This paper aims at building autonomous controllers for swarm robots, specifically aimed at enforcing a given shape formation, here a column formation. The proposed approach features two main characteristics. Firstly, a state-of-the-art evolutionary setting is used to achieve the on-board optimization of the controller, avoiding any simulator-based approach. Secondly, as the cost of physical experiments might be prohibitively high for plain evolutionary approaches, a data mining approach is achieved on the top of evolution; rule discovery is used to discover the most promising regions in the controller search space. The merits of the approach are experimentally validated using a 5 robot formation, showing that the hybrid evolutionary learning process outperforms evolution alone in terms of swarm speed and shape quality.