AAAI Publications, The Twenty-Seventh International Flairs Conference

Font Size: 
A Novel Methodology for Processing Probabilistic Knowledge Bases Under Maximum Entropy
Gabriele Kern-Isberner, Marco Wilhelm, Christoph Beierle

Last modified: 2014-05-03

Abstract


Probabilistic reasoning under the so-called principle of maximum entropy is a viable and convenient alternative to Bayesian networks, relieving the user from providing complete (local) probabilistic information and observing rigorous conditional independence assumptions. In this paper, we present a novel approach to performing computational MaxEnt reasoning that makes use of symbolic computations instead of graph-based techniques. Given a probabilistic knowledge base, we encode the MaxEnt optimization problem into a system of polynomial equations, and then apply Gröbner basis theory to find MaxEnt inferences as solutions to the polynomials. We illustrate our approach with an example of a knowledge base that represents findings on fraud detection in enterprises.


Full Text: PDF