My thesis involved developing a baseball player prediction system using ensemble learning. My focus was on predicting the season totals for six key offensive statistics for players who have played in at least 162 games prior to the season being predicted. I produced several different prediction systems using model trees, neural networks, and support vector machines as well as using the ensemble learning techniques bagging, boosting, and stacking. All prediction systems were developed in Java using the Weka package.
I am currently working in conjunction with USDA Forest Service on the NED-2 project, which is intended to aid resource managers develop goals, assess current and future conditions, and produce sustainable management plans for forest properties. Specifically, NED-2 is a robust, intelligent, goal-driven decision support system that integrates vegetation growth models, wildlife models, silvicultural models, GIS, and visualization tools for forest ecosystem management. NED-2 uses a blackborad architecture and a set of semi-autonomous agents to manage these tools for the user.
My tasks have included upgrading the software to integrate with the newest versions of the ESRI ArcView software. I updated the previous code for integration with older versions of ArcView and maintained backwards compatability with the previous versions. I am currently in charge of maintaining the regeneration agent which uses various models to determine how a forest regenerates over time. As well I have have implemented a User File Regeneration model which allows users to input specific regeneration results based on observed results which are triggered by user defined preconditions.
I have developed an interface between LPA-Prolog and the Microsoft Speech API.
I have developed a spam filter that uses a combination of Naive Bayes Classifiers and Evolutionary Computation.
I have performed a comparison of machine learning methods for resource allocation games. The set of resources allocation games I am focusing on has time dependent resource capacities, no communication, and user preference between resources.
I have implemented a real time personalized news webpage which uses a modified version of Naive Bayes Classifiers to learn what types of articles a user finds interesting.