Changeset 14301

Timestamp:

09/22/16 17:00:29 (8 years ago)

Author:

bburlacu

Message:

#2635: Use linear scaling in the evaluator to bring the target in the range of the estimated (work in progress)

File:

• branches/HeuristicLab.OSGAEvaluator/HeuristicLab.OSGAEvaluator/SymbolicRegressionSingleObjectiveOSGAEvaluator.cs (modified) (2 diffs)

branches/HeuristicLab.OSGAEvaluator/HeuristicLab.OSGAEvaluator/SymbolicRegressionSingleObjectiveOSGAEvaluator.cs

@@ -292 +292 @@
         #endregion
       } else {
-        var calculator = new OnlinePearsonsRCalculator();
         var trainingPartitionSize = problemData.TrainingPartition.Size;
         var interval = (int)Math.Floor(trainingPartitionSize * RelativeFitnessEvaluationIntervalSize);
@@ -299 +298 @@
         // use the actual estimated values for the first i * interval rows of the training partition and and assume the remaining rows are perfectly correlated
         // if the quality of the individual still falls below the parent quality, then we can reject it sooner, otherwise as i increases the whole estimated series will be used
+        var lsc = new OnlineLinearScalingParameterCalculator();
+        var rcalc = new OnlinePearsonsRCalculator();
+        var actualQuality = SymbolicRegressionSingleObjectivePearsonRSquaredEvaluator.Calculate(SymbolicDataAnalysisTreeInterpreterParameter.ActualValue, SymbolicExpressionTreeParameter.ActualValue, estimationLimits.Lower, estimationLimits.Upper, problemData, problemData.TrainingIndices, false);
         for (int i = 0; i < trainingPartitionSize; i += interval) {
-          calculator.Reset();
-          // save estimated values into the list (for caching)
+          var start = problemData.TrainingPartition.Start;
           int end = Math.Min(trainingPartitionSize, i + interval);
+          // cache estimated values 
+          // scale target values to the range of the estimated values
           for (int j = i; j < end && e.MoveNext(); ++j) {
             estimated.Add(e.Current);
-          }
-          var start = problemData.TrainingPartition.Start;
-          // add (estimated, target) pairs to the calculator
-          for (int j = 0; j < end; ++j) {
             var index = j + start;
-            calculator.Add(targetValues[index], estimated[j]);
-          }
-          // add (target, target) pairs to the calculator (simulate perfect correlation on the remaining rows)
+            // in the context of the linear scaling calculator, the target value becomes the "original" 
+            // while the estimated value becomes the "target" (because we want to scale the target in the range of the estimated)
+            lsc.Add(targetValues[index], e.Current);
+          }
+          var a = lsc.Alpha; // additive scaling term
+          var b = lsc.Beta;  // multiplicative scaling factor
+          // calculate the quality
+          for (int j = i; j < end; ++j) {
+            var index = j + start;
+            rcalc.Add(estimated[j], targetValues[index]);
+          }
+          var rcalc2 = (OnlinePearsonsRCalculator)rcalc.Clone();
           for (int j = end; j < trainingPartitionSize; ++j) {
             var index = j + start;
-            var v = targetValues[index];
-            calculator.Add(v, v);
-          }
-          var r = calculator.ErrorState == OnlineCalculatorError.None ? calculator.R : double.NaN;
+            var v = targetValues[index] * b + a;
+            rcalc2.Add(v, targetValues[index]);
+          }
+          var r = rcalc2.ErrorState == OnlineCalculatorError.None ? rcalc2.R : double.NaN;
           quality = r * r;
-          if (!(quality > parentQuality))
-            break;
+          bool falseReject = false;
+          if (!(quality > parentQuality)) {
+            if (actualQuality > parentQuality)
+              falseReject = true;
+          }
+          //          if (!(quality > parentQuality))
+          //            break;
         }
+
         return quality;
       }