Large Scale Elicitation
Date/Time: Sunday 14:45-16:15
Sponsor: INFORMS Decision Analysis Society
Chair: Bruce D. Abramson
Chair Address: Cambridge Research Assoc., Inc., 1430 Spring Hill Rd., Ste. 200, McLean, VA 22102 ,
Some Elicitation Issues in the Design of a Normative System Bruce D. Abramson --- Cambridge Research Assoc., Inc., 1430 Spring Hill Rd., Ste. 200, McLean, VA 22102 ,
- Domain models that underlie intelligent systems are often orders of¨ magnitude larger than traditional decision models. This shift in¨ scale raises a variety of issues about the construction of both a¨ domain's structure and its probability model. New elicitation¨ techniques are necessary to address these issues.
An Elicitee's Perspective on Large-Scale Elicitation John Brown --- Ntl. Oceanic & Atmospheric Admin., 325 Broadway, Boulder, CO 80303 ,
- Hailfinder is a Bayesian system that combines meteorological data¨ and models with expert judgment to forecast severe summer weather¨ in NE Colorado. Hailfinder's design required the domain expert to¨ provide the system designers with both a complex relational¨ structure and several thousand probabilities.
Tools to Guide Elicitation in Large Belief Networks Jayant Kalagnanam --- Carnegie Mellon Univ., Dept. of Eng. & Public Policy, Pittsburgh, PA 15213-3890,
- The elicitation of conditional probability matrices of a belief¨ network is probably the most expensive aspect of building belief¨ networks. We present sensitivity analysis tools to help identify the¨ most influential numbers for the decisions at hand. These tools can¨ be used to guide the elicitation process for belief networks.
Bayesian Network Engineering Suzanne Mahoney, Jonathan Weiss, Anne Martin, Kathryn B. Laskey, Tod Levitt --- Info. Extraction & Transport. Inc., PO Box 808, East Setauket, NY 11733 ,
- Constructing large Bayesian networks is a system engineering task¨ for which supporting tools are vital. While decomposition makes the¨ network 'understandable' to elicitor and expert alike, identifying¨ commonalities among the variables and parameters is the key to¨ reducing the KE burden. Verification proceeds from local level tests¨ through systems level tests as the network is integrated.