Relational and object oriented Bayesian networks



Large and complex Bayesian networks are usually hard to design. Once designed,
a Bayesian network may need to be adapted to changes in the domain it represents.
We investigate several approaches to facilitating design and maintenance of
Bayesian networks.
Several frameworks have been proposed in the past that ease the specification
of Bayesian networks using object oriented ideas. The group has been working
with so-called object oriented Bayesian networks, and has proposed a framework
that supports a top-down modeling approach.Moreover, we have proposed methods
that exploits the characteristics of object oriented domains when learning the
parameters as well as the structure of a network.
Adaptability is a central concern addressed by the language of relational Bayesian
networks, which we develop, investigate and implement. Relational Bayesian networks
are a logic-based representation language for high-level, adaptable probabilistic
models. These generic models can be instantiated over concrete domains,which leads to a domain-specific model that can then be represented by a Bayesian
network. A simple example of this two-level modeling approach is a general model
of (genetic) inheritance, which can be instantiated over any concrete domain consisting
of members of a given pedigree. The Bayesian networks for the domainspecific
models often become quickly intractable for inference with increasing
size of the domain. As an alternative to Bayesian networks we have investigated
arithmetic circuits as a computational data structure for probabilistic inference in
domain-specific models. Empirical results show that in typical examples we
can handle with arithmetic circuits domains that are about 2-3 times as large as the
largest domains amenable to inference with Bayesian networks.
The language of relational Bayesian networks is implemented in the Primula
system. Key components of Primula are a constructor for standard Bayesian
networks from a relational Bayesian network and a concrete input domain, and an
importance sampling algorithm for approximate inference for model instances not
amenable to exact inference.
Effektiv start/slut dato19/05/2010 → …