1996 Fiscal Year Final Research Report Summary
A Study on Fuzzy Symbolic Classifiers.
| Project/Area Number |
07680412
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Tokyo Denki University |
|---|
Principal Investigator |
ICHINO Manabu Faculty of Science and Engineering, Tokyo Denki University, Professor, 理工学部, 教授 (40057245)
|
|---|
| Project Period (FY) |
1995 – 1996
|
| Keywords | fuzzy classifier / pattern recognition / symbolic data / feature selection / class description / symbolic pattern / Cartesian System Model / membership function |
|---|
| Research Abstract |
The symbolic data analysis is a new research field which has been rapidly grown up guided by Professor E.Diday in France. In this field, the generalization of the data descriptions is one main stream, where each sample data may be described not only by numeric features but also by symbolic features.
The "fuzzy symbolic pattern classifiers" in this study can treat symbolic patterns which are simultaneously described by unmeric features and symbolic features. The following result are obtained in this study :
1.The approach treated here is based on the Cartesian system model (CSM) which is our own mathematical model to treat symbolic data. In this study we generalized the CSM itself in order to treat wider types of features used to describe sample patterns.
2.When samples are described only by numeric features, there exist fuzzy classifiers based on distance functions defined on the given set of sample patterns. We generalized this approach by introducing the generalized Minkowski metrics defined on the CSM.
3.As a different method, we developed region oriented fuzzy symbolic classifier in which each pattern class is described by "events" (rectangular regions in numeric feature cases) and the membership functions are defined between events and sample patterns. We found that, in this approach, "feature selection" is essentially important in order to take balance between the separability among patterm classes and the generality of class descriptions. The result was reported as an invited paper for the symbolic data analysis in Fifth Conference of the International Federation of Classification Societies (IFCS-96) held in Kobe.
|
