Approximately Optimal Teaching of Approximately Optimal Learners

We propose a method of generating teaching policies for use in intelligent tutoring systems (ITS) for concept learning tasks [37], e.g., teaching students the meanings of words by showing images that exemplify their meanings à la Rosetta Stone [30] and Duo Lingo [13]. The approach is grounded in control theory and capitalizes on recent work by [28], [29] that frames the “teaching” problem as that of finding approximately optimal teaching policies for approximately optimal learners (AOTAOL). Our work expands on [28], [29] in several ways: (1) We develop a novel student model in which the teacher’s actions can partially eliminate hypotheses about the curriculum; (2) With our student model, inference can be conducted analytically rather than numerically, thus allowing computationally efficient planning to optimize learning; and (3) We develop a reinforcement learning-based hierarchical control technique that allows the teaching policy to search through deeper learning trajectories. We demonstrate our approach in a novel ITS for foreign language learning similar to Rosetta Stone and show that the automatically generated AOTAOL teaching policy performs favorably compared to two hand-crafted teaching policies.