Artificial Intelligence Research Group (AIRG)
Topics - Spring 2005

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

Dates and topics for this semester are as follows:

Jan 20

Mingyu Feng
Progress report on reporting for Assistment project

Jan 27

The AIRG Seminar will join the Milestones in Computer Science Distinguished Lecture Series today

Feb 3

Video of Judea Pearl's talk at CMU on Nov. 12, 2004 available at: "Reasoning with Cause and Effect" (part 1 of 2) http://wean1.ulib.org/cgi-bin/meta-vid.pl?target=Lectures/Miscellaneous/Judea%20Pearl

Feb 10

Seminar cancelled due to snow storm

Feb 17

No meeting this day - Academic Advising Day

Feb 24

Video of Judea Pearl's talk at CMU on Nov. 12, 2004 available at: "Reasoning with Cause and Effect" (part 2 of 2) http://wean1.ulib.org/cgi-bin/meta-vid.pl?target=Lectures/Miscellaneous/Judea%20Pearl

March 3

Ed Nelson
Ogre Expert: An Expert System to Play Ogre

This talk will describe an expert system that plays the game Ogre, by Steve Jackson Games. This toy problem is interesting for these reasons: 1) The game Ogre models two opposing forces assumed to be of equal strength but composed of different units. 2) Spatial information (where units and terrain features are on the board) has to be integrated with text data (unit abilities and current Ogre status) in order to play a competent game. 3) Ogre loosely simulates battlefield conditions. If expert systems work well on this type of problem, the way is opened toward using expert systems in increasingly realistic simulations. The question that this research attempts to answer is whether or not an expert system is a good approach to use to "solve" this kind of problem.

March 10

No meeting this day - Term break

March 17

The AIRG seminar will join the IMGD talk scheduled for this week

Tuesday March 22, 1-2 pm. FL 141

Aparna Varde
"Learnmet: Learning a Domain-Specific Distance Metric for Graph Mining"

Experimental results in scientific domains such as Materials Science and Mechanical Engineering are often plotted as graphs to aid visual analysis of the corresponding processes for decision support. These graphs represent the behavior of process parameters thereby incorporating semantics. Performing a laboratory experiment consumes significant time and resources. This motivates the need to estimate the graph that would be obtained in an experiment given its input conditions, and vice versa. The estimation helps in decision support in the corresponding industry. In order to achieve accurate estimation it is important to capture the domain semantics in graphs. It is seldom known apriori which distance metric best preserves semantics. This motivates the need for learning such a metric. State-of-the-art learning techniques are either inapplicable or inaccurate in our context. We propose a technique called LearnMet to learn a domain-specific distance metric for graphs. LearnMet iteratively clusters graphs using a guessed initial metric and compares the obtained clusters with correct clusters given by domain experts. The metric is refined in each iteration based on the difference between the obtained and correct clusters until they match each other. The metric corresponding to the matching clusters best preserves domain semantics and is the learned distance metric. LearnMet is evaluated by comparing the estimation using the learned metric with the estimation using the default notion of distance in graphs. The evaluation is being done primarily in the domain "Heat Treating of Materials" that motivated this research. The learned metric enhances estimation accuracy as elaborated in the evaluation.

March 24

Janet E. Burge
Software Engineering Using design RATionale (SEURAT)

One of the many difficulties encountered while performing software maintenance is determining the impact of potential changes on what already exists. The problem gets even worse if the original software developers are not available. One way to address this difficulty is to give the maintainers access to the Design Rationale (DR) of the original system. This rationale would provide the intent behind the design and implementation decisions as well as a history of design alternatives that have been considered. Unfortunately, this information is difficult and time consuming to capture and therefore is rarely available. Our approach to this problem is to look at how the rationale could be used. Rationale needs to be useful to provide incentive for its initial capture. To investigate rationale uses, we have developed a system called SEURAT (Software Engineering Using RATionale) which integrates with a software development environment. SEURAT goes beyond mere presentation of rationale by inferencing over it to check for completeness and consistency in the reasoning used while a software system is being developed and maintained. We performed an experiment that evaluated this system by having it assist with three different types of maintenance tasks. This evaluation showed that using SEURAT decreased the time needed to perform maintenance tasks, especially for less experienced developers.

March 17

Jason A. Walonoski
Comparing Mathfacts and Learning Rate in the Assistments Project

Research in the AI Lab has yielded some preliminary findings on student learning rates in the Assistments project. One hypothesis that we wished to explore was the possibility of there being a correlation being mathfacts (knowing how to add, subtract, multiply, and divide) and the ability to learn higher-level math concepts (MCAS items) using Learning Opportunity Pairs (LOPS) and mathfact performances. I will cover our hypothesis, our experiment, our (inconclusive) results, and cover some possible explanations of our results. Time allowing, I will discuss future work.

April 7

Video Keynote Talk - KDD2004 Seattle, Aug. 22 2004
Dr. David Heckerman. Machine Learning and Applied Statistics Group. Microsoft Research
Graphical Models for Data Mining