A genetically optimized ensemble of is σ-FLNMAP neural classifiers based on non-parametric probability distribution functions

An advantage of the σ-FLNMAP neural network for classification is known to be the capacity to employ an underlying positive valuation function different than v(x) = x. This work demonstrates the effectiveness of a non-parametric probability distribution function used as an underlying positive valuation. Moreover a novel positive valuation function is introduced here analytically in a lattice of intervals. An ensemble of σ-FLNMAP classifiers is employed as a majority voting model whose parameters 1) a weight w_i, i = 1, ..., N for each constituent lattice, and 2) the number n_v of σ-FLNMAP voters, are estimated optimally from the training data using a genetic algorithm. Similarities and differences are delineated with various ensemble methods from the literature. The genetic-fuzzy-neural-computing techniques presented in this work, despite their computational complexity, imply significant comparative improvements in three benchmark classification problems. It is explained how both the adaptive resonance theory (ART) and the min-max neural networks can benefit from the tools presented here.