator";
-      solutionProc.AddSubOperator(evaluatorInjector);
+
+      LinearScaler trainingScaler = new LinearScaler();
+      trainingScaler.GetVariableInfo("Values").ActualName = "TrainingValues";
+
+      LinearScaler validationScaler = new LinearScaler();
+      validationScaler.GetVariableInfo("Values").ActualName = "ValidationValues";
+
+      LinearScaler testScaler = new LinearScaler();
+      testEvaluator.GetVariableInfo("Values").ActualName = "TestValues";
+
       solutionProc.AddSubOperator(trainingEvaluator);
       solutionProc.AddSubOperator(validationEvaluator);
       solutionProc.AddSubOperator(testEvaluator);
+      solutionProc.AddSubOperator(trainingScaler);
+      solutionProc.AddSubOperator(validationScaler);
+      solutionProc.AddSubOperator(testScaler);
       #endregion
 
@@ -171 +198 @@
 
       NodeBasedVariableImpactCalculator nodeImpactCalculator = new NodeBasedVariableImpactCalculator();
-      nodeImpactCalculator.GetVariableInfo("TreeEvaluator").ActualName = "ModelAnalysisTreeEvaluator";
       nodeImpactCalculator.GetVariableInfo("SamplesStart").ActualName = "ValidationSamplesStart";
       nodeImpactCalculator.GetVariableInfo("SamplesEnd").ActualName = "ValidationSamplesEnd";

trunk/sources/HeuristicLab.GP.StructureIdentification/3.3/HeuristicLab.GP.StructureIdentification-3.3.csproj

@@ -88 +88 @@
     <Compile Include="LinearScalingPredictorBuilder.cs" />
     <Compile Include="Evaluators\VariableFrequencyBasedImpactCalculator.cs" />
-    <Compile Include="ScalingTreeEvaluator.cs" />
-    <Compile Include="ScalingTreeEvaluatorInjector.cs" />
     <Compile Include="Evaluators\TreeComplexityEvaluator.cs" />
     <Compile Include="FunctionLibraryInjectors\ArithmeticFunctionLibraryInjector.cs" />

trunk/sources/HeuristicLab.GP.StructureIdentification/3.3/LinearScalingPredictorBuilder.cs

@@ -35 +35 @@
       : base() {
       AddVariableInfo(new VariableInfo("FunctionTree", "The function tree", typeof(IGeneticProgrammingModel), VariableKind.In));
-      AddVariableInfo(new VariableInfo("PunishmentFactor", "The punishment factor limits the estimated values to a certain range", typeof(DoubleData), VariableKind.In));
-      AddVariableInfo(new VariableInfo("Dataset", "The dataset", typeof(Dataset), VariableKind.In));
-      AddVariableInfo(new VariableInfo("TrainingSamplesStart", "Start index of training set", typeof(DoubleData), VariableKind.In));
-      AddVariableInfo(new VariableInfo("TrainingSamplesEnd", "End index of training set", typeof(DoubleData), VariableKind.In));
-      AddVariableInfo(new VariableInfo("TargetVariable", "Name of the target variable", typeof(StringData), VariableKind.In));
+      AddVariableInfo(new VariableInfo("Beta", "Beta parameter for linear scaling as calculated by LinearScaler", typeof(DoubleData), VariableKind.In));
+      AddVariableInfo(new VariableInfo("Alpha", "Alpha parameter for linear scaling as calculated by LinearScaler", typeof(DoubleData), VariableKind.In));
+      AddVariableInfo(new VariableInfo("UpperEstimationLimit", "Upper limit for estimated value (optional)", typeof(DoubleData), VariableKind.In));
+      AddVariableInfo(new VariableInfo("LowerEstimationLimit", "Lower limit for estimated value (optional)", typeof(DoubleData), VariableKind.In));
       AddVariableInfo(new VariableInfo("Predictor", "The predictor combines the function tree and the evaluator and can be used to generate estimated values", typeof(IPredictor), VariableKind.New));
     }
@@ -49 +48 @@
     public override IOperation Apply(IScope scope) {
       IGeneticProgrammingModel model = GetVariableValue<IGeneticProgrammingModel>("FunctionTree", scope, true);
-      double punishmentFactor = GetVariableValue<DoubleData>("PunishmentFactor", scope, true).Data;
-      Dataset dataset = GetVariableValue<Dataset>("Dataset", scope, true);
-      int start = GetVariableValue<IntData>("TrainingSamplesStart", scope, true).Data;
-      int end = GetVariableValue<IntData>("TrainingSamplesEnd", scope, true).Data;
-      string targetVariable = GetVariableValue<StringData>("TargetVariable", scope, true).Data;
-      IPredictor predictor = CreatePredictor(model, punishmentFactor, dataset, targetVariable, start, end);
+      //double punishmentFactor = GetVariableValue<DoubleData>("PunishmentFactor", scope, true).Data;
+      //Dataset dataset = GetVariableValue<Dataset>("Dataset", scope, true);
+      //int start = GetVariableValue<IntData>("TrainingSamplesStart", scope, true).Data;
+      //int end = GetVariableValue<IntData>("TrainingSamplesEnd", scope, true).Data;
+      //string targetVariable = GetVariableValue<StringData>("TargetVariable", scope, true).Data;
+      double alpha = GetVariableValue<DoubleData>("Alpha", scope, true).Data;
+      double beta = GetVariableValue<DoubleData>("Beta", scope, true).Data;
+      DoubleData lowerLimit = GetVariableValue<DoubleData>("LowerEstimationLimit", scope, true, false);
+      DoubleData upperLimit = GetVariableValue<DoubleData>("UpperEstimationLimit", scope, true, false);
+      IPredictor predictor;
+      if (lowerLimit == null || upperLimit == null)
+        predictor = CreatePredictor(model, beta, alpha, double.NegativeInfinity, double.PositiveInfinity);
+      else
+        predictor = CreatePredictor(model, beta, alpha, lowerLimit.Data, upperLimit.Data);
       scope.AddVariable(new HeuristicLab.Core.Variable(scope.TranslateName("Predictor"), predictor));
       return null;
     }
 
-    public static IPredictor CreatePredictor(IGeneticProgrammingModel model, double punishmentFactor,
-  Dataset dataset, string targetVariable, int start, int end) {
-      return CreatePredictor(model, punishmentFactor, dataset, dataset.GetVariableIndex(targetVariable), start, end);
-    }
-
-
-    public static IPredictor CreatePredictor(IGeneticProgrammingModel model, double punishmentFactor,
-      Dataset dataset, int targetVariable, int start, int end) {
+    public static IPredictor CreatePredictor(IGeneticProgrammingModel model, double beta, double alpha, double lowerLimit, double upperLimit) {
 
       var evaluator = new HL3TreeEvaluator();
-      // evaluate for all rows
-      evaluator.PrepareForEvaluation(dataset, model.FunctionTree);
-      var result = from row in Enumerable.Range(start, end - start)
-                   let y = evaluator.Evaluate(row)
-                   let y_ = dataset.GetValue(row, targetVariable)
-                   select new { Row = row, Estimation = y, Target = y_ };
-
-      // calculate alpha and beta on the subset of rows with valid values 
-      var filteredResult = result.Where(x => IsValidValue(x.Target) && IsValidValue(x.Estimation));
-      var target = filteredResult.Select(x => x.Target);
-      var estimation = filteredResult.Select(x => x.Estimation);
-      double a, b;
-      if (filteredResult.Count() > 2) {
-        double tMean = target.Sum() / target.Count();
-        double xMean = estimation.Sum() / estimation.Count();
-        double sumXT = 0;
-        double sumXX = 0;
-        foreach (var r in result) {
-          double x = r.Estimation;
-          double t = r.Target;
-          sumXT += (x - xMean) * (t - tMean);
-          sumXX += (x - xMean) * (x - xMean);
-        }
-        b = sumXT / sumXX;
-        a = tMean - b * xMean;
-      } else {
-        b = 1.0;
-        a = 0.0;
-      }
-      double mean = dataset.GetMean(targetVariable, start, end);
-      double range = dataset.GetRange(targetVariable, start, end);
-      double minEstimatedValue = mean - punishmentFactor * range;
-      double maxEstimatedValue = mean + punishmentFactor * range;
-      evaluator.LowerEvaluationLimit = minEstimatedValue;
-      evaluator.UpperEvaluationLimit = maxEstimatedValue;
-      var resultModel = new GeneticProgrammingModel(MakeSum(MakeProduct(model.FunctionTree, b), a));
-      return new Predictor(evaluator, resultModel, minEstimatedValue, maxEstimatedValue);
+      evaluator.LowerEvaluationLimit = lowerLimit;
+      evaluator.UpperEvaluationLimit = upperLimit;
+      var resultModel = new GeneticProgrammingModel(MakeSum(MakeProduct(model.FunctionTree, beta), alpha));
+      return new Predictor(evaluator, resultModel, lowerLimit, upperLimit);
     }
-
-    private static bool IsValidValue(double d) {
-      return !double.IsInfinity(d) && !double.IsNaN(d);
-    }
-
 
     private static IFunctionTree MakeSum(IFunctionTree tree, double x) {
@@ -134 +96 @@
       return constX;
     }
-
-    private static void CalculateScalingParameters(IEnumerable<double> xs, IEnumerable<double> ys, out double k, out double d) {
-      if (xs.Count() != ys.Count()) throw new ArgumentException();
-      double xMean = xs.Sum() / xs.Count();
-      double yMean = ys.Sum() / ys.Count();
-
-      var yEnumerator = ys.GetEnumerator();
-      var xEnumerator = xs.GetEnumerator();
-
-      double sumXY = 0.0;
-      double sumXX = 0.0;
-      while (xEnumerator.MoveNext() && yEnumerator.MoveNext()) {
-        sumXY += (xEnumerator.Current - xMean) * (yEnumerator.Current - yMean);
-        sumXX += (xEnumerator.Current - xMean) * (xEnumerator.Current - xMean);
-      }
-
-      k = sumXY / sumXX;
-      d = yMean - k * xMean;
-    }
   }
 }

trunk/sources/HeuristicLab.GP.StructureIdentification/3.3/OffspringSelectionGPRegression.cs

@@ -198 +198 @@
       VariableInjector injector = base.CreateGlobalInjector();
       injector.AddVariable(new HeuristicLab.Core.Variable("PunishmentFactor", new DoubleData()));
+      // NB: LowerEstimationLimit and UpperEstimationLimit should replace the direct use of PunishmentFactor in the algorithm (gkronber 9 March, 2010)
+      injector.AddVariable(new HeuristicLab.Core.Variable("LowerEstimationLimit", new DoubleData(double.NegativeInfinity)));
+      injector.AddVariable(new HeuristicLab.Core.Variable("UpperEstimationLimit", new DoubleData(double.PositiveInfinity)));
       injector.AddVariable(new HeuristicLab.Core.Variable("BestValidationSolutionAge", new IntData()));
       injector.AddVariable(new HeuristicLab.Core.Variable("MaxBestValidationSolutionAge", new IntData()));

trunk/sources/HeuristicLab.GP.StructureIdentification/3.3/StandardGPRegression.cs

@@ -41 +41 @@
     public virtual string TargetVariable {
       get { return ProblemInjector.GetVariableValue<StringData>("TargetVariable", null, false).Data; }
-      set { ProblemInjector.GetVariableValue<StringData>("TargetVariable", null, false).Data = value; }
+      set {
+        ProblemInjector.GetVariableValue<StringData>("TargetVariable", null, false).Data = value;
+      }
     }
 
     public virtual Dataset Dataset {
       get { return ProblemInjector.GetVariableValue<Dataset>("Dataset", null, false); }
-      set { ProblemInjector.GetVariable("Dataset").Value = value; }
+      set {
+        ProblemInjector.GetVariable("Dataset").Value = value;
+      }
     }
 
@@ -101 +105 @@
     public virtual double PunishmentFactor {
       get { return GetVariableInjector().GetVariable("PunishmentFactor").GetValue<DoubleData>().Data; }
-      set { GetVariableInjector().GetVariable("PunishmentFactor").GetValue<DoubleData>().Data = value; }
+      set {
+        GetVariableInjector().GetVariable("PunishmentFactor").GetValue<DoubleData>().Data = value;
+      }
+    }
+
+    public virtual double UpperEstimationLimit {
+      get { return GetVariableInjector().GetVariable("UpperEstimationLimit").GetValue<DoubleData>().Data; }
+      set { GetVariableInjector().GetVariable("UpperEstimationLimit").GetValue<DoubleData>().Data = value; }
+    }
+    public virtual double LowerEstimationLimit {
+      get { return GetVariableInjector().GetVariable("LowerEstimationLimit").GetValue<DoubleData>().Data; }
+      set { GetVariableInjector().GetVariable("LowerEstimationLimit").GetValue<DoubleData>().Data = value; }
     }
 
@@ -193 +208 @@
       VariableInjector injector = base.CreateGlobalInjector();
       injector.AddVariable(new HeuristicLab.Core.Variable("PunishmentFactor", new DoubleData()));
+      // NB: LowerEstimationLimit and UpperEstimationLimit should replace the direct use of PunishmentFactor in the algorithm (gkronber 9 March, 2010)
+      injector.AddVariable(new HeuristicLab.Core.Variable("LowerEstimationLimit", new DoubleData(double.NegativeInfinity)));
+      injector.AddVariable(new HeuristicLab.Core.Variable("UpperEstimationLimit", new DoubleData(double.PositiveInfinity)));
       injector.AddVariable(new HeuristicLab.Core.Variable("BestValidationSolutionAge", new IntData()));
       injector.AddVariable(new HeuristicLab.Core.Variable("MaxBestValidationSolutionAge", new IntData()));
@@ -235 +253 @@
 
       IOperator combinedTerminationCriterion = AlgorithmBase.CombineTerminationCriterions(base.CreateTerminationCondition(), bestSolutionAge);
-      
+
       terminationCritertion.OperatorGraph.AddOperator(combinedTerminationCriterion);
       terminationCritertion.OperatorGraph.InitialOperator = combinedTerminationCriterion;

trunk/sources/HeuristicLab.Modeling/3.2/HeuristicLab.Modeling-3.2.csproj

@@ -84 +84 @@
     <Compile Include="AnalyzerModel.cs" />
     <Compile Include="BestSolutionStorer.cs" />
+    <Compile Include="LinearScaler.cs" />
     <Compile Include="SimpleStableCorrelationCoefficientEvaluator.cs" />
     <Compile Include="SimplePearsonCorrelationCoefficientEvaluator.cs" />

trunk/sources/HeuristicLab.Modeling/3.2/SimpleEvaluatorBase.cs

@@ -9 +9 @@
 namespace HeuristicLab.Modeling {
   public abstract class SimpleEvaluatorBase : OperatorBase {
-    protected const int ORIGINAL_INDEX = 0;
-    protected const int ESTIMATION_INDEX = 1;
+    public const int ORIGINAL_INDEX = 0;
+    public const int ESTIMATION_INDEX = 1;
 
     public virtual string OutputVariableName {