Boosting of Fuzzy Rules with Low Quality Data

An extension of the Adaboost algorithm is proposed for obtaining fuzzy rule based classifiers from imprecisely perceived data. Isolated fuzzy rules are regarded as weak learners, and knowledge bases as ensembles. Rules are iteratively added to a base, and the search of the best rule at each iteration is carried out by a genetic algorithm driven by a fuzzy fitness function. The successive weights of the instances are also fuzzy, however each rule is assigned a crisp number of votes, interpreted as degrees of importance of these rules. The results of the new algorithm are compared to those of other genetic algorithms for low quality data, in both accuracy and linguistic qualilty.