IDEA Lab Research Projects

Members - Faculty, students, and collaborators
Research - Details about our current research projects
Theses and Dissertations - Publications and code releases for student theses and disserations
Publications - Recent technical papers and presentations
Software - Recent software releases

Developing Spatiotemporal Relational Models for Hazardous Weather Prediction

3D image of simulated storm Three dimensional view of a simulated supercell thunderstorm. Strong updraft regions are shown in red, strong downdrafts in blue and strong vorticity in green and yellow.

Severe weather events such as floods, tornados, lightning, and hail result in loss of life, property, and in disruptions to transportation systems. The annual economic damage from these mesoscale events is estimated to be greater than $13B a year (Pielke and Carbone 2002). In collaboration with the Center for Adaptive Sensing of the Atmosphere (CASA) Engineering Research Center (ERC) and the Linked Environments for Atmospheric Discovery (LEAD) Information Technology Research (ITR), we are improving the prediction of severe weather phenomena through the development of new spatiotemporal relational knowledge discovery techniques for use on mesoscale meteorological data with a specific focus on tornadoes and drought.

Knowledge discovery methods focus on making sense of data by identifying salient patterns. These patterns will enable meteorologists to improve their understanding of the causes of severe weather events. Statistical models will be used to predict the arrival of mesoscale weather events based on the current and past weather readings. The models are being developed with human-understanding in mind. A key challenge is developing new techniques for modeling spatiotemporal relational data, as current techniques generally ignore the temporal aspects of the data. However, temporal relationships are critical for weather data. In this research, we are specifically developing novel dynamic, or temporal, relational models.


Relational reinforcement learning and autonomous discovery of abstractions

Reinforcement learning techniques are a perfect fit for many real-world learning tasks, where it is possible to specify the goal but very difficult to define how the goal should be accomplished. For example, no computer Go program has been able to best the top human players. Current RL techniques are often limited in scope due to knowledge representations. The real world is noisy, dynamic, and relational and we are developed a novel representation that can handle each of these issues and applying it to the task of computer Go.

We have introduced Relational U-tree, a relational adaptation of McCallum's U-tree model and are currently expanding it to multi-modal representations. This approach autnomously identifies salient abstractions from a dyanamic relational data set. Relational U-tree is constructed incrementally, which means that it is constructed online and that the tree can be dynamically rearranged based on new evidence.


Learning the structure of student retention data using Bayesian Networks

Composite Bayesian Network identified in retention data collected from students in the College of Engineering at the University of Oklahoma.


Student retention in STEM (Science, Technology, Engineering and Mathematics) fields is of critical national importance. In conjunction with researchers at the University of Oklahoma's Research Institute for STEM Education, we are applying machine learning techniques to authentic data sets collected from students in the College of Engineering to improve our understanding of why some students persist while others leave engineering and computer science majors.



Created by amcgovern [at]

Last modified June 19, 2013 1:59 PM