Download presentation
Presentation is loading. Please wait.
1
Parameterising Bayesian Networks: A Case Study in Ecological Risk Assessment Carmel A. Pollino Water Studies Centre Monash University Owen Woodberry, Ann Nicholson, Kevin Korb Computer Science and Software Engineering Monash University
2
Ecological Risk Assessment Process for improving environmental management to protect ECOLOGICAL values/assets Focus: Managing physical, chemical and biological processes to protect ecological endpoints Ecological sustainability/integrity poorly dealt with in many integrative analyses (not just environmental sustainability – human focus)
3
Ecological Risk Assessment Problem Formulation What are the environmental values of the system? What is it you want to protect? Conceptual Model Risk Analysis (Risk = Likelihood x Consequence) Risk Characterisation Risk Management
4
Ecological Risk Assessment Stakeholder engagement (adoption) technical non-technical Modular (multiple stressors - multiple models in single framework) Promotion of Iterative and Adaptive approaches to environmental management Inform future monitoring and targeted research needs
5
Models required for multiple stressor / hazard problems (complex issues) Models need to: incorporate information with high uncertainties incorporate disparate information be able to prioritise risks to endpoint be applicable in risk management Modelling Complex Issues (ERA)
6
Address Uncertainty and Complexity Increasingly being used in ecological applications Modular DSS Complex system composed of simpler parts (or multiple models) Inputs: expert opinion, literature, data, other models Predictions able to be tested (test robustness of predictions) and models easily updated Models simple (pragmatic), transparent and easily interpreted (adoption into risk management) Bayesian Networks
7
Two components Structure (Graph Theory) Probabilities (Probability Theory) Links in graph represent relationships between variables (as with a conceptual model) Probabilistic relationships (strengths) between variables Bayesian Networks
8
ERA Case Study Goulburn Catchment, Victoria, Australia
9
Important irrigation Area Important habitat for endangered and threatened native fish species Problem Formulation: Reduced abundance and diversity of native fish in the Goulburn Catchment, Victoria, Australia Adoption: Goulburn Murray Water Goulburn Broken Catchment
11
Model Structure Expert (Workshop 1) and Automated
13
Fish Network -5 sub-networks Water Quality Flow Structural Habitat Biological Interactions -2 query nodes Fish Abundance Fish Diversity -23 sites 6 reaches -2 temporal scales 1 and 5 year changes
14
Model Parameterisation Expert (Workshop 2) and Automated
15
Parameterisation Expert Elicitation used to parameterise aspects of model not represented by data (lack of variability in data set) Iterative process of updating expert derived probabilities (prior probabilities) using data (automated process) Needed this process to be supervised
16
KEBN BUT For parameterisation of a BN, there are no detailed methodologies for combining: qualitative and quantitative derived probabilities expert elicitation and automated discovery Develop KEBN – formalised process for parameterisation
17
Parameter Estimation KEBN has 3 “paths”: (1) Elicited from experts (2) Learned from data Routine Monitoring Data Targeted Research (3) Generated from a combination of sources Evaluation process essential to assess parameterisation process
18
Quantitative KEBN Spiral
19
Fish Network
20
Parameter Estimation Data variables: initially given uniform distribution Sparse or no data variables: elicited. Experts were asked to report confidence in estimates (equivalent sample size – ESS), to be used by data learning/training method: EM (Expectation Maximisation) algorithm.
21
Parameter Estimation Parameterisation after learning compared to original in “Assess Degree of Change” process. Identify where large changes occur Changes focus on where there are discrepancies in expert elicited probabilities and data derived probabilities.
22
Model Evaluation Quantitative Sensitivity Analyses Predictive Accuracy Qualitative Expert Real data vs. Model Prediction
23
Quantitative Evaluation Sensitivity Analysis Identify if a variable is either too sensitive or insensitive to other variables in particular contexts Identify errors in either BN structure or Conditional Probabilities Identify knowledge gaps Case study: Sensitivity analyses found that when water quality is low, other variables have less impact on Future Abundance. This agrees with expert’s understanding.
24
Quantitative Evaluation
25
Predictive Accuracy Data split (80% training, 20% testing) Error Rate (Future Abundance) = 5.8% Limited data Lack of variability in abundance of fish communities throughout catchment (mostly low – poor condition)
26
Figure 5: Relative Abundance Data (left axis - bars) versus BN Model Predictions (right axis - line) for Sites in the Goulburn Main Channel. Qualitative Evaluation
27
Test aspects of network not represented in data set Conditions required for ‘healthy’ native fish communities Robustness of network Fish Ecologists Environmental Managers / Natural Resource Managers Expert Evaluation
28
Risk Management framework Prioritise risks Identify knowledge gaps Allocate resources for: Further monitoring and research Risk mitigation GBC Bayesian Networks
29
Adaptive Management framework Monitor and Update Test assumptions in model Adopt and learn as: New information becomes available New situations arise GBC Bayesian Networks
30
Model specific for different communities of fish in Murray-Darling Basin Currently biased towards Low Flow Specialists Model represent dynamic changes (temporal changes) Future Improvements
31
Acknowledgements National Program for Sustainable Irrigation Goulburn Murray Water Stakeholders
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.