Reward-punishment editing

In this work a novel editing technique is proposed. The basic idea of the algorithm is to reward patterns that contribute to a correct classification and to punish those that provide a wrong one. Reward-punishment is performed according to two criteria: the former operates at very local level while the latter analyses the training set at coarser scales in a multi-resolution fashion. A score is calculated for each pattern according to the two criteria and patterns whose score is lower than a predefined threshold are edited out. Experiments carried out on two difficult classification problems show the superiority of this method with respect to other well known approaches.