In this paper, we present a general method for optimizing a tutoring system with a target application in the domain of second language acquisition. More specifically, the optimisation process aims at learning the best sequencing strategy for switching between teaching and evaluation sessions so as to maximise the increase of knowledge of the learner in an adapted manner. The most important feature of the proposed method is that it is able to learn an optimal strategy from a fixed set of data, collected with a hand-crafted strategy. This way, no model (neither cognitive nor probabilistic) of learners is required but only observations of their behavior when interacting with a simple (non-optimal) system. To do so, a particular batch-mode approximate dynamic programming algorithm is used, namely the Least Square Policy Iteration algorithm. Experiments on simulated data provide promising results.

Index Terms. Tutoring systems, approximate dynamic programming