=Paper=
{{Paper
|id=None
|storemode=property
|title=ABC-BN: A Tool for Building, Maintaining and Using Bayesian Networks in an Environmental Management Application
|pdfUrl=https://ceur-ws.org/Vol-818/paper14.pdf
|volume=Vol-818
}}
==ABC-BN: A Tool for Building, Maintaining and Using Bayesian Networks in an Environmental Management Application==
https://ceur-ws.org/Vol-818/paper14.pdf
ABC-BN: A tool for building, maintaining and using Bayesian
networks in an environmental management application
Ann E. Nicholson, Owen Woodberry, Adrian Moorrees, Alicia Lucas
Steven Mascaro, Kevin Korb Biodiversity and Ecosystem Services
Bayesian Intelligence Pty Ltd. 2/8 Nicholson Street
2/21 The Parade, Clarinda East Melbourne, VIC 3002, Australia
VIC, 3169, Australia
Abstract tion (ABC) as a central resource for managing threat-
ened species and communities (DSE, 2009). It facili-
The Victorian state government Department tates the management of actions documented in Action
of Sustainability and Environment has a web- Statements prepared under the Flora and Fauna Guar-
based application called Actions for Biodi- antee Act 1988 and Recovery Plans prepared under
versity Conservation (ABC) as a central re- the Environment Protection and Biodiversity Conser-
source for managing over 400 threatened vation Act 1999. ABC currently holds information on
species and communities. ABC maintains in- more than 400 threatened species and communities1
formation for species and communities at in- and over 8,000 management actions at approximately
dividual locations, including lists of threats 2000 locations across Victoria. ABC maintains infor-
(to species and communities), actions (to mation for species and communities at individual lo-
mitigate threats) and population and habi- cations, including lists of threats (to species), actions
tat factors. Here we describe an extension to (to mitigate threats) and population and habitat fac-
ABC, a tool for building Bayesian networks tors. By bringing together information from a range of
for selected populations of threatened species sources, ABC is intended to make significant improve-
and occurrences of threatened communities, ments in both knowledge about threatened species and
to model the interactions between actions, communities and the transfer of that knowledge. The
threats and population and habitat factors. recording of actions implemented for populations and
The tool, called ABC-BN, also allows users communities enables the development of prioritized
to do what-if scenario reasoning, and is inte- lists of actions to be made on an increasingly sound
grated with ongoing monitoring and report- basis by land and water managers. Reporting on out-
ing. Unlike most ecological BN modelling to comes provides a basis for applying adaptive manage-
date, which typically takes months or years ment, whereby the effectiveness of management can
to produce a single, often complex, BN, for be improved based on the current and previous out-
a specific problem, the aim here was to pro- comes for a population. By 2004, ABC Stage 1 was
duce a tool that would facilitate the develop- operational and able to report on whether an action to
ment, maintenance and use of a large number ameliorate a threat had been carried out. ABC Stage
of quite simple standardized BNs, specialized 2 (deployed in 2006), incorporated the facility to pre-
for particular instances. We describe the in- pare actions plans within the system. However ABC
cremental development of ABC-BN over 3 did not support any modelling of interactions between
years. A prototype version was evaluated actions, threats and outcome factors and had no pre-
in 2009 by building models for 100 species, dictive capability. More importantly, it was not able
and the completed ABC-BN was deployed in to report on the status of a species or community for
April, 2011. which there was little or no data collected, in a way
that was comparable across the state.
Bayesian networks (BNs) (Pearl, 1988; Jensen and
1 INTRODUCTION Nielsen, 2007), are becoming increasingly popular for
The Victorian state government Department of Sus- 1
The FFG Act 1988 defines a community as “a type
tainability and Environment (DSE) has a web-based of assemblage which is wholly or substantially made up of
application called Actions for Biodiversity Conserva- taxa of flora or fauna existing together in the wild”.
environmental and ecological monitoring and risk as- ABC is a web-based interface to a database applica-
sessment (see § 5.2.3 in Korb and Nicholson, 2010 for tion. It is based around entities called items, which
a recent survey). There have been a number of mod- may be flora or fauna species, communities, or poten-
elling guidelines published (e.g., Varis and Kuikka, tially threatening processes. Within ABC, there are
1999; Borsuk et al., 2004; Renken and Mumby, 2009), different types of users, with different responsibilities
while Uusitalo (2007) reviews their features and use in and access levels. ABC is based on recording informa-
modelling environmental applications. In 2008, DSE tion about each item at each location (linked with a
decided to extend ABC with BN technology, to pro- GIS) it is found; there is a location monitor respon-
vide the capacity to model the interactions between sible for recording the management and monitoring.
actions, threats and outcomes and overall status, and The item monitor oversees the information about
allow users to do what-if predictive and diagnostic sce- each item, while system administrators have ad-
nario reasoning. These additional capacities would be ditional responsibilities such as appointing monitors
integrated with ongoing monitoring and reporting. and adding actions and threats to the global lists. Fi-
nally, there are various stakeholders (e.g., members
In most ecological BN modelling to date the knowledge
of other government organizations, NGOs, community
engineering paradigm is to develop a single, often com-
groups, academics and researchers), who can be given
plex BN, for a specific problem, which typically takes
access to the information in ABC, but may not make
months or years to complete (e.g., Pollino et al., 2007;
changes.
Smith et al., 2007; Chee et al., 2005). Here, in con-
trast, the aim was to produce a tool that would fa- For each item, the ABC database contains informa-
cilitate the development, maintenance and use of BNs tion about the threats that have been identified at
for specific populations and occurrences of species and each location, which are ranked along two dimensions,
communities with a high priority within ABC, rather “‘likelihood” and “impact”, representing priorities.
than entire species and communities. It records actions (from a central list of possible ac-
tions) that have been identified as potentially useful
This aim led to some key design decisions. First, the
to reduce threats to the item at that location. Fi-
tool would be for use by the DSE so-called monitors
nally, monitors must record monitored information in
– DSE scientists and Biodiversity Officers – already
actions taken, the status of threats, and various out-
maintaining information in ABC about the species
comes, which are observations relating to the species
for which they were responsible. The BN technology
populations or community occurrences, or the habitat
would be hidden from these users. Second, the tool
which can be monitored. This information facilitates
would support the building of simple BNs using a tem-
the prioritizing of species and communities based on
plate that enforced a certain causal structure between
the importance of the location and contributions of
actions, threats and outcomes.
actions to mitigate threats.
In this paper, we describe the incremental develop-
ment of the tool, called ABC-BN, over a 3 year pe-
riod, with the challenges of changes to the underlying 3 ARCHITECTURE
BN template, and a number of substantial additions
to the tool requirements. We give an overview of the The key requirement for ABC-BN was that it be com-
ABC-BN’s functionality, which is based on the itera- pletely aligned with the existing ABC application,
tive, incremental knowledge engineering of BNs Laskey with the same web interface, the same types of users,
and Mahoney (2000) A prototype version was evalu- and model the same action, threat and outcome infor-
ated in 2009 by building models for 100 species, and mation in the ABC database.
the completed ABC-BN was deployed in April 2011. Figure 1 shows the system architecture for integration
of ABC and ABC-BN. ABC-BN is invoked via a menu
tab from ABC, with the current item as well as the
2 ACTIONS FOR BIODIVERSITY
user ID, passed to ABC-BN. From that point, ABC-
CONSERVATION (ABC) BN interacts directly with the users via their browser,
with uploads and downloads of BN models allowed via
The first version of ABC was developed for DSE by a shared DSE file system. ABC-BN is a TOMCAT
Spatial Vision, a software company specialising in GIS application, implemented using JavaServer Pages and
applications. Stage I became operational in 2004, with the NeticaJ Java API.2 The BNs and ABC-BN specific
the Stage II in place from 2006. The addition of ABC- information are stored in new tables added to the ABC
BN, to be developed by Bayesian Intelligence, became Oracle database.
part of Stage III, along with enhancements to ABC
2
itself, to be delivered by Spatial Vision. www.norsys.com
Figure 2: ABC-BN development phases
Upload (Strict Standard)
Upload (Standard) Upload (Non-Standard)
Add New
Edit Existing
Edit
Edit
Draft Draft
(Strict Standard / Standard) (Non-Standard)
Submit
Submitted Reject
(Strict Standard / Standard)
Recommend
Recommended
(Strict Standard / Standard)
Approve
Approved
(Strict Standard / Standard)
Supercede
Archived
(Strict Standard / Standard)
Figure 1: System architecture for the integration of Figure 3: Model lifecycle
ABC and ABC-BN
3.1 Incremental and Iterative Development
(by a new approved model) it becomes archived. Fig-
ABC-BN was designed and built incrementally, follow- ure 3 shows the model lifecycle.
ing the prototype-based spiral software development Two user roles were added for ABC-BN: (1) Model
cycle advocated by Brookes (1995) and Boehm (1988). administrators have privileged access to the ABC-
There were three main phases (see Figure 2): a pre- BN module, notably approving models and the abil-
liminary scoping phase, then two main design and de- ity to edit the ABC-BN global settings, upload Netica
velopment phases. These were due to (1) changes to BNs to the system, and delete any model; (2) Out-
the underlying BN template and (2) a number of sub- come monitors are nominated by the model admin-
stantial additions to the tool functionality. istrators for each item location.3 Outcome monitors
may build and submit models for approval, while item
monitors are responsible for recommending models.
4 USER ROLES AND MODEL
MANAGEMENT Table 1 summarizes the user roles within ABC-BN.,
while Figure 12 shows the model management page.4
ABC-BN provides model management and an approval
workflow, together with an audit history. The BN
models are managed for each item location; there may
be many draft models, but only one approved model 3
This is because they may be someone other than the
at any one time. Models must go through a three step location monitor.
submission process: submitted → recommended 4
Figures showig screen shots of ABC-BN are placed to-
→ approved. When an approved model is superseded gether at the end of the paper.
Table 1: User roles
Role Abbr. Comments
Model Ad- MA Full access to the ABC-BN module. Table 2: Elicitation algorithm
ministrator Able to create, submit, recommend
and approve models.
Item Moni- IM Access all models. Able to submit Module Sub-module Algorithm
tor and recommend models for locations -create standard template
under the item. model
Outcome OM Access all models. Able to create and Structure: Name: -name the model
Monitor submit models for item location. Threat: -select threats to be included
Any (Other) M Access all models. in the model
Monitor Actions -for each threat variable
Stakeholder SH Access only Approved models. & Asset -select action variables affect-
ing this threat
Factors: -select asset factor variables
affected by this threat
5 MODEL BUILIDNG States: -for all action, threat and asset
factor variables
-select a state space
ABC-BN supports all stages of the BN elicitation pro- -if optional state description
cess. The elicitation algorithm is shown in Table 2. -give description
Additional -for each threat and asset fac-
Navigation around the elicitation process is flexible, tor variable
the user may move to any module, and any sub-module Links: -select other variables, of
the same type, affected by this
within that module (using tab menus). Thus it pro- variable
vides the iterative and incremental BN construction CPTs: -for all threat, asset factor and
query variables
process advocated by many (e.g., Laskey and Ma- -for each parent state combi-
honey, 2000; Korb and Nicholson, 2010; Boneh, 2010). nation
-select verbal cues for each
Figure 4 shows an example of a BN created within the state
system.5 Scenarios: -test model with scenarios tool
5.1 Structure
ABC-BN supports building models specific to each
item location, based on a pre-defined structure tem-
plate, incorporating actions, threats and asset factors
in the ABC database. This template evolved over
ABC-BN’s development phases as shown in Figure 5:
Phase 1: Template contains only threats, which are par-
ents of a ThreatTrend node, with values
{Improving, Stable, Worsening}
Phase 2: Template contains Action nodes (all root nodes)
which are parents of the Threat nodes, which
are in turn parents of AssetFactor nodes, as
well as two Trend nodes.
Phase 3: Trend nodes modelling change now replaced
by Status nodes representing absolute val-
ues: {Severe,Moderate,Negligible} for Threat-
Status and {Good, Fair, Poor} for AssetSta-
tus.
The elicitation process builds only BNs that follow this
so-called strict standard template, which limits the
complexity of the networks and ensures they are easily
analyzed for reporting (based on the Status nodes).6 Figure 5: ABC-BN’s templates from Phase 1 (above)
and Phase 3 (below). The Phase 2 template was as for
5
This BN was produced by Phil Papas and Di Crowther Phase 3, but contained Trend nodes, instead of Status
of DSE, Biodiversity and Ecosystem Services, Arthur Ry- nodes.
lah Institute.
6
ABC-BN also allows uploading of BNs (for example,
constructed in Netica) that do not follow the strict stan-
Figure 4: An example complete BN created within ABC-BN.
ure 13). Once the Threat nodes are selected, the
user then selects the associated Action and Asset-
Factor nodes for each (see Figure 14).
5.2 Parameter Elicitation
Eliciting the parameters of the BN, the conditional
probabilities, is recognised as a difficult task. For each
node, there is distribution for each combination of val-
ues of the parent variables; this is exponential in the
Figure 6: Threat Ranking page. number of parent variables. It can be hard for experts
to express their experience/opinion in numbers, and
they are often inconsistent. Hence the decision was
The BN nodes are selected from the values already made for ABC-BN to support the alternative qualita-
in the database for that item and location. The dis- tive assessment of probabilities using verbal anchors,
crete state space for each node is selected from a global developed by van der Gaag et al. (1999), and imple-
list (intended to provide consistency across the system, mented in the Verbal Elicitor tool (Hope et al., 2002).
and maintained by the Model Administrator), divided The user is presented with scenario descriptions (about
into “types”; for example the “Choice” type could have combination of parent values) and a selection of com-
alternatives {Yes,No} and {Done,NotDone}, while an mon chance tags, e.g., “certain”, “likely” or “impos-
example “Amount” type is {High, Medium, Low}. sible”. In this way actual probability values are not
The user first selects threats from a list of threats for required or shown to the expert. ABC-BN contains
that item. Administrators can assign a threat to one a so-called verbal map (which is maintained by the
of two ranks based on the impact and likelihood of a Model Administrator) which maps verbal cues to prob-
threat, (see Figure 6), which changes the way threats abilities (see Figure 15).
are presented on the Threat Selector page (see Fig- For Phase 2, ABC-BN provided only this qualitative
elicitation, and the users could not see the probabil-
dard: (1) standard models, which are similar to the strict ities stored into the CPT.7 Feedback from the users
standard structure, but without any constraints on the during the Phase 2 evaluation indicated that this was
links, and may be submitted for approval; (2) non-standard
7
models can be any Netica BN, but may not be submitted That is, unless they downloaded the BN for viewing in
for approval, as they can’t be analyzed for reporting. Netica!
Figure 8: Example of table parameter elicitation:
(above) Verbal cue; (below) Number
“any” state.8 Users can switch between a single ques-
tion (i.e., a parent state combination) and the full table
view. ABC-BN also provides visualization of the pa-
rameter elicitation progress, incorporating a summary
table (not shown) with colour coding for questions that
are either complete, incomplete or in need of revision
(i.e., when there have been structural changes).
6 Reasoning in ABC-BN
Figure 7: Example of single question parameter elici- ABC-BN’s reasoning component (developed in Phase
tation: (above) Verbal cue; (below) Number 2), called “Scenarios”, allows the user to enter any
combination of exact or likelihood evidence and pro-
vides a visualization of the posterior distributions af-
ter belief updating. The screen is layed out in four
quite frustrating. Also, by their very nature, the prob- columns – Actions, Threats, Asset Factors, and Sta-
abilities elicited using verbal tags will be inaccurate, tus – corresponding to the layers in the BN template,
as the cues are limited to only some probabilities, while the arcs in the underlying network are hidden.
and they are nearly always changed during normaliza- is shown in Figure 9.
tion. Hence in Phase 3 the parameter elicitation was To enter evidence, the user clicks on a node’s “Add
changed to a hybrid system, combining both qualita- Evidence...” button, which changes the node view to
tive and quantitative elicitation, catering for an itera- show a slider for evidence. This rectangular slider is
tive stepwise refinement of the probabilities. split into as many parts as there are states, where
Basic elicitation is done via single questions, where the the length of each part represents the probability of
user can switch between “Words” and “Numbers”; an a state. The user can either drag the dividers, or en-
example of these alternatives, for a Fire threat node, ter number values, to enter the desired likelihood evi-
with a single Backburning action parent node, is dence. Alternatively, the user can simply select a state
shown in Figure 7. In single question mode, the “Con- if the evidence is exact.
tinue” option takes the user to the next combination of In Phase 3, the reasoner was extended with a simi-
parent values - that is, the next row in the CPT. Note larity analysis report, which summarizes the current
that the user is told the question is “X/Y”, meaning scenario, that is, the model and the entered set of ev-
Xth question out of Y total questions. idence; an example is shown in Figure 10. As well as
Feedback on the Phase 2 prototype also indicated that reporting the calculated probabilities for all the nodes,
the users found it hard to “calibrate” their answers column 1 shows the difference between the posteriors
when answering these single questions, so Phase 3 in- and the initial distribution (before evidence was en-
cluded an alternative table interface, allowing visual- tered), while column 2 shows differences between the
isation and flexible navigation of the CPT (in either final distribution and distribution given only the Ac-
word or number form). An example is shown in Fig- tion node evidence.
ure 8. Navigation of the table can be done by either 8
The “any” state is useful if one parent (or a set of
clicking on a summary table, or by selecting a parent parents) dominate in a particular combination, rendering
state combination via drop down boxes , including an other parents inconsequential.
Figure 9: Scenario tester page
Figure 10: Similarity analysis report
terior probabilities of the two Status nodes (given the
evidence), plus metadata on all included factors (ac-
tions, threats and asset factors), such as the source of
the information (observed data, expert or/and litera-
ture).
As well as reports on individual items, the system also
produces a summary table with categorization, across
a user-specified set of items; an example is shown in
Figure 11. This table is an aggregation of the clas-
sification using the highest posterior probabilities for
Status nodes computed for each BN, given the out-
comes evidence added by the monitor. Results are
defined as Inconclusive if the difference between the
two most probable states is less than the “inconclusive
threshold”, a global variable (modifiable by adminis-
trators).
Figure 11: Outcome summary page
8 EVALUATION
7 OUTCOME SUMMARY
The prototype developed in Phase 2 was trialled in
This component was added in Phase 3, integrating the May-June, 2009 in 5 elicitation workshops around the
BN models with the monitoring and reporting func- state. It was used to generate BNs of 100 threat-
tions of ABC. For each model, the user enters evidence ened species and communities. In these trials, the
for the BN, for any of the Action, Threat or Asset- facilitators were DSE personnel9 overseeing the devel-
Factor nodes, along with meta-data. This evidence is opment of ABC-BN, while the workshop participants
stored in the database with the BN. The model is then were other ABC monitors. These gave very positive
re-run with this evidence, and an Outcome Report is 9
The facilitators were authors Moorrees and Lucas –
produced. This report contains a summary of the pos- ABC administrators, as well as item and location monitors.
feedback, as well as suggestions for improvement that DSE (2009). Actions for biodiversity conservation (abc):
were adopted for Phase 3. Managing our threatened species and communities.
http://www.dse.vic.gov.au/DSE/nrenpa.nsf/LinkView
Phase 3 was deployed early April, 2011, with the first /487AAFDF542DDAC6CA25703F001C16FEB44D7EEB86E4BDFD
round of outcomes for 100 models due at the end of CA257115001408B8.
June. We intend to undertake a more formal evalua- Hope, L., A. Nicholson, and K. Korb (2002). Knowledge
tion of both the tool, and of the individual models de- engineering tools for probability elicitation. Technical
veloped by it. The tool can be evaluated by measuring report 2002/111, School of Computer Science and Soft-
ware Engineering, Monash University.
the rate of elicitation (e.g. number of models, number
of parameters), with qualitative feedback from the ex- Jensen, F. V. and T. D. Nielsen (2007). Bayesian networks
and decision graphs (2nd ed.). New York: Springer Ver-
perts regarding their level of comfort using the tool. lag.
The quality of the models is harder to assess, but we
Korb, K. B. and A. E. Nicholson (2010). Bayesian artificial
intend to have a selection (10%) reviewed by experts intelligence (2nd ed.). Chapman & Hall/CRC.
not involved in the elicitation or approval process.
Laskey, K. and S. Mahoney (2000). Network engineering
for agile belief network models. IEEE: Transactions on
9 CONCLUSION Knowledge and Data Engineering 12 (4), 487–498.
Pearl, J. (1988). Probabilistic Reasoning in Intelligent Sys-
tems. San Mateo, CA: Morgan Kaufmann.
We have described ABC-BN, a tool for building, main-
taining and using Bayesian networks in an existing en- Pollino, C., O. Woodberry, A. Nicholson, K. Korb, and
B. T. Hart (2007). Parameterisation of a Bayesian
vironmental management application. In contrast to network for use in an ecological risk management case
most knowledge engineering of ecological applications study. Environmental Modelling and Software 22 (8),
to date, ABC-BN supports the construction of a large 1140–1152.
number of simple, standardized BNs over a relatively Renken, H. and P. J. Mumby (2009). Modelling the dy-
short time period. It supports iterative and incremen- namics of coral reef macroalgae using a Bayesian belief
tal model construction, including hybrid qualitative network approach. Ecological Modelling 220 (9-10), 1305
and quantitative parameter elicitation. It allows users – 1314.
to do what-if predictive and diagnostic scenario rea- Smith, C., A. Howes, B. Price, and C. McAlpine (2007).
soning, and the BN models are integrated with mon- Using a Bayesian belief network to predict suitable habi-
tat of an endangered mammalThe Julia Creek dunnart
itoring data and reporting, providing outcome status (Sminthopsis douglasi). Biological Conservation 139,
reports for both individual items as well as status sum- 333–347.
maries for sets of items. Overall, the tool allows man- Uusitalo, L. (2007). Advantages and challenges of Bayesian
agers to estimate the costs associated with actions to networks in environmental modelling. Ecological Mod-
mitigate threats and objectively assess the relative im- elling 203 (3-4), 312–318.
portance of actions for different species populations van der Gaag, L. C., S. Renooij, C. L. M. Witteman,
and community occurrences. B. M. P. Aleman, and B. G. Taal (1999). How to elicit
many probabilities. San Francisco, CA, pp. 647–654.
Varis, O. and S. Kuikka (1999). Learning Bayesian decision
References analysis by doing: lessons from environmental and nat-
Boehm, B. W. (1988). A spiral model of software develop- ural resources management. Ecological Modelling 119,
ment and enhancement. IEEE Computer , 61–72. 177–195(19).
Boneh, T. (2010). Ontology and Bayesian decision net-
works for supporting the meteorological forecasting pro-
cess. Ph. D. thesis, Clayton School of Information Tech-
nology, Monash University.
Borsuk, M., C. Stow, and K. Reckhow (2004). A Bayesian
network of eutrophication models for synthesis, pre-
diction, and uncertainty analysis. Ecological Mod-
elling 173 (2-3), 219–239.
Brooks, F. (1995). The Mythical Man-Month: Essays
on Software Engineering (Second ed.). Reading, MA:
Addison-Wesley.
Chee, Y., M. Burgman, and J. Carey (2005). Use of a
Bayesian network decision tool to manage environmental
flows in the Wimmera river, Victoria. Report No. 4,
LWA/MDBC Project UMO43: Delivering Sustainability
Through Risk Management, University of Melbourne,
Australia.
Figure 12: Model management page
Figure 13: Threat selector page
Figure 14: Action and Asset Factor selector page
Figure 15: Verbal cues page
Figure 16: Example outcome page