Building Cognitive Model for Tutoring by Demonstration

Speaker
Noboru Matsuda
Postdoctoral Fellow, Human-Computer Interaction Institute, Carnegie Mellon University

When
-

Where
Newell-Simon Hall 1305 (Michael Mauldin Auditorium)

Description

The aim of this work is to extend an intelligent authoring environment for Cognitive Tutors. It is well know that Cognitive Tutors are very effective tools to facilitate student’s learning. However, the tutor’s effectiveness depends on the quality of a cognitive model representing domain principles (i.e., skills to be learned). Writing a cognitive model for a real-world tutor requires cognitive task analysis, programming, and testing. This can take a significant amount of effort—many days, weeks, or months depending on task complexity, even for a well-trained cognitive scientist. For novice (untrained) authors, these tasks often present an insurmountable barrier. To help novice authors in building Cognitive Tutors, we have developed a machine learning agent—the Simulated Student—that automatically generates a cognitive model from sample solutions demonstrated by the author. We have successfully integrated the Simulated Student into the Cognitive Tutor Authoring Tools (CTAT), a powerful authoring environment for Cognitive Tutors. Using the Simulated Student as a building block of CTAT, we are exploring whether this cutting-edge technology can indeed help authors to build a fully functional Cognitive Tutor with a fully equipped cognitive model as a generalization of examples demonstrated. In this talk, I will discuss the overall framework of the Simulated Student and explain how a Cognitive Tutor can be authored by demonstration. I will then provide some results from the studies on effectiveness and generality of the Simulated Student. The major findings include (1) that the order of training problems does not affect the quality of the cognitive model, (2) that ambiguities in the interpretation of demonstrations might hinder machine learning, and (3) that more detailed demonstrations can both avoid difficulties with ambiguity and prevent search complexity from becoming impractical.

Speaker's Bio

Dr. Noboru Matsuda is a post doctoral fellow at Human-Computer Interaction Institute, Carnegie Mellon University. He received his Ph.D in Intelligent Systems from the University of Pittsburgh in 2004. Prior to that he was a visiting scholar at the Learning Research and Development Center at Pitt from University of Electro-Communications in Tokyo, Japan. His research interests are at the intersection of artificial intelligence and cognitive theories of learning and teaching. Since completing a master’s degree in education in 1988, he has been studying intelligent tutoring systems in a wide variety of domains including mathematics, programming languages, and physics. As a member of NSF-funded Pittsburgh Science of Learning Center (PSLC), his current research interest is in a computational model of human learning and application of it to explore how people learn.

Speaker's Website
http://www.cs.cmu.edu/~mazda/