Knowledge Engineering a Bayesian Network for an Ecological Risk Assessment KEBNERA

1
Knowledge Engineering a Bayesian Network for an
Ecological Risk Assessment (KEBN-ERA)

• Owen Woodberry
• Supervisors
• Ann Nicholson
• Kevin Korb
• Carmel Pollino

2
Overview

• Background
• Ecological Risk Assessment (ERA)
• Bayesian Networks (BNs)
• Knowledge Engineering (KE)
• The Project (KEBN-ERA)

3
Ecological Risk Assessment (ERA)

• Objective
• To develop and test a generic framework to be
  used in the assessment of ecological risks
  associated with Australian irrigation activities
• This study is specific to the Goulburn Broken
  catchment
• The Bayesian Network is being developed under Dr
  Carmel Pollino of the Water Studies Center
  (Monash University)

4
Bayesian Networks (BNs)

• A BN is a probabilistic reasoning tool
• Enter evidence/observations/interventions
• Query to find the probability of an event

5
Knowledge Engineering (KE)

• What we need
• A set of variables (nodes) and their states
• A graphical structure
• Conditional probability tables for each node
• This is difficult to elicit from experts

6
Goals of the Knowledge Engineering field

• To make the task of knowledge engineering easier
  by
• Formalising the expert elicitation process
• Developing automated tools to aid in these tasks
• Integrate the knowledge obtained from each of
  these methods
• Ensure that the BN is created correctly
• Maintenance and continued development of the BN

7
The Project (KEBN-ERA)

• Objectives
• Analysis of KE process to date
• Identify and undertake improvements
• Evaluation of the BN
• Implement possible support tools

8
(No Transcript)
9
Some improvements already identified

• Improving the spatial representation of the
  network
• Improving the temporal representation of the
  network
• Exploring the effects of combining experimental
  data with elicited values

10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
Evaluation Methods

• Using Domain experts
• Elicitation review
• Sensitivity analysis
• Case based evaluation
• MATILDA
• Using Automated methods
• Predictive accuracy
• Expected Value
• Kullback-Leibler divergence
• Bayesian Information reward

14
Conclusion

• The Project (KEBN-ERA)
• Analysis of KE process to date
• Improvements
• Evaluation
• Implement possible support tools