Presentations, Tutorials and Archived Program

Bayesian networks are increasingly used as a decision support tool in environmental management to explore the consequences of management scenarios and policies. They are particularly useful because of the capacity of BNs to quantify uncertainty and to incorporate prior information, either as data or expert opinion. To maximise the effectiveness of these models, it is essential that the decision maker understand the reliability of the model's outputs. Sensitivity analysis is one approach to evaluate the influence of any imprecision of data in the reliability of the predictions the model produces. In this study, we aimed to improve the reliability of a published BN model's output by applying expert opinion to modify the network structure and the prior information on the conditional probability tables. The former study developed a BN with the purpose of managing nuisance aquatic macrophytes in rivers; the objective was to evaluate benefits of various restoration options. It consisted of factors that determine light conditions and availability (resources) and those that influence flow conditions (disturbances), the aquatic macrophyte cover in the streams, and their cause effect relationships. We constructed the BN of the same case study using the same dataset, while adding extra nodes of nutrient factors (NH3, FRP) as further resources to the macrophyte cover. A comparison between the two BNs showed that the sensitivity of the target node towards other nodes of the network was decreased by a minimum of 15% up to 61%. The results revealed that the new network was less sensitive towards the precision of data (measurements, expert opinions, etc.) in other nodes. Decreasing the sensitivity makes the model more robust against any imprecisions in available data and increases the extent of reliability of any decision-making based on the model's outputs. This is the first step in a larger research to contribute to BN development in determining the validity of the model for decision making at different management levels. more >

In Queensland, as in the rest of the world, the demand for electricity continues to grow. But, of greater importance to electricity agencies has been the even faster growth in peak demand. The requirement to have adequate generation capacity and distribution networks for this peak leads to inefficiencies and higher costs. Stakeholders are interested in reducing the need for the provision of expensive network capacity. Interventions in the system that incorporated only technical solutions have been found not to succeed and there is a need to combine aspects about the customers and customer-industry engagement activities. Modelling this electricity system is complex with difficulties in gaining an understanding of the interactions of the elements. However, these disparate variables can be integrated using a Bayesian network (BN) model.

To meet a need to providing an agency partner with a model that they could manipulate without having to learn to use specialist modelling software, the BN was developed using MS Excel. To reflect a BN, the model was implemented using a separate worksheet for each node of the model. The navigation in the workbook was aided by hyperlinks, both textual and graphical. Data entry was facilitated by consolidating scenario selection items and key, changing parameter values on easily accessed sheets with these then being transmitted to their respective sheets. Of particular interest to people considering using Excel will be the construction of the calculations of output probabilities for the states of each node. Impacts and intermediate values for the states of nodes were provided through tables and graphically.

Using Excel as a modelling approach allowed the easy integration of the probabilistic social dimension with data on appliance use with the externally modelled responses of changes in heat loads. Further it will allow non-specialists to use the model to aid policy and marketing decisions. more >

Management agencies are often tasked with providing advice regarding risk to residents. Risk is based on probabilities and uncertainties and conveying such information to the general public is often difficult. Bayesian Networks are an ideal tool for risk assessment, but what is the best way to convey outputs of these models to the public?

Adequately preparing for wildfire is one of the key factors in determining the survival of property and residents. Fire management agencies provide advice regarding how to prepare, but this information has traditionally been in the form of checklists. Residents commonly tick off the cheapest and easiest actions but these do not necessarily result in the greatest reduction in risk. As a result residents can have an inflated sense of security as they perceive their level of preparation to be adequate.

In an attempt to improve the situation, we were contracted to prepare a preparedness tool using a Bayesian Network approach. While the approach seemed the best method for a risk based assessment, a range of external factors meant that the model became largely deterministic. In this presentation, we outline these factors. Of particular concern, was the issue of litigation should advice result in negative outcomes. more >

The NSW Rural Fire Service (RFS) provides a web-based service that allows householders to assess their ability to defend their homes from fire. The existing tool asks the user a series of questions aimed at two ends: 1) assessing the maximum possible intensity of fire at the home and 2) assessing the individual's capacity for dealing with a fire of that intensity. Based on this assessment, along with simple if-then rules, the tool gives advice as to whether or not it is safe for the householder to stay and defend their home in the event of a bush fire.

RFS engaged a team from the University of Wollongong to replace this simple rule-driven approach with a more comprehensive risk-based model in the form of a Bayesian network. RFS required a user-friendly web-based interface to this model, which educates the householder on the fire risks as well as on the impact of taking action to reduce those risks. The RFS was keen to ensure that the new tool was both informative and engaging.

Bayesian Intelligence has developed this web-based tool for the RFS, which has a strong focus on intuitive representations of complex risk assessments. The tool keeps the user engaged and informed throughout the assessment by providing immediate graphical feedback on the impact of user responses. While the underlying Bayesian network is largely deterministic, the tool nonetheless manages to better convey the uncertainties involved in such assessments, and provides a robust interface as future improvements are made to the model. more >

Wildfires occur frequently in the Mediterranean region especially in hot and dry summer periods. Fires are ignited mainly by humans, whereas fire spread is influenced by weather conditions (e.g. wind speed), topography (e.g. slope) and vegetation type. Wildfire risk estimation can prove valuable for planning preventive measures (e.g. tree thinning, prescribed burning, insurance actions) and mitigation measures (e.g. evacuation, suppression). Wildfire risk estimation consists of two components: the probability of an occurring wildfire event and the respective consequences. The latter can be expressed by vulnerability and exposure of the biotic and abiotic systems affected by the wildfire. Vulnerability is defined as the degree of expected damage as a function of hazard intensity. Exposure refers to the items at risk, such as people and property.

This study focuses on the development of a wildfire risk assessment system at the meso scale for a Mediterranean area. We achieve this with the construction of a Bayesian network (BN), due to its ability to facilitate the explicit modeling of all relevant parameters, their causal relationships and the involved uncertainties. Standardization and weighting of risk indicators ensures comparability of results. The BN model consists of three sub-models: the fire occurrence, the fire behavior and the consequence model. Consequences refer to residential building damage costs and human safety. Numerical investigations are made with spatio-temporal datasets (stored and managed in a Geodatabase) with a 1km² spatial and 1 day temporal resolution (time series 2006-2010) for the Mediterranean island of Cyprus. Prior probability distributions are additionally obtained from literature and expert knowledge. The results are shown in risk maps. more >

Sensitivity Analysis is a major part of the knowledge engineering process when designing Bayesian Networks. This analysis is concerned about how changes in the network will affect the outcomes. This is particular important in Bayesian networks where many of the parameters have been elicited from domain experts. We will be concerned with how findings or changes in parameter values at various nodes can change the beliefs at another node. Due to computational issues this analysis is normally restricted to changes in a single node. Therefore, in general, it is not possible to analyse possible interaction effects due to changes between different nodes or multiple parameters. We present our investigations into how the techniques from Experimental Design in Computer Simulations, can be adapted to perform sensitivity analysis in Bayesian Networks. These techniques enable us to perform sensitivity analysis involving multiple nodes or parameters. We consider a range of networks. Using Experimental Design theory together with information about d-separation, we show how we can determine what parameters should be changed and what values they should take, in these networks. We then describe how we can fit a meta-model to represent how the outcome depends upon the changes in the parameters. For the outcomes, we consider the standard measures used in sensitivity analysis, such as mutual information and variance, but we also show other measures, such as average response, for some networks can be analysed. Final we show how the techniques of sensitivity analysis in experimental design can be used to obtain the required analysis of the networks. more >

Anomaly detection has become an important issue as "Big Data" has gotten bigger. Application range across the spectrum, from network intrusion detection, fraud identification and forensic accounting to spotting health problems and epidemic outbreaks. Anomalies come in three main flavors: bad luck (pseudo-anomalies), changes in the dynamics of the data generation process (climate change) and cuckoos' eggs (genuine intrusions). The most common anomaly detection methods are data-driven, using fluctuations in densities or distance from unsupervised clusters to spot outliers. While these methods can spot outliers, they are very weak at discriminating between the three kinds of outliers. Model-driven detection can do better. Here we describe the use of Bayesian networks to drive the anomaly detection process. more >

Tennis Australia places high importance on producing athletes who are able to reach and sustain a professional ranking within the top 100. The aim of this study is to use Bayesian networks (BNs) as a means to model historical rankings data, thereby allowing for the prediction of an athletes likelihood of reaching a top 100 ranking during their career based on early career results. Historical data relating to the professional rankings of all athletes (worldwide) ranked between the years 1984 and 2011 has been sourced. We have built two BNs, both with a simple temporal structure, for predicting rankings in the future. The network structures are hand-crafted, while the parameters are learnt from the data. In the first BN, the ranking in each year is dependent on the previous year and the starting age. In this network, the overall career peak ranking is represented as a logical function of the yearly rankings. In order to avoid having all the yearly rankings as a parents of the career peek ranking node (which cannot be compiled), we use a simple modeling “trick”, introducing a set of intermediate nodes to do the logical combination of pairs. The second structure conditions the yearly rankings on career peak ranking only, implying that training (i.e., parameterization of the network) can only be done using retired athletes (with known peak ranking). From both these models, we can perform inference on the network to determine the probability an athlete will ever reach a top 100 ranking, given the results they have achieved thus far in their careers. We evaluate and compare the models using the data, doing a 80-20 training test split. We will present results, showing how the different models behave. We will also discuss the practical issues, including the data pre-processing and with the BN software, that arose during the model construction. more >

Information Theory is the new black. We used it to dress up causation, as in Korb, Nyberg & Hope (2011): the amount of information that causes give about effects can measure their influence. But more exotic colours can emerge when two causal influences interact. This may produce some flashy 'synergy', where the joint effect is more than if the two effects were independent, or it may produce some dull 'redundancy' (i.e. overdetermination), where the joint effect is less. But measuring these quantities is difficult, partly because both may occur in the same interaction. There have been several ingenious proposals for a plausible measure that overcomes this difficulty, such as Williams & Beer (2010), Griffith & Koch (2012), and Harder, Salge & Polani (2013). Such measures are potentially very useful, for example, in computational genetics. However, we argue that these recent attempts are not satisfactory, since they don't accord with intuition in some simple cases. Furthermore, there's evidently some conceptual confusion. For instance, these authors can't even agree about whether there's any redundancy in a firing squad. We suggest a way forward. more >

The story of how a team of 30 including engineers, data analysts, statisticians, business analysts and software engineers built 24 Bayesian Belief Networks (BBNs) into a corporate data warehouse in nine months. The models calculate monthly risk scores for over four million individual assets. Western Power runs the electricity distribution network in the south west of Western Australia, servicing over one million customers. An initiative was started in 2011 to use BBNs to calculate risk scores for individual assets in the network. After a prototyping phase, an aggressive project was launched in 2012 to build models that predict asset failure and consequence of failure for 24 assets classes including poles, transformers, conductors and insulators. The project included building the models into the corporate data warehouse to extract data and calculate risk scores on a monthly basis. The project overcame many challenges including scope creep, modeller capability, data availability, governance, and change management. Success was dependent on detailed scope documents, training, model templates and project management discipline. The result was a project that delivered to scope, on time and under budget. The models are now providing key data to support network investments of over $1 billion per year. more >

The River BN was designed to show how land-use and riparian vegetation condition influence river hydrology, sediment and nutrient regimes, and in turn stream benthic macroinvertebrates and algae. Insufficient data was available to make the model strictly numerical/predictive. However, the model does capture the dominant relationships between key ‘drivers’ of ecological change in catchments. The River BN is therefore a useful tool for exploring how river ecological condition can respond to changes driven by environmental planning (e.g. land and water use at catchment scale), large scale investment (riparian rehabilitation across a catchment), or small scale investment (restoring riparian forest in a single river reach). In this talk I will discuss construction and validation of the model, the model's limitations and what I learnt about modelling for environmental management through its construction. more >

Willow encroachment into the naturally mixed landscape of vegetation types in the Upper St. Johns River Basin in Florida, USA, impacts upon biodiversity, aesthetic and recreational values. To control the extent of willows and their rate of expansion into other extant wetlands, spatial context is critical to decision making. Modelling the spread of willows requires spatially explicit data on occupancy, an understanding of seed production, dispersal and how the key life-history stages respond to environmental factors and management actions. Nicholson et al. (2012) outlined the architecture of a management tool to integrate GIS spatial data, an external seed dispersal model and a state-transition dynamic Bayesian network (ST-DBN) for modelling the influence of environmental and management factors on temporal changes in willow stages. That paper concentrated on the knowledge engineering and expert elicitation process for the construction and scenario-based evaluation of the prototype ST-DBN. This paper extends that work by using object-oriented techniques to generalise the knowledge organisational structure of the willow ST-DBN and to construct an object-oriented spatial Bayesian network (OOSBN) for modelling the neighbourhood spatial interactions that underlie seed dispersal processes. We present an updated architecture for the management tool together with algorithms for implementing the dispersal OOSBN and for combining all components into an integrated tool. more >

The Victorian Government is reserving 15,000 hectares of land to protect native grasslands and offset urban growth in Melbourne. The Department of Environment and Primary Industries has been tasked with developing an adaptive management approach to guide the restoration and management of grasslands.

The adaptive management process relies critically on the ability to model the system and generate predictions of the outcomes and costs of management strategies. Bayesian networks were selected for this purpose as they are powerful tools for representing and reasoning with uncertainty and can incorporate expert-derived knowledge with empirical data produced via field experiments.

Ecological systems are highly complex, and it is difficult to simplify and model them without losing sight of important interactions. In order to reflect ecological processes like competition, selective stock grazing and herbicide responses, we chose to track groups of species over time, to make temporal predictions about changes in the system. To do this the model employed a number of extensions to basic Bayesian network methodologies: decision network utility nodes, to report the costs associated with interventions and the environmental value of the site; dynamic nodes and links to model the changes in state variables over time; and object oriented techniques to manage the complexity of the number of species and seasonal transitions.

The model was developed over a series of 7 expert workshops. In collaboration with the lead domain expert, we used the information gathered during these workshops to inform the construction, parameterisation and validation of the Bayesian network, forming the basis for iterative and incremental development during adaptive management. more >

Urban and peri-urban vegetation or forests provide important ecosystem services for people living in the oasis towns of the Taklamakan desert in Xinjiang, NW China. In this research project, we apply Bayesian Networks to elicit and integrate experts' knowledge on the role of vegetation to mitigate two local environmental problems: 1) Dust weather and 2) urban heat stress. A Bayesian Network has been developed during a workshop series consisting of 3 workshops with institutional stakeholders and scientists from various disciplines. As most workshop participants had difficulties with conditional probability tables, we elicited their knowledge in various ways (e.g. with units such as - to +++ or 0-1) and systematically converted this information into conditional probability tables afterwards. Before we started the workshop series, interviews with institutional stakeholders in the cities of Aksu and Korla revealed the need for them to know which tree and plant species are most effective in mitigating dust weather while needing the least irrigation. We addressed this need by developing two simple Bayesian Decision Networks with utility nodes for the irrigation needed (as cost) and plant-specific dust weather mitigation (as benefit). The overall stakeholder dialogue and the development of the Bayesian (Decision) Networks aim at 1) jointly identifying ways how to optimize these ecosystem services in order to mitigate dust weather and urban heat stress in oasis towns and 2) applying, adapting and evaluating Bayesian Networks for transdisciplinary knowledge integration. more >

Intertidal macroalga provides an important habitat for a wide variety of marine species, especially invertebrates. Algae become detached from shore by an assortment of means and drift at sea where they continue to serve as an important habitat, food source and mode of dispersal. Many species of alga are particularly susceptible to climate change, leading to their rapid decline along urbanized coastlines. This has already been witnessed in Sydney where Phyllospora comosa, a key habitat-forming species, has disappeared from 70 kilometres of coastline. I aim to use a Bayesian network (BN) to model the factors that influence the abundance of attached macroalga and its subsequent breaking from shore to become drifting wrack. I then plan to predict the potential impact of changes in the abundance of wrack to the associated invertebrate community. Data corresponding to current and future climate change was sourced from Australian government agencies (CSIRO, Bureau of Meteorology and Marine Parks Authority). Information regarding invertebrate communities was collected from two New South Wales locations. Major determinants of algal detachment include (1) grazing effort (2) storm frequency and intensity and (3) susceptibility to disease. These factors are forecasted to increase under climate change and may result in algal detachment at a rate that cannot maintain existing algal populations. The 71 wrack-associated invertebrate species I have identified to date are predicted to be effected differently by the change in abundance of drift algae; negative impacts are particularly likely for obligate rafters and those with long larval dispersal phases. Early versions of the BN indicate an overall loss of biodiversity. I anticipate this model will be a useful tool for predicting the impacts of climate change, providing insight into factors that are fundamental to preserving future biodiversity. more >

Over the past 10 years, the Bureau of Meteorology has been developing Bayesian networks to assist with the prediction of fog at airports (Boneh, 2010). These tools are now deployed and in use at Melbourne and Perth airports, and have shown good predictive performance. However the existing models do not explicitly represent time; there is a different network for predicting fog at each forecast point throughout the night. In addition, the current models only predict whether a fog will occur within the forecast period (e.g., before 12 noon the next day) and do not provide any prediction of time of fog onset or clearance. In this work, we describe a dynamic Bayesian network template that explicitly models the progression of conditions for fog formation over time, together with onset and clearance times. We show how this DBN can be used throughout the forecasting period, to update fog prediction as observations come in for various weather variables (including of course, fog itself). The template also allows weather and fog predictions from other tools (e.g., numerical weather prediction models) to be incorporated into the DBN fog forecasting process. The template for the DBN structure has been built with input from expert elicitation workshops held in both Melbourne and Perth). The template has been validated against a suite of hand-crafted fog scenarios. We are currently preparing the data to parameterize and test the models on real fog situations. more >

New Zealand currently exports $4.7 billion of wood products, including more than 12.7 million m3 of logs (almost all Pinus radiata). Currently all logs are treated to eliminate potential infestations by phytosanitary pests. The most common treatment used at present is chemical fumigation with methyl bromide, or in the case of China, phosphine is permitted. Scion has just embarked on a four year programme of research to evaluate the necessity of current mandatory end point phytosanitary treatments, such as fumigation. As an alternative we propose an ecologically based assessment process that determines actual phytosanitary risk so that the need for pre-export treatments can be evaluated. This concept uses ecological information, e.g., pest phenology, habitat requirements, developmental biology, and dispersal capabilities, to determine if the potential pest pressure at a given time and place warrants the need for the application of an end point phytosanitary treatment. This programme adopts a Bayesian Network approach to model infestation risk both spatially and temporally. The models rely heavily on: 1) quantifying the thermal development of pest species so that phenology can be predicted from current and future meteorological conditions, 2) understanding the influence of landscape context on best abundance, and 3) accurate estimates of pest dispersal abilities throughout the landscape. The programme is supported by a national Quarantine Pest Trapping Network (QPTN) that will provide 4 years of pest abundance data from sites in both forests and ports. The QPTN data will be used to make an initial case for a winter pest free area of production, contribute to the validate our Bayesian Network models, and provide the backbone of future official assurance monitoring programme to support the adoption of an ecologically based assessment of phytosanitary risks to reduce the need for phytosanitary treatments of export logs. more >

Dynamic Bayesian networks (DBNs) are a useful tool for modelling and prediction of dynamic systems. The present work describes a new DBN learning algorithm based on the causal discovery program CaMML. The developed algorithm is experimentally evaluated and compared with a number of existing approaches for learning dynamic Bayesian networks.

An additional experimental investigation of a broad class of DBN learning algorithms is also presented. Some algorithms learn the static network structure and temporal arcs in a dynamic Bayesian network separately (‘two-step’ algorithms), whereas other approaches learn the entire DBN structure at once. Under certain conditions, these two-step algorithms can learn incorrect network structures, even with arbitrarily large data sets. We explore the type of failures that can occur, and the extent which these two-step algorithms are subject to such failures. more >

One of the aims of the Better Border Biosecurity (B3) programme is to develop Bayesian Network (BN) models as decision support tools that aims to objectively evaluate biosecurity risks and help make decisions about importation of an agricultural commodity and managing the risks associated with commodities.

Developing a BN risk model for fresh produce imports to NZ we are faced with some important issues: (1) Putting together a network structure that captures the import pathway as completely as possible, potentially going back to orchard infestation levels in the exporting country, yet keeping the complexity of the network manageable; (2) Identifying the different units of aggregation from fruit to boxes to consignments and how the relationships at different unit levels can be implemented in the network structure. To this extent we have frequently used mathematical relationships to construct CPT where the unit of aggregation of a node is different to that of its parents; (3) It is important to understand what the stakeholder, in our case the Ministry for Primary Industries (MPI), needs and their perception of risk, and to determine how these can be incorporated into the functional BN constructed through a consultative process. We will present a draft BN risk model we have developed using Netica, and report on limitations and opportunities for improvement. more >