Changeset 15859 for branches/2886_SymRegGrammarEnumeration/HeuristicLab.Algorithms.DataAnalysis.SymRegGrammarEnumeration

Timestamp:

03/23/18 15:41:02 (7 years ago)

Author:

lkammere

Message:

#2886: Add simple constant optimization to evaluation of solutions.

File:

• branches/2886_SymRegGrammarEnumeration/HeuristicLab.Algorithms.DataAnalysis.SymRegGrammarEnumeration/Analysis/RSquaredEvaluator.cs (modified) (3 diffs)

branches/2886_SymRegGrammarEnumeration/HeuristicLab.Algorithms.DataAnalysis.SymRegGrammarEnumeration/Analysis/RSquaredEvaluator.cs

@@ -1 +1 @@
 using System;
+using System.Diagnostics;
 using HeuristicLab.Algorithms.DataAnalysis.SymRegGrammarEnumeration.GrammarEnumeration;
 using HeuristicLab.Common;
@@ -63 +64 @@
 
     private void EvaluateSentence(GrammarEnumerationAlgorithm algorithm, SymbolString symbolString) {
+      var problemData = algorithm.Problem.ProblemData;
+
       SymbolicExpressionTree tree = algorithm.Grammar.ParseSymbolicExpressionTree(symbolString);
+      Debug.Assert(SymbolicRegressionConstantOptimizationEvaluator.CanOptimizeConstants(tree));
 
-      var problemData = algorithm.Problem.ProblemData;
+      // TODO: Initialize constant values randomly
+      // TODO: Restarts
+
+      double r2 = SymbolicRegressionConstantOptimizationEvaluator.OptimizeConstants(expressionTreeLinearInterpreter,
+        tree,
+        problemData,
+        problemData.TrainingIndices,
+        applyLinearScaling: false,
+        maxIterations: 200,
+        updateVariableWeights: true,
+        updateConstantsInTree: true);
+
+      foreach (var symbolicExpressionTreeNode in tree.IterateNodesPostfix()) {
+        ConstantTreeNode constTreeNode = symbolicExpressionTreeNode as ConstantTreeNode;
+        if (constTreeNode != null && constTreeNode.Value.IsAlmost(0.0)) {
+          constTreeNode.Value = 0.0;
+        }
+      }
+
       SymbolicRegressionModel model = new SymbolicRegressionModel(
         problemData.TargetVariable,
@@ -71 +93 @@
         expressionTreeLinearInterpreter);
 
-      var target = problemData.TargetVariableTrainingValues;
-      var estVals = model.GetEstimatedValues(problemData.Dataset, problemData.TrainingIndices);
-      OnlineCalculatorError error;
-      var r2 = OnlinePearsonsRCalculator.Calculate(target, estVals, out error);
-      if (error != OnlineCalculatorError.None) r2 = 0.0;
+      //var target = problemData.TargetVariableTrainingValues;
+      //var estVals = model.GetEstimatedValues(problemData.Dataset, problemData.TrainingIndices);
+      //OnlineCalculatorError error;
+      //r2 = OnlinePearsonsRCalculator.Calculate(target, estVals, out error);
+      //if (error != OnlineCalculatorError.None) r2 = 0.0;
 
       var bestR2Result = (DoubleValue)algorithm.Results[BestTrainingQualityResultName].Value;
-      if (r2 > bestR2Result.Value) {
+      bool better = r2 > bestR2Result.Value;
+      bool equallyGood = r2.IsAlmost(bestR2Result.Value);
+      bool shorter = algorithm.BestTrainingSentence != null && symbolString.Count() < algorithm.BestTrainingSentence.Count();
+      if (better || (equallyGood && shorter)) {
         bestR2Result.Value = r2;
         algorithm.Results.AddOrUpdateResult(BestTrainingModelResultName, model);