The simplified fuzzy rules have the advantage that the system can output real number values directly and are frequently applied to the field of control systems. In general, the precision of the fuzzy rules becomes higher as fuzzy partitions increase in number. However, when the teaching data is in both cases below, tuning of the fuzzy rules by conventional methods is not satisfactory. In the first case, where the teaching data is arranged unequally, the reasoning error by the fuzzy rules in sparse domains becomes large because the information necessary to generate the fuzzy rules is insufficient, and, in extreme cases the fuzzy rules can't be generated. Moreover, even if the fuzzy partitions increase in number, the reasoning error never comes down. On the contrary, it goes up, because the fuzzy rules not based on teaching data increase. In the second case, where the teaching data contains disturbance, the reasoning error of the fuzzy rules becomes larger. In this case, the sensitivity to the disturbance becomes higher as the fuzzy partitions increase. Of course, we cannot decrease the error by increasing the fuzzy partitions. In this paper, we propose a fuzzy reasoning model of multiple division layers using iterative transfer window to solve the above problems. This model has a multi-layer structure having both the advantages of rough fuzzy partitions and smooth fuzzy partitions. Moreover, the redundant fuzzy division layers can be decreased by an iterative transfer window. Finally, through identification experiments of two nonlinear functions having sparse domains and disturbance as teaching data, we will show the fuzzy rules generated by the proposed method have high identification precision even for those problems, and also have a low sensitivity to teaching data obtained from actual environment such as the above problems. That is to say, it was verified that the proposed method has a high generalization ability.
View full abstract