A Construction Method of Fuzzy Systems Using Vector Quantization

Abstract

We propose a learning method of fuzzy inference rulesusing a vector quantization, neural gas network. Somemodels using self-organization or vector quantization byneural networks have been proposed in previous studies.These models show good results for the number of fuzzyinference rules in high dimensional problems. However, mostof these models determine a distribution of initial fuzzyinference rules by considering only input data. In thispaper, so as to make a more proper distribution of theinitial fuzzy inference rules in input space, we propose amethod considering not only input data but output data.Further, the number of fuzzy inference rules is determinedto an objective value (threshold of inference error) in aconstructive way. In order to demonstrate the validity ofthe proposed method, some numerical examples areperformed.