To select an adequate coding is one of the main problems in applications based on Evolutionary Algorithms. Many codings have been proposed to represent the search space for obtaining decision rules. A suitable representation of the individuals of the genetic population can reduce the search space, so that the learning process is accelerated by decreasing the number of necessary generations to complete the task. In this sense, natural coding achieves such reduction and improves the results obtained by other codings. This paper justifies the use of natural coding by comparing it with hybrid coding that joins well-known binary and real representations. We have tested both codings on a heterogeneous subset of databases from the UCI Machine Learning Repository. The experiments' results show that natural coding improves the quality of the obtained knowledge-model using only one third of the generations that hybrid coding needs as well as a smaller population. © Springer-Verlag Berlin Heidelberg 2003.