Learning algorithms with regularization criteria for fuzzy reasoning model

This paper describes two learning algorithms using regularization terms for fuzzy reasoning model. The proposed algorithms are learning ones for parameters of the antecedent and the consequent parts. In the first algorithm, the entropy function as a regularization term for parameters of the antecedent part in fuzzy reasoning rule is introduced. Then, the parameters of the center and the width are adjusted so as to be 0.5 for each membership value. In the second algorithm, the weight elimination function as a regularization term for the weights of the consequent part is proposed. The model to have a few weights of large absolutes or many weights of small absolutes is constructed. Some numerical simulations are performed to show the validity of the proposed methods. Specially, it is shown that the use of the weight elimination function decreases learning time.