fixed bias calculation (it was an older version than the final one in

stable-clean)

bundles/fr.irisa.cairn.gecos.typeexploration/src/fr/irisa/cairn/gecos/typeexploration/dse/TabuExploration.java

@@ -244,23 +244,34 @@ public class TabuExploration extends AbstractExplorationAlgorithm {
 		List<Integer> lastModifiedWLs = nextSolutionCandidates.stream()
 				.map(e -> e.getCurrentW(e.lastModifiedSymbol))
 				.collect(toList());
-
+		
 		List<Double> accuracyValues = nextSolutionCandidates.stream()
-			.map(e -> getAccuracy(e.getLastCreatedSolution()).get())
-			.collect(toList());
-		ValuesNormalizer accuracyNormalizer = new ValuesNormalizer(accuracyValues);
+				.map(e -> getAccuracy(e.getLastCreatedSolution()).get())
+				.collect(toList());
 		
 		List<Double> costValues = nextSolutionCandidates.stream()
 				.map(e -> getCost(e.getLastCreatedSolution()).get())
 				.collect(toList());
-		ValuesNormalizer costNormalizer = new ValuesNormalizer(costValues);
 		
+		ValuesNormalizer accuracyNormalizer = new ValuesNormalizer(accuracyValues);
+		ValuesNormalizer costNormalizer = new ValuesNormalizer(costValues);
+		//FIXME these values are currently arbitrary
+		{
+			//max must be obtained before setting min or range, since it is computed from min and range
+			final double accMax = accuracyNormalizer.getMax();
+			final double costMax = costNormalizer.getMax();
+			accuracyNormalizer.setRange(Math.max(accuracyMetric.getNormalizationRangeLB(), accuracyNormalizer.getRange()));
+			accuracyNormalizer.setMin(accMax-accuracyNormalizer.getRange());
+			costNormalizer.setRange(Math.max(costMetric.getNormalizationRangeLB(), costNormalizer.getRange()));
+			costNormalizer.setMin(costMax-costNormalizer.getRange());
+		}
+
 		Double normalizedCurrentAccuray = getAccuracy(currentSolution).map(accuracyNormalizer::normalize).orElse(null);
 		Double normalizedCurrentCost = getCost(currentSolution).map(costNormalizer::normalize).orElse(null);
-
+		
 		List<Double> normalizedAccuracyValues = accuracyNormalizer.normalize(accuracyValues);
 		List<Double> normalizedCostValues = costNormalizer.normalize(costValues);
-		
+
 		List<Double> criteria = new LinkedList<>();
 		for (int i = 0; i < nextSolutionCandidates.size(); i++) {
 			Integer lastModifiedWL = lastModifiedWLs.get(i);
@@ -269,15 +280,23 @@ public class TabuExploration extends AbstractExplorationAlgorithm {
 			Double ratio = (1 + computeImprovement(normalizedCurrentAccuray, normalizedAccuracyValue, accuracyMetric.isHigherBetter())) /
 			(1 + computeDegradation(normalizedCurrentCost, normalizedCostValue, costMetric.isHigherBetter()));
 			
-			Double WLbias = lastModifiedWL.doubleValue() / (10000);
+			//FIXME bias strength is currently not well thought out
+			Double WLbias = Math.pow(lastModifiedWL.doubleValue(), 1.2) / (350); //TODO try 250
 			
 			final Double criterion;
+			
 			if (directionUp)
 				criterion = ratio - WLbias;
 			else
 				criterion = ratio + WLbias;
 			
-			logger.warning(String.format("TabuSearchMetric: %f -> %f", ratio, criterion));
+			WordLengthsExplorer candidate = nextSolutionCandidates.get(i);
+			logger.info(String.format("TabuSearchMetric (%s; WL:%d, Acc:%f, Cost:%f, NormAcc:%f, NormCost:%f): %f -> %f", 
+					candidate.getLastModifiedSymbol().getName(), lastModifiedWL.intValue(),
+					accuracyValues.get(i), costValues.get(i),
+					normalizedAccuracyValue, normalizedCostValue,
+					ratio, criterion));
+			
 			criteria.add(criterion);
 			nextSolutionCandidates.get(i).currentSolution.setNote("gradient", criterion+"");
 			nextSolutionCandidates.get(i).currentSolution.setNote("selected", "F");