Integrating perceptual representation learning and skill learning in a simulated student

One of the fundamental goals of artificial intelligence is to understand and develop intelligent agents that simulate human-level intelligence. This fundamental goal complements another essential goal in education, improving understanding of how humans acquire knowledge and how students may vary in their abilities to learn. Contributing to both goals, a lot of efforts have been made to develop intelligent agents that simulate human learning of math and science. However, constructing such a learning agent currently requires manual encoding of prior domain knowledge, which is both inefficient and less cognitively plausible. Previous cognitive science research has shown that one of the key factors that differentiates experts and novices is their different representations of knowledge. Moreover, for many existing learning algorithms, “better” representations often lead to more effective learning. We [1] recently proposed an efficient algorithm that acquires representation knowledge in the form of “deep features”. In this paper, we integrate this algorithm into a simulated student, SimStudent, which learns procedural knowledge from example solutions and problem solving experience. We show that with the integration, prior knowledge engineering effort is reduced, learning performance is as good or better, and SimStudent becomes a more plausible simulation of human learning.