A learning methodology for coherent hybrid probabilistic fuzzy classifiers

We aim at redesigning the hybrid fuzzy classifier proposed in \cite{demeloetal} that joins together probabilistic inference with classical Wang-Mendel fuzzy rule bases. We will profit from coherent probabilistic fuzzy IF-THEN rules, as already described in \cite{colettipetturitivantaggi}, with a novel elicitation strategy based on a new learning methodology. This will lead us to propose a probabilistic fuzzy rule based classification algorithm. The methodology for constructing and drawing inferences from a probabilistic fuzzy rule based classifier guarantees the global coherence of the probability evaluations and allows to take into account potentially imprecise (lower-upper) probabilistic conclusions. The proposed classification algorithm will be tested on a doping alert problem and compared with two other fuzzy IF-THEN rule based classifiers on artificial datasets.