Accepted:
15 Oct 1998
| |
/vol06/ishibuchi-nii/ | © Copyright 1998 | |||
| Volume 06 | Received: Accepted: |
01 Jul 1998 15 Oct 1998 |
|||
|
Linguistic rule extraction from neural networks for high-dimensional classification problems
Hisao Ishibuchi, Manabu Nii, Kimiko Tanaka |
|
| Abstract | |
| In this paper, we show how linguistic rules Êcan be extracted from trained neural networks Êfor high-dimensional pattern classification Êproblems. In our rule extraction method, antecedent linguistic values such as " small " and " large " are presented to a trained neural network for extracting linguistic rules. Since input values are handled as fuzzy Ênumbers, outputs from the trained neural network are also calculated as fuzzy numbers by fuzzy arithmetic. When linguistic values in the antecedent part of a linguistic rule is presented to the trained neural network, the consequent class and the certainty grade are specified by the corresponding fuzzy output vector. In this manner, all combinations of antecedent linguistic values are examined as input vectors to the trained neural network for extracting linguistic rules. While our rule extraction method works very well for low-dimensional pattern classification problems, there exist several difficulties in the application to high-dimensional problems. One difficulty is the exponential increase of the number of possible combinations of antecedent linguistic values. Another difficulty is excess fuzziness of the calculated fuzzy output vector, which prevents our rule extraction method from appropriately specifying the consequent class and the certainty grade. From the viewpoint of the understandability of extracted knowledge, it is also another difficulty that a large number of linguistic rules are extracted from the trained neural network. In this paper, we show how these difficulties can be remedied in our rule extraction method. Simulation results on a real-world pattern classification problem with many continuous attributes show that classification knowledge can be extracted from trained neural networks in an understandable form. | |
| Full Text |
|
|
Multimedia Links (none) Reference Links (none) Citation Reference |
Get viewers for PS & PDF   |
