Published in Proceedings from the 4th Bayesian Modelling Applications Workshop at UAI'06, July 1, 2006.
Copyright © 2006 by Krol Kevin Mathias, Cynthia Isenhour, Alex Dekhtyar, Judy Goldsmith, Beth Goldstein. The definitive version is available at http://dx.doi.org/10.1.1.113.3875.
NOTE: At the time of publication, the author Alex Dekhtyar was not yet affiliated with Cal Poly.
The project described in this paper originated with an observation by the AI group at the University of Kentucky, that, individually, stochastic planning and constraint satisfaction are well-studied topics that resulted in efficient software, but stochastic planning in the presence of constraints on the domains and actions is an open area of investigation.
We were interested in an advising scenario, and chose the US social welfare system, a.k.a. “Welfare to Work” as our test domain. This required computer scientists to learn more than expected about social science as well as the local welfare system. This paper discusses the discipline specific assumptions we brought to this project, and how they served as impediments to research. We also show how the different perspectives have sparked new ideas in knowledge elicitation.