Additive Belief-Network Models

From Tetherless World Wiki

Jump to: navigation, search

Citation: Paul Dagum and Adam Galper. (1993) Additive Belief-Network Models. In KSL-93-01, 1993.

Publication techreport ( Edit )
type Technical Report
bibtype techreport
Bibtex basics
author Paul Dagum and Adam Galper
title Additive Belief-Network Models
number KSL-93-01
institution Knowledge Systems, AI Laboratory
address Washington, D.C
year 1993
Bibtex more
Access Paper
abstract The intractability of available probabilistic inference algorithms hinders belief network applications to large domains. Researchers have shown that both exact and approximate probabilistic inference is NP-hard, and therefore, we do not hope to find tractable solutions to inference in large applications. The intractability of inference, known implicitly to designers of large applications, and the formal proofs of its complexity that came afterwards, together motivated alternative research directions in hopes of tractable solutions to the impasse. From this work arose, for example, noisy OR-gates used in QMR-DT and probabilistic similarity networks.Motivated by recent developments in belief network models for time-series analysis and forecasting, we define "additive belief network models" (ABNM). We (1) discuss the nature and implications of the approximations made by an additive decomposition of a belief network, (2) prove greater efficiency in the induction of additive models when available data is scarce, (3) generalize the Lauritzen-Spiegelhalter inference algorithm to exploit the additive decomposition of ABNMs (4) prove greater efficiency of inference, and (5) present implementation results on induction and on inference of belief networks.

KSL Technical Report ID: KSL-93-01
Facts about Additive Belief-Network ModelsRDF feed
Abstract The intractability of available probabilis … The intractability of available probabilistic inference algorithms hinders belief network applications to large domains. Researchers have shown that both exact and approximate probabilistic inference is NP-hard, and therefore, we do not hope to find tractable solutions to inference in large applications. The intractability of inference, known implicitly to designers of large applications, and the formal proofs of its complexity that came afterwards, together motivated alternative research directions in hopes of tractable solutions to the impasse. From this work arose, for example, noisy OR-gates used in QMR-DT and probabilistic similarity networks.Motivated by recent developments in belief network models for time-series analysis and forecasting, we define "additive belief network models" (ABNM). We (1) discuss the nature and implications of the approximations made by an additive decomposition of a belief network, (2) prove greater efficiency in the induction of additive models when available data is scarce, (3) generalize the Lauritzen-Spiegelhalter inference algorithm to exploit the additive decomposition of ABNMs (4) prove greater efficiency of inference, and (5) present implementation results on induction and on inference of belief networks. ction and on inference of belief networks.
Address Washington, D.C  +
Author Paul Dagum and Adam Galper  +
Bibtype techreport  +
Has author Paul Dagum and Adam Galper  +
Has identifier KSL-93-01  +
Has publishing details 1993  +
Has title Additive Belief-Network Models  +
Has where published KSL-93-01  +
Has year 1993  +
Institution Knowledge Systems, AI Laboratory  +
Ksl tr id KSL-93-01  +
Number KSL-93-01  +
Process note YES  +
Title Additive Belief-Network Models  +
Year 1993  +
Retrieved from "http://tw.rpi.edu/wiki/Additive_Belief-Network_Models"
Personal tools