This paper presents the principles and the architecture of PROTO-TEG, a self-improving tutor in geometry. This system is able to discover criteria useful for selecting the didactic strategies it has at its disposal. These criteria are expressed as characteristics of the student model. They are elaborated by comparing student model states recorded when a strategy was effective and those recorded when the same strategy was not effective. This comparison is performed by machine learning methods, or, more precisely, by learning concepts from examples. An empirical experiment was performed in order to assess the self-improving functions; conditions were discovered for five of the nine didactic strategies. However, this new knowledge did not lead to PROTO-TEG being more efficient in terms of student performance.