Supervised Machine Learning by Generation of Rules : Optimization of the classification rate by mixed correlation

We present a method of supervised learning by automatic generation of production rules. This learning method through examples is at the junction of the statistical methods and those based on techniques of Artificial Intelligence. Classification rules are automatically generating rules and show the membership of an object to a class. However, this membership is attached with uncertainty; each conclusion is accompanied by a belief degree. Our approach is polythetic, or multi-feature, insofar as the features which appear in the premises of the rules are selected in block, and not in a successive way. The idea is to locate privileged correlations between the features describing the examples and to gather them by taking account of these correlations. In this article, we propose a new form of correlation search "the Method of Mixed Search "; the originality of this method is to integrate the discriminating capability of classes membership. the search for correlations is based on the union of the classes. Experimental tests were carried out. The principal results obtained as well as their comparison with two other methods of correlations search are presented. These results are satisfactory and allow us to extend the validation of our approach to other data.