Complex motor learning
If research in artificial intelligence has shown anything, it is that the physical world behaves little like the conveniently symbolic simulations that make things tractable. Many of us have taken the gap between simulation and reality as a provocation to interesting research. My target in this attack is a control architecture for the micromechanical flying insect (MFI) that builds on the work of Andrew Ng, David Andre, and others. I intend to extend their work to support continuous state/action spaces, dynamic controllers, and complex (potentially hierarchical) control programs that combine multiple objectives.
Recommender systems
The recommendation problem is to predict, given a user's history of expressed preferences and the preference histories of other users, the relevance of unseen items to that user. It emerges not only in online commerce applications, where retailers recommend items to consumers based on such data as purchase histories, but in "information marketplaces" like Usenet, where users confront far more postings than they could ever hope (or care) to read.
In the past, I've explored the value of time order for improving the accuracy of recommendations. Anecdotally, this approach tends to expose structural aspects of the domain (such as the link structure of a Web site or the time order of a television schedule) as much as temporal features of user behavior. My current work in this field considers the combination of spatial information (for example, the location of a mobile-phone user relative to several restaurants) with user preference histories for location-based services. Stay tuned ...
Relevant Papers
Zimdars, A. L., Chickering, D. M., and Meek, C. (2001). Using temporal data for making recommendations. In Breese, J., and Koller, D. (eds.), Uncertainty in Artificial Intelligence: Proceedings of the Seventeenth Conference. San Francisco: Morgan Kaufmann Publishers.
