A new method for system modelling and pattern classification (Q2570203)

This study introduces a new class of neurofuzzy systems endowed with some generalized models of logic connectives realized in terms of t-norms and s-conorms. The generalization of the connectives is realized in terms of their weighted versions. Given the ``original'' t-norm of ``\(n\)'' arguments, \(\text{t}(a_1, a_2,\dots,a_n)\), it is augmented by a vector of weights \({\mathbf w} = [w_1 w_2\dots w_n]^T\) such that \(\mathop T^n_{i=1}(1 -w_i(1 -a_i))\). For any s-conorm, \(\text{s}(a_1, a_2,\dots,a_n)\), its weighted version comes in the form \(\mathop S^n_{i=1}(w_ia_i)\) with the weights denoted by \(w_1,w_2,\dots,\) and \(w_n\). Two main categories of inference schemes are discussed (that is the Mamdani-type and logical type). Covered are essential computational details along with pertinent optimization aspects and gradient-based learning mechanisms involved in these networks. A suite of numerical experiments is used to illustrate the performance of the resulting neurofuzzy networks and contrast their performance with some other fuzzy and neurofuzzy models available in the literature.