<p>This thesis presents the results of the implementation and evaluation of two machine learning algorithms [Baz98, GB97]based on notions from Rough Set theory [Paw82]. Both algorithms were implemented and tested using the Weka [WF00]software framework. The main purpose for doing this was to investigate whether the experimental results obtained in [Baz98]could be reproduced, by implementing both algorithms in a framework that provided common functionalities needed by both. As a result of this thesis, a Rough Set framework accompanying the Weka system was designed and implemented, as well as three methods for discretization and three classi cation methods. </p><p>The results of the evaluation did not match those obtained by the original authors. On two standard benchmarking datasets also used previously in [Baz98](Breast Cancer and Lymphography), signi cant results indicating that one of the algorithms performed better than the other could not be established, using the Students t- test and a con dence limit of 95%. However, on two other datasets (Balance Scale and Zoo) differences could be established with more than 95% signi cance. The Dynamic Reduct Approach scored better on the Balance Scale dataset whilst the LEM2 Approach scored better on the Zoo dataset.</p>

This thesis presents the results of the implementation and evaluation of two machine learning algorithms [Baz98, GB97]based on notions from Rough Set theory [Paw82]. Both algorithms were implemented and tested using the Weka [WF00]software framework. The main purpose for doing this was to investigate whether the experimental results obtained in [Baz98]could be reproduced, by implementing both algorithms in a framework that provided common functionalities needed by both. As a result of this thesis, a Rough Set framework accompanying the Weka system was designed and implemented, as well as three methods for discretization and three classi cation methods. The results of the evaluation did not match those obtained by the original authors. On two standard benchmarking datasets also used previously in [Baz98](Breast Cancer and Lymphography), signi cant results indicating that one of the algorithms performed better than the other could not be established, using the Students t- test and a con dence limit of 95%. However, on two other datasets (Balance Scale and Zoo) differences could be established with more than 95% signi cance. The Dynamic Reduct Approach scored better on the Balance Scale dataset whilst the LEM2 Approach scored better on the Zoo dataset.