Identification algorithms for fuzzy relational matrices. II: Optimizing algorithms (Q1971888)

The paper reviews and evaluates four fuzzy identification algorithms, known in the literature: (1) the recursive parameter tuning method, (2) the probabilistic descent method, (3) the neural learning method, and (4) the quasi-Newton minimization method. The methods under study include a user-specified fuzzy performance criterion in the development and identify fuzzy relational matrices by numerical minimization of an adequate performance index. The same basis as in part one of the paper (i.e. the same input-output data, membership functions and defuzzification procedure) is assumed for comparison of the algorithms. All of the reviewed identification procedures give comparable results, and are better than the non-optimizing algorithms studied in part one, except the recursive parameter tuning method which is judged to be unacceptable. The relationship between fuzzy and non-fuzzy (discrete) performance criteria is also studied. The results of simulations show that there is an inverse relationship between minimization of a fuzzy and a discrete criterion for the algorithms under consideration. The conclusions of the paper refer to the conditions assumed for evaluating the algorithms.