To design Bayesian networks


  • Ann-Sofie Stenérus
  • Anna Lindberg
  • Alexander Ahl

Publish date: 2008-12-29

Report number: FOI-R--2680--SE

Pages: 56

Written in: Swedish


  • Bayesian network theory
  • Bayesian modelling
  • probability
  • distribution
  • conditional distribution
  • decision support
  • decision support models
  • expert knowledge


This report has been written as part of the project Operational Analysis Methods (OAM) to build competence on the applications of Bayesian network theory at the Division of Defence Analysis, FOI. The report is a continuation on previous work and is mainly based upon experiences from within the division when working with the method as a decision support tool for Swedish defence applications. Interviews with colleagues and the authors´personal experiences constitute the main basis of the report but articles and other sources of information have also been used. To exemplify knowledge, experiences and discussions, two Bayesian network models developed at the division are described continuously throughout the report. One of the models is used to assess risk in proving grounds namely The risk assessment model for UnExploded Ordnance, (RVM for UXO). The other model, A Military Analysis method for Reliable Tactical Assessments, (MARTA), combines AHP analysis with Bayesian theory and is used to assess an army battalion´s capabilities during an exercise. The report initially discusses the types of problems suitable for Bayesian network modelling and when and how the networks can be used. The process of developing a Bayesian network can be separated into three phases. FFirst, to build a network a defined problem is required. Different methods to obtain this and to find variables important for the problem are therefore described. Second, once the problem has been limited and the variables found, a clear structure must be created. There are several critical issues the modeller should be aware of and which should, if possible, be avoided. Third, to create the final Bayesian network structure the variable conditions must be found and their probabilities determinded. At the Division of Defence Analysis, external experts are often invaluable sources of information in all phases. The needs of an interface often discriminate between the modeller and end users during the whole process, from development to implementation of the network. To solve this user related problem different interfaces can be developed. Equally important is the validation and verification to secure the model structure and data. The above mentioned problems can be solved in different ways, in part by the software used for modelling and this is discussed in the final sections of the report. Several conclusions can be drawn from the experiences of working with and using Bayesian networks. One conclusion is that a simplified structure with limited uncertainties often is preferred over a complex and more correct structure with larger uncertainties in the data and tables.