In number recognition, one of the challenges is to deal with the high dimensionality of data that affects the performance of algorithms. On the other hand, pattern recognition allows establishing fundamental properties among sets of objects. In this context, Rough Set Theory applies the concept of super-reducts in order to find subsets of attributes that preserve the capability of the entire set to distinguish objects that belong to different classes. Nevertheless, finding these reducts for large data sets has exponential complexity due to the number of objects per class and attributes per object. This paper proposes a new approach for dealing with this complex problem in real data sets to obtain a close enough to a minimal discriminator. It takes advantage of the theoretical background of Rough Set Theory, especially considering those super-reducts of minimal length. In literature, there is an algorithm for finding these minimal length reducts. It performs well for a small sampling of objects per class of the entire data set. An evolutionary algorithm is performed to extend it over a huge data set, taking a subset of the entire list of super-reducts as the initial population. The proposed discriminator is evaluated and compared against state-of-the-art algorithms and data set declared performance for different models.