LIP6 1999/008
- Thesis
Apprentissage supervisé par génération de règles : le système SUCRAGE - A. Borgi - Ben Bouzid
- 269 pages - 01/15/1999- document en - http://www.lip6.fr/lip6/reports/1999/lip6.1999.008.ps.tar.gz - 2,474 Ko
- Contact : Amel.Borgi (at) nulllip6.fr
- Ancien Thème : APA
- Keywords : Supervised learning, production rules, imprecision, uncertainty, approximate reasoning, image processing
- Publisher : Valerie.Mangin (at) nulllip6.fr
Facing the increase of data amount recorded daily, the detection of both structures and specific links between them, the organisation and the search of exploitable knowledge in this information become a strategic stake for decision holding and prediction task.
This complex problem known as Data Mining has multiple aspects. We focus on one of them : supervised learning. We propose a learning method from examples situated at the junction of statistical methods and those based on Artificial Intelligence techniques. Our modelisation is based on automatic generation of classification rules and on an original use of approximate reasoning. The classification function is directly given in the form of production rules base. This ensures the transparency and easy interpretation of the classifier.
The proposed learning method is multi-features, it allows to take into account the possible predictive power of a simultaneously considered features conjunction. The feature space partition allows a multi-valued representation of the features and data imprecision integration. The rule uncertainty is managed in the learning phase as well as in the recognition one. To introduce more flexibility and overcome the boundary problem due to the discretisation, we propose to use approximate reasoning. The originality of our approach lies in using the proposed approximate reasoning not only as an inference mode and to manage imprecise knowledge, but also to refine learning and validate the rule base.
The proposed method was implemented in a tool called SUCRAGE and confronted with a real application in the field of image processing (multi-components image segmentation). The obtained results are very satisfactory. They validate our approach and allow us to consider other application fields.
This complex problem known as Data Mining has multiple aspects. We focus on one of them : supervised learning. We propose a learning method from examples situated at the junction of statistical methods and those based on Artificial Intelligence techniques. Our modelisation is based on automatic generation of classification rules and on an original use of approximate reasoning. The classification function is directly given in the form of production rules base. This ensures the transparency and easy interpretation of the classifier.
The proposed learning method is multi-features, it allows to take into account the possible predictive power of a simultaneously considered features conjunction. The feature space partition allows a multi-valued representation of the features and data imprecision integration. The rule uncertainty is managed in the learning phase as well as in the recognition one. To introduce more flexibility and overcome the boundary problem due to the discretisation, we propose to use approximate reasoning. The originality of our approach lies in using the proposed approximate reasoning not only as an inference mode and to manage imprecise knowledge, but also to refine learning and validate the rule base.
The proposed method was implemented in a tool called SUCRAGE and confronted with a real application in the field of image processing (multi-components image segmentation). The obtained results are very satisfactory. They validate our approach and allow us to consider other application fields.