On the Applicability of the NetFAN-Approach to Function Approximation

Fuzzy systems were shown to be universal approximators, so are their trainable variant, the neuro-fuzzy systems. But fuzzy systems suffer from the curse of dimensionality, i.e. a very strong increase in computational and memory demands with an increasing number of input variables. This paper describes the NetFAN-approach to reduce this drawback by decomposition. It also proofs that such decomposed systems are universal approximators. The benchmark example of modeling the energy and water consumption of a building not only demonstrates that it achieves approximation capabilities like artificial neural networks. It also gives a notion how to utilize abstract background knowledge.