Personalization according to specific requirements of an individual student is one of the most important features in adaptive educational systems. Considering learning and how to improve a student's performance, these systems must know the way in which an individual student learns best. In this context, the current work outlines a new approach to automatically and dynamically discover students learning styles, considering its non-deterministic and non-stationary aspects, and taking into account that learning styles may change during the learning process in an unexpected and unpredictable way. Our approach is mainly based in genetic algorithms and reinforcement learning, and it has been tested through computer simulation of students. Promising results have been obtained through experiments. Some of them are presented in this paper.