untitled

(ff) #1

346 14 Bayesian Networks



  • Frequentist (ML) techniques make no a priori assumptions.

  • Bayesian (maximum a priori) techniques start with a prior distribution
    and gradually improve it as data become available.

  • EM is used for determining the distribution of a random variable that is
    not directly observable.

  • Connectionist networks are special kinds of statistical models for which
    there are efficient machine learning techniques.


14.3.3 Building BNs from Components


Modern software development methodologies make considerable use of com-
ponents that have already been developed and tested. This greatly reduces
the amount of effort required to develop large systems. An approach to BN
development that makes use of previously developed BN components has
been proposed in (Koller and Pfeffer 1997). In addition to reducing the BN
development effort, this approach may also be able to improve performance
during inference. This approach is known asobject-oriented Bayesian networks
(OOBNs). The basic OOBN concept is called an “object.” An OOBN object
can be just a random variable, but it can also have a more complex structure
via attributes whose values are other objects. An OOBN object can corre-
spond to an entity or it can be a relationship between entities.
AsimpleOOBN object corresponds to a BN node. It has a set of input
attributes (i.e., the parent nodes in the BN) and an output attribute (i.e., its
value). AcomplexOOBN object has input attributes just as in the case of a
simple OOBN object, but a complex object can have more than one output
attribute. The input and output attributes define the interface of the OOBN.
The interface defines the formal relationship of the object to the rest of the
BN. It can also have encapsulated attributes that are not visible outside the
object. A complex object corresponds to several BN nodes, one for each of
the outputs and encapsulated attributes. The notion of a complex object is a
mechanism for grouping nodes in a BN. The JPD of an OOBN, as well as the
process of inference, is exactly the same whether or not the grouping is used.
However, by grouping (encapsulating) nodes into objects, one gains a num-
ber of significant advantages:


  1. Complex objects can be assigned to classes which can share CPDs. Reusing
    CPDs greatly simplifies the task of constructing a BN.

Free download pdf