In the conventional learning methods of fuzzy inference systems using vector quantization, only input data is ordinary used to determine distribution of initial fuzzy rules. In order to improve the performance of vector quantization, it is needed to consider not only input data but output data. Therefore, some methods have already proposed. Though these methods show good performance compared with the conventional methods, sufficient effect is not shown yet. One of the problems is to use vector quantization only to determine the initial assignment of fuzzy rules. In this paper, we propose a new method composed of iterating two phases of vector quantization and learning. In the first step of the proposed method, the appearance frequency like probability density using input and output data is determined, and distribution of fuzzy rules is made properly by vector quantization. Secondly, the parameters of fuzzy rules are updated by the conventional learning method. Further, the first and second phases are iterated for the maximum learning time. In order to demonstrate the validity of the proposed method, numerical simulations for function approximation and pattern classification problems are performed.

fuzzy inference systems, vector quantization, neural gas, learning algorithm, appearance frequency.