General and Efficient Cognitive Model Discovery Using a Simulated Student
In order to better understand how humans acquire knowledge, one of the essential goals in cognitive science is to build a cognitive model of human learning. Moreover, a cognitive model that better matches student behavior will often yield better instruction in intelligent tutoring systems. However, manual construction of such cognitive models is time consuming, and requires domain expertise. Further, manually-constructed models may still miss distinctions in learning which are important for instruction. Our prior work proposed an approach that finds cognitive models using a state-of-the-art learning agent, SimStudent, and we demonstrated that, for algebra learning, the agent can find a better cognitive model than human experts. To ensure the generality of that proposed approach, we further apply it to three domains: algebra, stoichiometry, and fraction addition. To evaluate the quality of the cognitive models discovered, we measured how well the cognitive models fit to student learning curve data. In two of those domains, SimStudent directly discovers a cognitive model that predicts human student behavior better than the human-generated model. In fraction addition, SimStudent supported discovery of a better cognitive model in combination with another automated cognitive model discovery method.