Fuzzy adaptive multi-module approximation network

The paper presents a fuzzy version of the Adaptive Multi-module Approximation Network. New modules are generated when performance of existing modules is inadequate for some training data, and the applicability of a module to each input vector is determined based on the fuzzy membership of that vector in the possibly asymmetric clusters represented by the reference vectors associated with different modules. The main idea is that for neural networks that rely on a reference vector (for vector quantization, clustering, and similar tasks), the use of fuzzy membership criterion based on the distribution of data (inside different Voronoi cells) may be more appropriate than the traditional approach using a Euclidean metric to determine to which cell each data point belongs.