One of the most important aspects when we are evaluating a machine learning system is its capability for synthesizing knowledge available in input data. Also it is to have the right proce-dures for measuring and comparing the levels of absorption of such knowledge. Another point to take in consideration is the format this knowledge is showed to us for being applied. However, synthesizing capabilities of a system are often reduced by the way input data is presented to the algorithm. Here is shown a method that tries to decrease this factor through reusing less-successful previously generated classifiers. This method, as seen in experimental tests, is efficient in most of the examples tested. We also analyze advantages and problems encountered when using this type of methods.