For a better human friendly man-machine interaction, it has been interested to study on classifiers which is essential in recognition systems for man-machine interaction.
Among many classifiers, Fuzzy Min-Max Neural Network (FMMNN) has attracted attention since the seminar paper by Simpson [1]. FMMNN has a very simple structure and a fast learning speed, and thus it can be simply implemented as hardware. It has, however, some problems in the sense that learning result depends on the ordering of input data and the training parameter that limits the size of hyperbox. To overcome the latter problem, multi-resolution approach may be an alternative.
In this thesis, a new method to alleviate the latter problem by using Committee Machine scheme is proposed. This method achieves the multi-resolution FMMNN without loss of incremental learning ability. Each expert is a FMMNN with a fixed training parameter. The gating network controls the output of experts according to the input.
The proposed method outperformed the single FMMNN with any training parameter and has the incremental learning ability.
