This study takes place in the context of multi-agent systems (MAS), and especially reactive ones. In such a system, interactions are essential, and trigger a collective behaviour that is not directly linked to the individual ones. Whereas the evolution of the system is unknown if not tried, the regularity of emergent structures in the system is observable and forms a global behaviour. In this paper, we propose to control the global behaviour of a MAS thanks to reinforcement learning tools applied at its global level. We also highlight the choice of the features taken into account to achieve this control, that is the information considered to decide which action to perform.