Self-configuring hybrid evolutionary algorithm for fuzzy classifier design with active learning for unbalanced datasets

The paper describes an active training example selection for a self-configured hybrid evolutionary algorithm for fuzzy rule bases design for classification problems. This method is related to instance selection methods, which allow not only decreasing of required computational recourses, but also increasing the quality of the obtained classifiers. The method changes the probabilities of instances which are selected into the training subsample depending on how good they are classified by the algorithm. After several generations the sample is changed and probabilities are recalculated. Those instances which were not used before and those which were misclassified by the algorithm had higher probabilities of getting into the training sample. The probabilities of instance selection were calculated using a procedure similar to proportional selection in the genetic algorithm. The idea of training instance selection described here was implemented for the fuzzy classifiers forming. This algorithm uses the combination of Pittsburg and Michigan approach for fuzzy rule base design with fixed terms, and the Michigan approach is used together with the mutation operator. The size of the rule base is not fixed, and may change during the algorithm run, and a corresponding class number and the rule weight were calculated heuristically for every rule. Moreover, the algorithm uses an initialization procedure that uses instances from the sample to generate more accurate rules. In the Michigan part the operators of adding rules, deleting rules and replacing rules has been implemented. The creation of new rules could be performed by genetic approach - using the existing rules, and heuristically - using those instances which were misclassified. The efficiency of the algorithm was shown on a set of complex classification problems with several classes, as an efficiency measure the overall accuracy and the average accuracy among classes was used.