Strategies to identify fuzzy rules directly from certainty degrees: a comparison and a proposal

With identification methods that learn fuzzy rules directly from certainty degrees, we refer to methods that select the most promising rules from the training examples in only one pass. In order to do that, these methods employ a certainty measure to assess the goodness of each rule. This paper aims to analyze in depth the behaviors and features of two different strategies for identifying fuzzy models from certainty degrees, each of both combined with one of two well-known alternatives for measuring the certainty degrees of the rules. With this aim, the advantages and drawbacks of each method are analyzed experimentally by considering the model error when applied to several systems. Besides, the robustness of the results is investigated by applying the methods to noisy data. As a conclusion, a new method combining the best components of the previously considered methods is proposed and its results are analyzed. The achieved performance in accuracy and computational cost shows the benefit of this new method.