Risk Assessment and Decision Analysis with Bayesian Networks

171

7

Defining the Structure of

Bayesian Networks

7.1 Introduction

We have found that experts, in a range of application domains, often

experience common difculties and apply common strategies when

building Bayesian network (BN) models. The aim of this chapter is to

help you both avoid the common problems and benet from the com-

mon solutions. One particular difculty experienced concerns cau-

sality and the decisions about direction of edges in a BN; this is

described in Section 7.2. In Section 7.3 we focus on exploiting the

very similar types of reasoning that experts apply in very different

application domains. We have identied a small number of natural

and reusable patterns in reasoning to help when building BNs. We

call these patterns idioms. In Section7.4 we discuss another aspect of

BN modeling, called asymmetry, which presents some difculty.

Typically this is the result of conditional causality, where the exis-

tence of a variable in the model depends on a state of some other. We

look at the most common types of asymmetry and propose pragmatic

solutions in each case.

Although idioms are a powerful method for helping to model the

structure of a risk assessment problem as an individual BN, we need

a different method to help us build models that “scale-up” to address

large-scale problems. The most natural way to build realistic large-scale

BN models is to link individual models using ideas that originate from

object-oriented design. These multiobject BN models are described in

Section 7.5.

Finally in Section 7.6 we discuss the missing variable fallacy. This

occurs when we neglect to include some crucial variable in the model

and can only really reason about its absence by interrogating the model

results, in terms of predictions and inferences. The key argument here

concerns visibility of assumptions in the model.

Chapter 6 provided an overview of

the process involved in building a

BN and also provided some impor-

tant theoretical guidelines to help

you build the correct structural

relationship between variables in a

BN. The objective of this chapter

is to provide more practical guide-

lines to help you build a suitable

BN structure.

Visit www.bayesianrisk.com for your free Bayesian network software and models in

this chapter