ISSN:
1573-7470
Source:
Springer Online Journal Archives 1860-2000
Topics:
Computer Science
,
Mathematics
Notes:
Abstract A Bayesian network is a knowledge representation framework for encoding both qualitative and quantitative probabilistic dependencies among a set of propositional (or random) variables. An important type of probabilistic inference in a Bayesian network is the derivation of the most probable composite hypotheses — a set of hypotheses composed of multiple variables in a network. Such a type of probabilistic inference, however, is computationally intractable. In this paper an adaptive reasoning approach based on qualitative interval arithmetic is proposed as a method of dealing with the computational problem. Using this approach, a qualitative boundary, which reflects the upper and lower limits of a posterior likelihood, can be derived for each composite hypothesis. The advantage ofbounding each composite hypothesis qualitatively is that the quantitative values of the posterior likelihoods are not all necessary in the course of an inference. Consequently, an exhaustive evaluation can be avoided. The complexity of the proposed approach can be demonstrated to be no worse than that of a direct computation and in some cases, the computation is only a small fraction of that required in a straightforward direct computation.
Type of Medium:
Electronic Resource
URL:
http://dx.doi.org/10.1007/BF01530955
Permalink