Change-glasses Pattern Classification with a Fuzzy Neural Network

INTRODUCTION "Change-glasses" approach in pattern recognition [1] relies on the assumption that there are subspaces of the initial feature space where the chosen classification rule can be replaced by another, more competent, one (see, e.g. [2]). This will hopefully lead to a better classification accuracy in comparison with that obtained through one rule only. This classification strategy bears an analogy with the process of working up a diagnosis in medicine. If the physician does not feel competent to resolve a special case, he summons a consultation team of professionals in that particular field. The main question that arises here concerns the way of partitioning the feature space. There are different approaches underlying some heuristic classification paradigms. In fact, every rule-based classifier performs a partitioning through antecedent clauses and assigns a classification rule to each region through the implication. A partition may be based on the geometric properties of the classes detected by a preliminary clustering [3] or by sequential groping about for the class boundaries [4]. In the fuzzy classification rule described in [5,6] the partitioning is uniform, i.e. the regions continue to be split until a sufficiently high certainty of the rule, generated by each region, is achieved. In this way, the decision boundary is approximated as precisely, as necessary. The problem is how to guarantee that the generalization capability of the classifier is sufficiently high if the regions contain only few objects. Trying to prevent this case, we suggest to use only two regions designated as 'doubtful' and 'undoubtful', respectively. Note that each of these may not necessarily be compact and may consist of more than one disjoint subregions. These regions are then treated by different classification rules. For the purposes of partitioning we used a previously developed technology based on a neural network with fuzzy inputs [7,8]. The details are summarized in Section 2. Section 3 contains an experimental illustration.