In this work, we propose several algorithms for learning first-order TSK fuzzy rules. These methods, based on two stages, first generate a set of candidate rules with an adaptation of the apriori algorithm for frequent item set search. Then, they select a subset of such rules by means local search or a genetic algorithm. The results obtained show that, genetic algorithms tends to converge to systems with a high number of rules which minimize the training error, but also overfit. On the other hand, local search gets stuck in configurations with few rules which, despite producing a higher training error, avoid overfitting and produce the best results in terms of error.
