AIRG Topics - Fall 2008

Our group meets on Thursdays at 11:00 a.m., FL 246, Beckett Conf. Room.

Dates and topics for this semester are as follows:

Sept 4

AIRG Organizational Meeting (Coordinator: tbd)

Sept 11

tbd

Sept 18

tbd

Sept 25

tbd

Oct 2

tbd

Oct 9

tbd

Oct 16

tbd

Oct 23

Ugrad break - No meeting

Oct 30

Jason Wilson

"Towards an Affective Embodied Conversational Agent"

Emotions play a daily role in human conversations and interactions. We are able to recognize others' emotional state through audio and visual cues and use this information while we interact. Embodied Conversational Agents are computer agents that incorporate a variety of non-verbal communication mechanisms. An Affective Embodied Conversational Agent addresses how emotions play a role in communicating and reasoning. This talk reviews some of the recent work in affect recognition, decision-making, and presentation and discusses examples of recent affective agents performing in each of these domains. It concludes with some insight on the direction this research is and could be leading.

Nov 6

Zachary Pardos

"Effective Skill Assessment Using Expectation Maximization in a Multi Network Temporal Bayesian Network"

We propose a temporal independent skill model for effective assessment and prediction of student skills that is more accurate than multi mapped conjunctive models while requiring only a fraction of the computational resources to run. Using the Expectation Maximization algorithm we define steps for effective learning of model parameters over 150,000 responses of real student data that reveal important skill knowledge and learning trends. This skill report is exhibited in the paper. Lastly we focus on how to harness the power of the learned parameters to accurately predict an end of year standardized state math test. Our results of prediction using the independent skill network model beat out previous best prediction errors using a multi mapped conjunctive model. We believe these results could encourage wider use of these machine learning techniques that can now be effectively run on standard computing machines such as PCs found in school computer labs.

Nov 13

No AIRG presentation.

Group members are encouraged to attend the IMGD Seminar by Tom Hunter at 11:00 am in Salisbury Labs, Room 406, entitled:

"Evidence Based Game Design"

Glymetrix designs games that front end software that helps people manage diabetes. To do this effectively we base our designs on well articulated theory. Theories of learning styles, of self efficacy, and of needs all underscore the design of the games. Medical evidence is also a critical part of good design when a game is part of a therapeutic application. Research into diabetes care, the effect of gambling on the brain, nutrition, diet and exercise all influence design. This talk will look at Glymetrix general approach to theory and evidence, and go in depth on one or two examples.

Nov 20

Michael Sao Pedro

"Towards Predicting, Scaffolding, and Assessing Middle School Students' Scientific Inquiry Skills: Teaching the Control of Variables Strategy"

Scientific inquiry is highly regarded in our national educational standards and is viewed as critical to science education reform efforts. However, in part due to the difficulties in implementing and assessing it in practice, inquiry is typically substituted with rote learning. Additionally, research on inquiry learning and assessment is hindered by several factors including the lack of available assessments and the difficulty in measuring these skills due to their complexity. To address these problems, we will extend the Assistment system to support real-time assisting and assessing of inquiry skills using open-ended, interactive microworlds.
     In this talk, I will present our ongoing efforts towards this endeavor, emphasizing our study currently in progress at a Fitchburg, MA middle school involving over 170 students. This study, an extension of Klahr and Nigam (2004), focuses on students' acquisition of a cornerstone skill in scientific inquiry, the control of variables strategy (CVS). We compare of the effects of three pedagogical approaches, direct instruction with reification, direct instruction, and discovery learning, on learning this strategy. We also survey students' epistemologies of models as well as their own self-assessment of their perseverance to determine if these interact with learning gains across the three conditions. Finally, I will discuss some of our software design decisions and ideas for supporting more complex inquiry activities.

Nov 27

Thanksgiving Break - No Meeting

Dec 4

Mingyu Feng

"Towards building a better cognitive model" (Ph.D. progress talk)

A Cognitive Model models student knowledge and is a key component of intelligent tutors. Much work has been done on constructing and improving cognitive models. In our previous work, we presented our effort to build a fine grained model, and evaluated the model comparing it to coarse grained models. Generally speaking, creating an accurate model of student knowledge is hard due to various sources of uncertainty. The first model made, such as ours, is usually just a best guess and should be refined. We are engaged in an effort to refine the current model. Our subject matter expert is improving the model based on her domain knowledge, and the model-fit information provided by our models could allow this process to be data-driven. Additionally, we plan to apply and extend the approach of learning factor analysis to improve the existing model, and will use educational data mining techniques (learning decomposition) to detect inappropriate problem-taggings in the current model.

Dec 11

Leena Razzaq

"Is scaffolding worth the extra time?"

Razzaq and Heffernan (2006) showed that scaffolding compared to hints on demand in an intelligent tutoring system could lead to higher averages on a middle school mathematics post-test. In Razzaq, Heffernan & Lindeman (2007), we reported on an experiment that examined the effect of math proficiency and the level of tutor-student interaction on learning. We found an interesting interaction between the level of interaction and math proficiency where less-proficient students benefited from more tutor-student interaction and more-proficient students benefited from less interaction. However, these two studies suffered from a "time on task" confound. The conditions with less tutor-student interaction take significantly less time than those that require more interaction. In this talk, I will discuss how we will control for time in a new experiment to examine whether it is worth the extra time that scaffolding takes. I will also share some preliminary results.

Dec 18

End of Semester - No Meeting