Phase 1

Navigation:

Introduction

Background
- Knowledge Engineering Bayesian Networks  (KEBN)
- Ecological Risk Assessment (ERA)

Events
- Phase 1
- Phase 2
- Phase 3

Conclusions

Downloads

Links

References

In this development phase, evaluation, and subsequent implementation was centered around a two-day stakeholder workshop.  Problems with the model identified before and during the workshop were addressed and possible solutions decided upon.  Following the workshop, an in-depth investigation of issues raised was conducted, identifying where possible improvements could be implemented within the time constraints of the project.  Improvements identified included better representation of spatial and temporal components of the model.

Methods:

Stakeholder Workshop:

In order to evaluate and improve the existing model a two-day workshop was held involving experts in fish ecology, management and modeling.  The workshop was conducted at the Water Studies Centre (Monash University) in May 2003 and was run by the domain expert developer, Dr Carmel Pollino.  The objectives of the workshop were:

• To briefly review the progress of the project to date.

• To review the structure of the Bayesian network produced for native fish abundance and diversity in the Goulburn Broken catchment.

• To incorporate the knowledge of the expert panel into the network.

• To identify key knowledge gaps in the model and specify how these gaps might be filled.

The Domain Expert Developer:

As with all iterative phases of development, improvements to the model were identified, implemented and evaluated in collaboration with the domain expert developer.  In the evaluation phase, the domain expert was consulted before and after the stakeholder workshop, and throughout the implementation of improvements.

Spatial and Temporal Components:

Originally, separate networks were prepared for each spatial scale of interest.  Each of the networks represented the physical and fishery data particular for each site, making the site dependent nodes the root nodes of the networks.  In order exploit the inter-site redundancies possible methods of incorporating networks into a single global network where examined.  One possible method identified was to include an additional Site and/or Type variable, which could be a parent node to all Site/Type dependent nodes.  In doing this all the data cases could be incorporated into a single case file with an additional Site/Type field. 

A Bayesian Network normally represents a single time step.  This was the case with the original networks.  In order to consider changes in nodes over multiple time steps the network needed to be extended.  One possible method identified was the addition of a new time node influencing the query nodes directly, in much the same way as the spatial representation alteration proposed above.  Another possible method was to repeat the Bayesian Network over multiple time steps, called a Dynamic Bayesian Network [2, Chapter 17].  Links between variables in different time steps, called dynamic relationships, need to be identified to use this type of extension.  In order to extend the model by either of these methods it was necessary to define the time scales of changes and predictions.  By matching temporal differences, the case data could be extended to include cases for the different time scales.

Model Changes:

The stakeholder workshop and subsequent evaluation with the developer resulted in changes to the model ontology and causal structure.  Following the model evaluation at the stakeholder workshop, and subsequent consultation with the domain expert, the complete structure, including Site, Type and Time Scale nodes was generated.