FRINK Linked Data Fragments server

Matches in SemOpenAlex for { <https://semopenalex.org/work/W117871444> ?p ?o ?g. }

• W117871444 endingPage "805" @default.
• W117871444 startingPage "799" @default.
• W117871444 abstract "Importance sampling-based inference algorithms have shown excellent performance on reasoning tasks in Bayesian networks (Cheng & Druzdzel 2000; Moral & Salmeron 2003; Yuan & Druzdzel 2005). In this paper, we argue that all the improvements of these algorithms come from the same source, the improvement on the quality of the importance function. We also explain the requirements that a good importance function should satisfy, namely, it should concentrate its mass on the important parts of the target density and it should possess heavy tails. While the first requirement is subject of a theorem due to Rubinstein (1981), the second requirement is much less understood. We attempt to illustrate why heavy tails are desirable by studying the properties of importance sampling and examining a specific example. The study also leads to a theoretical insight into the desirability of heavy tails for importance sampling in the context of Bayesian networks, which provides a common theoretical basis for several successful heuristic methods. Introduction Importance sampling is used in several areas of modern statistics and econometrics to approximate unsolvable integrals. It has become the basis for several successful Monte Carlo sampling-based inference algorithms for Bayesian networks (Cheng & Druzdzel 2000; Moral & Salmeron 2003; Yuan & Druzdzel 2005), for which inference is known to be NP-hard (Cooper 1990; Dagum & Luby 1993). This paper argues that all the improvements of these algorithms come from the same source, the improvement on the quality of the importance function. A good importance function can lead importance sampling to yield excellent results in a reasonable time. It is well understood that we should focus on sampling in the areas where the value of the target function is relatively large (Rubinstein 1981). Thus, the importance function should concentrate its mass on the important parts of the target density. However, unimportant areas should by no means be neglected. Several authors pointed out that a good importance function should possess heavy tails (Geweke 1989; MacKay 1998). In other Copyright c © 2005, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. words, we should increase the sampling density in those unimportant areas. These two requirements seem to contradict one another. Why heavy tails are important and how heavy they should be are not well understood. This paper addresses these questions by studying the properties of importance sampling and discussing what conditions an admissible importance function should satisfy. We also try to illustrate why heavy tails are important by examining an example in which the conditions can be verified analytically. When analytical verification is impossible, we recommend to use two techniques to estimate how good an importance function is. The study also leads to a theoretical insight into the desirability of heavy tails for importance sampling in the context of Bayesian networks, which provides a common theoretical basis for several successful heuristic methods in Bayesian networks, including -cutoff (Cheng & Druzdzel 2000; Ortiz & Kaelbling 2000; Yuan & Druzdzel 2005), if-tempering (Yuan & Druzdzel 2004), rejection control (Liu 2001), and dynamic tuning (Shachter & Peot 1989; Ortiz & Kaelbling 2000; Moral & Salmeron 2003). Importance Sampling In our notation, a regular upper case letter, such as X , denotes a single variable, and x denotes its state. A bold upper case letter, such as X, denotes a set of variables. Their states are denoted by x. Now, let p(X) be a probability density of n variables X = (X1, ..., Xn) over domain Ω ⊂ R. Consider the problem of estimating the multiple integral" @default.
• W117871444 created "2016-06-24" @default.
• W117871444 creator A5005313203 @default.
• W117871444 creator A5066611124 @default.
• W117871444 date "2005-01-01" @default.
• W117871444 modified "2023-09-23" @default.
• W117871444 title "How Heavy Should the Tails Be" @default.
• W117871444 cites W1489119587 @default.
• W117871444 cites W1558800271 @default.
• W117871444 cites W1561981064 @default.
• W117871444 cites W1578411263 @default.
• W117871444 cites W1606274310 @default.
• W117871444 cites W1630331242 @default.
• W117871444 cites W1919989817 @default.
• W117871444 cites W1980452149 @default.
• W117871444 cites W1999432334 @default.
• W117871444 cites W2009169570 @default.
• W117871444 cites W2057565703 @default.
• W117871444 cites W2065412317 @default.
• W117871444 cites W2069739265 @default.
• W117871444 cites W2133722387 @default.
• W117871444 cites W2147632348 @default.
• W117871444 cites W2163698695 @default.
• W117871444 cites W2173497437 @default.
• W117871444 cites W2951959563 @default.
• W117871444 hasPublicationYear "2005" @default.
• W117871444 type Work @default.
• W117871444 sameAs 117871444 @default.
• W117871444 citedByCount "2" @default.
• W117871444 countsByYear W1178714442016 @default.
• W117871444 crossrefType "proceedings-article" @default.
• W117871444 hasAuthorship W117871444A5005313203 @default.
• W117871444 hasAuthorship W117871444A5066611124 @default.
• W117871444 hasConcept C105795698 @default.
• W117871444 hasConcept C106131492 @default.
• W117871444 hasConcept C107673813 @default.
• W117871444 hasConcept C111472728 @default.
• W117871444 hasConcept C119857082 @default.
• W117871444 hasConcept C138885662 @default.
• W117871444 hasConcept C14036430 @default.
• W117871444 hasConcept C140779682 @default.
• W117871444 hasConcept C151730666 @default.
• W117871444 hasConcept C154945302 @default.
• W117871444 hasConcept C160234255 @default.
• W117871444 hasConcept C162376815 @default.
• W117871444 hasConcept C173801870 @default.
• W117871444 hasConcept C19499675 @default.
• W117871444 hasConcept C2776214188 @default.
• W117871444 hasConcept C2779343474 @default.
• W117871444 hasConcept C2779530757 @default.
• W117871444 hasConcept C31972630 @default.
• W117871444 hasConcept C33923547 @default.
• W117871444 hasConcept C41008148 @default.
• W117871444 hasConcept C52740198 @default.
• W117871444 hasConcept C78458016 @default.
• W117871444 hasConcept C86803240 @default.
• W117871444 hasConceptScore W117871444C105795698 @default.
• W117871444 hasConceptScore W117871444C106131492 @default.
• W117871444 hasConceptScore W117871444C107673813 @default.
• W117871444 hasConceptScore W117871444C111472728 @default.
• W117871444 hasConceptScore W117871444C119857082 @default.
• W117871444 hasConceptScore W117871444C138885662 @default.
• W117871444 hasConceptScore W117871444C14036430 @default.
• W117871444 hasConceptScore W117871444C140779682 @default.
• W117871444 hasConceptScore W117871444C151730666 @default.
• W117871444 hasConceptScore W117871444C154945302 @default.
• W117871444 hasConceptScore W117871444C160234255 @default.
• W117871444 hasConceptScore W117871444C162376815 @default.
• W117871444 hasConceptScore W117871444C173801870 @default.
• W117871444 hasConceptScore W117871444C19499675 @default.
• W117871444 hasConceptScore W117871444C2776214188 @default.
• W117871444 hasConceptScore W117871444C2779343474 @default.
• W117871444 hasConceptScore W117871444C2779530757 @default.
• W117871444 hasConceptScore W117871444C31972630 @default.
• W117871444 hasConceptScore W117871444C33923547 @default.
• W117871444 hasConceptScore W117871444C41008148 @default.
• W117871444 hasConceptScore W117871444C52740198 @default.
• W117871444 hasConceptScore W117871444C78458016 @default.
• W117871444 hasConceptScore W117871444C86803240 @default.
• W117871444 hasLocation W1178714441 @default.
• W117871444 hasOpenAccess W117871444 @default.
• W117871444 hasPrimaryLocation W1178714441 @default.
• W117871444 hasRelatedWork W134960476 @default.
• W117871444 hasRelatedWork W145458542 @default.
• W117871444 hasRelatedWork W1558909906 @default.
• W117871444 hasRelatedWork W1948853487 @default.
• W117871444 hasRelatedWork W2182999079 @default.
• W117871444 hasRelatedWork W2464224693 @default.
• W117871444 hasRelatedWork W2472863411 @default.
• W117871444 hasRelatedWork W2559982436 @default.
• W117871444 hasRelatedWork W2593574362 @default.
• W117871444 hasRelatedWork W2765292055 @default.
• W117871444 hasRelatedWork W2782852064 @default.
• W117871444 hasRelatedWork W2950236988 @default.
• W117871444 hasRelatedWork W2950733259 @default.
• W117871444 hasRelatedWork W2951141273 @default.
• W117871444 hasRelatedWork W2991157682 @default.
• W117871444 hasRelatedWork W3004944559 @default.

• next