Neuro-fuzzy approach for modeling complex functional mappings

Abstract

Function approximators are needed in lots of applications to model nonlinear functional mappings. Where no formal description exists, computational intelligence methods may be used. But knowledge-based systems suffer from the knowledge engineering bottleneck as well as from the curse of dimensionality if the number of input variables increases. Artificial neural networks can handle such complex applications, but they are a black box-approach. Hence learned knowledge cannot be analyzed or improved manually. In this article the NetFAN-approach (Network of Fuzzy Adaptive Nodes) is described which combines decomposition of a neuro-fuzzy system and learning in order to apply neurofuzzy methods to applications with an increased number of input variables while keeping the advantages of neuro-fuzzy systems like interpretability of learned knowledge. Its applicability as a function approximator is demonstrated by The Great Energy Predictor Shootout benchmark problem. In this example, results were achieved which are comparable to the top benchmark candidates.