Evolution of adaptive discretization intervals for rule-based genetic learning system

The traditional classi er rules evolved in genetic based machine learning (GBML) systems need a discretization process to handle problems with real-valued attributes. A good discretization procedure is needed to generate a solution with good accuracy because the alternative of a high number of simple uniform-width intervals is bad due to the big search space being hardly explorable in a reasonable time. There exist some good discretization algorithms, like the Fayyad & Irani method [Fayyad and Irani, 1993], but they fail in some problems. The work being summarized here deals with a rule representation with adaptive discrete intervals which split or merge through the evolution process, nding a robust and correct discretization intervals as the learning process is done with a reasonable computational cost.