diff --git a/python/boomer/common/cpp/rule_refinement.cpp b/python/boomer/common/cpp/rule_refinement.cpp
index 4b5c9ada84..f135778ff5 100644
--- a/python/boomer/common/cpp/rule_refinement.cpp
+++ b/python/boomer/common/cpp/rule_refinement.cpp
@@ -482,6 +482,72 @@ ApproximateRuleRefinementImpl::ApproximateRuleRefinementImpl(AbstractStatistics*
 Refinement ApproximateRuleRefinementImpl::findRefinement(IHeadRefinement* headRefinement,
                                                          PredictionCandidate* currentHead,
                                                          uint32 numLabelIndices, const uint32* labelIndices) {
+    uint32 numBins = binArray_->numBins;
     Refinement refinement;
+    refinement.featureIndex = featureIndex_;
+    refinement.head = NULL;
+    refinement.indexedArray = NULL;
+    refinement.indexedArrayWrapper = NULL;
+
+    PredictionCandidate* dynamicCurrentHead = currentHead;
+
+    PredictionCandidate* bestHead = currentHead;
+
+    std::unique_ptr<IStatisticsSubset> statisticsSubsetPtr;
+    statisticsSubsetPtr.reset(statistics_->createSubset(numLabelIndices, labelIndices));
+
+    uint32 r = 0;
+    while(binArray_->bins[r].numExamples == 0 && r < numBins){
+        r++;
+    }
+    statisticsSubsetPtr.get()->addToSubset(r, 1);
+    uint32 previousR = r;
+    float32 previousValue = binArray_->bins[r].maxValue;
+    uint32 numCoveredExamples = binArray_->bins[r].numExamples;
+
+    r += 1;
+    for(r; r < numBins; r++){
+        uint32 numExamples = binArray_->bins[r].numExamples;
+
+        if(numExamples > 0){
+            numCoveredExamples += numExamples; //Das sollten wir, anders wie im Pseudo Code, besser schon hier machen, oder?
+            float32 currentValue = binArray_->bins[r].minValue;
+            dynamicCurrentHead = headRefinement->findHead(bestHead, refinement.head, labelIndices,
+                                                          statisticsSubsetPtr.get(), false, false);
+
+            if(dynamicCurrentHead != NULL){
+                bestHead = dynamicCurrentHead;
+                refinement.comparator = LEQ;
+                refinement.threshold = (previousValue + currentValue)/2.0;
+                refinement.start = 0;
+                refinement.end = r;
+                refinement.previous = previousR;
+                refinement.coveredWeights = numCoveredExamples;
+                refinement.covered = true;
+            }
+
+            dynamicCurrentHead = headRefinement->findHead(bestHead, refinement.head, labelIndices,
+                                                          statisticsSubsetPtr.get(), false, false);
+
+            if(dynamicCurrentHead != NULL){
+                bestHead = dynamicCurrentHead;
+                refinement.comparator = GR;
+                refinement.threshold = (previousValue + currentValue)/2.0;
+                refinement.start = 0;
+                refinement.end = r;
+                refinement.previous = previousR;
+                refinement.coveredWeights = numCoveredExamples;
+                refinement.covered = false;
+            }
+
+            previousValue = binArray_->bins[r].maxValue;
+            previousR = r;
+            statisticsSubsetPtr.get()->addToSubset(r, 1);
+
+        }
+
+    }
+
+
     return refinement;
 }