Spectrum Curricula: Design and Initial Results

Part of human adaptability is our ability to learn from instruction by other humans. Such instruction is typically an informal mix of demonstration and telling, presented by a situational expert who is not a teaching expert. If machines are to cooperate effectively with humans across a broad range of situations, they must exhibit human-like learning flexibility, even in the face of poor or ambiguous instruction. A key challenge, however, is how to effectively measure and compare the teachability of machine learners. To address this challenge, we have developed the instrument of a spectrum curriculum(Beal, Leung, and Laddaga 2010): a suite of lessons, incrementally varying along a dimension of interest and presented in order from hardest to easiest. The adaptivity of a student is then expected to be characterized by its performance curve across the suite of lessons, with more adaptive students expected to show a smooth increase of comprehension as the quality of teaching improves. We have developed seven spectrum curricula for simulated RoboCup, and preliminary tests of human-adapted versions of four of those curricula support this theory.