In the RuleSEEKER system a new conception of knowledge extraction from data has been developed. The characteristic feature of this approach was the multiple (i.e. by means of different machine learning tools) analysis of a primary source of knowledge (e.g. decision table), which supplied multiple learning models, called here secondary sources of knowledge (decision rules).

       Two different methods is used in the system. The first one relies on a separate optimization of each developed secondary source. The second method is distinctly different. All developed secondary sources are merged together and then, the entire joined (large) model, is optimized using the set of generic operations. Improved models for both methods, are then evaluated – via testing the classification accuracy. It might be stated that the optimization of learning models, using generic operations, yielded quite interesting and satisfactory results. Namely, the error rate, number of rules decreased, and average value of rule parameters increased. The improvement of learning models will play a significant role in a case of very extended models, that contain very large set of rules.