Network fragments
Suzanne M. Mahoney, Kathryn Blackmond Laskey · 1999
Bayesian networks are widely used for reasoning about uncertainty. Small fixed models over a predefined set of variables work for many domains in which probabilistic relationships do not vary from one problem instance to the next. In complex domains, such as military situation assessment, every problem instance is unique, large and composed of many repeated patterns among sets of variables. This dissertation presents a representation, network fragments, that captures these repeated patterns of probabilistic knowledge. Network fragments form the knowledge base from which a problem-specific Bayesian network, called a situation-specific network, is constructed to reason about a particular problem instance. Each fragment may represent not only a subset of the variables within a situation-specific network, but it may also represent a subset of the distribution for a dependent variable given its parents. Network fragments capture both the structure and parameters of probabilistic knowledge at a granularity that is conducive to efficient and effective knowledge elicitation, automated network construction, learning and inference. In particular, network fragments can depict the structural and parametric richness of probabilistic dependencies within Bayesian networks. Toward that end, this dissertation presents asymmetry networks, a representation that captures the structure an parameters of conditional probability tables. Subsets of an asymmetry network map to the partially specified distributions of network fragments. During automated construction, these fragments may be combined into a situation specific model. To support that task, this dissertation defines context algebra, an algebra for factoring and combining partially specified conditional probability tables. In addition, this dissertation presents a definition for situation-specific belief networks, which comprise the output of automated Bayesian network construction algorithms.