David J. Finton

Email: my last name, at cs dot wisc dot edu

Current Projects:

I'm a machine learning technologist for Boeing Phantom Works. My current projects include applied research in data mining and the validation of machine learning algorithms, as well as ongoing research in reinforcement learning. I continue to look for better ways of representing the values of actions in control tasks, and better ways of learning good features as the task itself is being learned.

Previous lives

I was a postdoc in the Laboratory for Molecular and Computational Genomics, as part of the Genomic Sciences Training Program. My work there included a hidden Markov algorithm that separated yeast haplotypes on the basis of the likelihood of observing their alignment features. I'm thinking of ways of generalizing this technique for pattern classification problems. I also proposed a state-space analysis of genomic change; I never had the opportunity to begin this work, but hope one day to revisit these ideas.

Other Research Interests

I want to understand the nature of intelligent behavior, how it arises in brains, and how it can be implemented in machines. In a nutshell, I'm interested in feature extraction, classification, and control, not only for static collections of data, but data-sets that change with time. The tricky part is recognizing which features are relevant to our task, and which may be ignored; we want to represent the world "as simply as possible--but no simpler." Here are some examples of my past work touching on cognitive economy (how we can make learning easier by adopting simpler and more effective representations) and reinforcement learning.

My dissertation on Cognitive Economy and the Role of Representation in On-Line Learning addresses the question `How can an intelligent agent learn to categorize its perceptions according to its goals.' The abstract page contains the abstract and has links to the Introduction and to an Afterword giving the big picture. The full dissertation is also available.

The dissertation is based on the idea that we can simplify the state-space for a problem by learning to filter out irrelevant information. It applied the psychological principle of cognitive economy to the domain of reinforcement learning to develop objective criteria for representational adequacy. Then it proved that these criteria allow tasks to be learned, and presented case studies of a system based on these ideas that successfully learned its representation along with its task. This work grounds feature relevance on principled criteria regarding the rewards associated with particular actions.
My recent paper When Do Differences Matter? On-Line Feature Extraction Through Cognitive Economy describes how cognitive economy may be worked out in a reinforcement learning system, and presents experimental studies based on the puck-on-a-hill task. Contributions of this paper include a description of how the idea of cognitive economy addresses the open problem of knowing how accurately we need to learn the action values, and a new methodology for testing and camparing representations by means of learning curves. [Paper now accepted for publication by Cognitive Systems Research, in press]

Last modified: March 12, 2006