Index: trunk/CrossPare/src/de/ugoe/cs/cpdp/training/GPTraining.java =================================================================== --- trunk/CrossPare/src/de/ugoe/cs/cpdp/training/GPTraining.java (revision 103) +++ trunk/CrossPare/src/de/ugoe/cs/cpdp/training/GPTraining.java (revision 104) @@ -1,4 +1,5 @@ package de.ugoe.cs.cpdp.training; +import java.util.LinkedList; import java.util.List; @@ -47,5 +48,5 @@ public class GPTraining implements ISetWiseTrainingStrategy, IWekaCompatibleTrainer { - private GPVClassifier classifier = new GPVClassifier(); + private GPVVClassifier classifier = new GPVVClassifier(); private int populationSize = 1000; @@ -342,5 +343,7 @@ /** * GP Multiple Data Sets Validation-Voting Classifier - * + * + * As the GP Multiple Data Sets Validation Classifier + * But here we do keep a model candidate for each training set which may later vote * */ @@ -360,21 +363,48 @@ // then is evaluated on the rest for(int i=0; i < traindataSet.size(); i++) { - Classifier classifier = new GPRun(); - - // one project is training data - classifier.buildClassifier(traindataSet.get(i)); - + + // candidates we get out of evaluation + LinkedList candidates = new LinkedList<>(); + + // 200 runs + + for(int k=0; k < 200; k++) { + Classifier classifier = new GPRun(); + + // one project is training data + classifier.buildClassifier(traindataSet.get(i)); + + double[] errors; + // rest of the set is evaluation data, we evaluate now + for(int j=0; j < traindataSet.size(); j++) { + if(j != i) { + // if type1 and type2 errors are < 0.5 we allow the model in the final voting + errors = this.evaluate((GPRun)classifier, traindataSet.get(j)); + if((errors[0] / traindataSet.get(j).numInstances()) < 0.5 && (errors[0] / traindataSet.get(j).numInstances()) < 0.5) { + candidates.add(classifier); + } + } + } + } + + // now after the evaluation we do a model selection where only one model remains for the given training data + double smallest_error_count = Double.MAX_VALUE; double[] errors; - - // rest of the set is evaluation data, we evaluate now - for(int j=0; j < traindataSet.size(); j++) { - if(j != i) { - // if type1 and type2 errors are < 0.5 we allow the model in the final voting - errors = this.evaluate((GPRun)classifier, traindataSet.get(j)); - if((errors[0] / traindataSet.get(j).numInstances()) < 0.5 && (errors[0] / traindataSet.get(j).numInstances()) < 0.5) { - classifiers.add(classifier); + Classifier best = null; + for(int ii=0; ii < candidates.size(); ii++) { + for(int j=0; j < traindataSet.size(); j++) { + if(j != i) { + errors = this.evaluate((GPRun)candidates.get(ii), traindataSet.get(j)); + + if(errors[0]+errors[1] < smallest_error_count) { + best = candidates.get(ii); + } } } } + + // now we have the best classifier for this training data + classifiers.add(best); + } } @@ -407,5 +437,5 @@ } - if(vote_positive >= 3) { + if(vote_positive >= (classifiers.size()/2)) { return 1.0; }else { @@ -430,6 +460,7 @@ public class GPVClassifier extends AbstractClassifier { + private List classifiers = null; private Classifier best = null; - + private static final long serialVersionUID = 3708714057579101522L; @@ -449,19 +480,66 @@ // then is evaluated on the rest for(int i=0; i < traindataSet.size(); i++) { - Classifier classifier = new GPRun(); - - // one project is training data - classifier.buildClassifier(traindataSet.get(i)); - - // rest of the set is evaluation data, we evaluate now + + // candidates we get out of evaluation + LinkedList candidates = new LinkedList<>(); + + // 200 runs + for(int k=0; k < 200; k++) { + Classifier classifier = new GPRun(); + + // one project is training data + classifier.buildClassifier(traindataSet.get(i)); + + double[] errors; + + // rest of the set is evaluation data, we evaluate now + for(int j=0; j < traindataSet.size(); j++) { + if(j != i) { + // if type1 and type2 errors are < 0.5 we allow the model in the final voting + errors = this.evaluate((GPRun)classifier, traindataSet.get(j)); + if((errors[0] / traindataSet.get(j).numInstances()) < 0.5 && (errors[0] / traindataSet.get(j).numInstances()) < 0.5) { + candidates.add(classifier); + } + } + } + } + + // now after the evaluation we do a model selection where only one model remains per training data set + // from that we chose the best model + + // now after the evaluation we do a model selection where only one model remains for the given training data double smallest_error_count = Double.MAX_VALUE; double[] errors; - for(int j=0; j < traindataSet.size(); j++) { - if(j != i) { - errors = this.evaluate((GPRun)classifier, traindataSet.get(j)); - if(errors[0]+errors[1] < smallest_error_count) { - this.best = classifier; + Classifier best = null; + for(int ii=0; ii < candidates.size(); ii++) { + for(int j=0; j < traindataSet.size(); j++) { + if(j != i) { + errors = this.evaluate((GPRun)candidates.get(ii), traindataSet.get(j)); + + if(errors[0]+errors[1] < smallest_error_count) { + best = candidates.get(ii); + } } } + } + + // now we have the best classifier for this training data + classifiers.add(best); + } + + // now determine the best classifier for all training data + double smallest_error_count = Double.MAX_VALUE; + double error_count; + double errors[]; + for(int j=0; j < classifiers.size(); j++) { + error_count = 0; + Classifier current = classifiers.get(j); + for(int i=0; i < traindataSet.size(); i++) { + errors = this.evaluate((GPRun)current, traindataSet.get(i)); + error_count = errors[0] + errors[1]; + } + + if(error_count < smallest_error_count) { + best = current; } } @@ -472,5 +550,5 @@ final Classifier classifier = new GPRun(); classifier.buildClassifier(traindata); - best = classifier; + classifiers.add(classifier); }