FGP: Using Statistics to Drive an Expert Database,
Abstract
We describe the FPG machine, which uses similarity-based retrieval and simulated speculation to convert pools of data directly into quasi-expert advice. The central operation is the retrieval of a small set of records similar to a partially-instantiated new record. The system uses two statistical techniques to improve on the standard Euclidean measure for calculating distance between two records represented as a vector of features. One is a facility to automatically weight the importance of features which will add or subtract to those features' contribution to the overall distance score. The other is a means for separating the most relevant records from the rest by finding a natural break in the ordering of the records by distance from the input. We explain the role these techniques play in the overall operation of the system in the next section; the algorithms used for the calculations are described in the appendices.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1992
- Accession Number
- ADP007189
Entities
People
- Scott Fertig
Organizations
- Yale University